Studies in Systems, Decision and Control 175

Nabil Derbel · Jawhar Ghommam

Quanmin Zhu Editors

New
Developments
and Advances
in Robot
Control
Studies in Systems, Decision and Control

Volume 175

Series editor
Janusz Kacprzyk, Polish Academy of Sciences, Warsaw, Poland
e-mail: [email protected]
The series “Studies in Systems, Decision and Control” (SSDC) covers both new
developments and advances, as well as the state of the art, in the various areas of
broadly perceived systems, decision making and control–quickly, up to date and
with a high quality. The intent is to cover the theory, applications, and perspectives
on the state of the art and future developments relevant to systems, decision
making, control, complex processes and related areas, as embedded in the fields of
engineering, computer science, physics, economics, social and life sciences, as well
as the paradigms and methodologies behind them. The series contains monographs,
textbooks, lecture notes and edited volumes in systems, decision making and
control spanning the areas of Cyber-Physical Systems, Autonomous Systems,
Sensor Networks, Control Systems, Energy Systems, Automotive Systems,
Biological Systems, Vehicular Networking and Connected Vehicles, Aerospace
Systems, Automation, Manufacturing, Smart Grids, Nonlinear Systems, Power
Systems, Robotics, Social Systems, Economic Systems and other. Of particular
value to both the contributors and the readership are the short publication timeframe
and the world-wide distribution and exposure which enable both a wide and rapid
dissemination of research output.

More information about this series at http://www.springer.com/series/13304

Nabil Derbel • Jawhar Ghommam • Quanmin Zhu
Editors

New Developments and

Advances in Robot Control

123
Editors
Nabil Derbel Jawhar Ghommam
Sfax Engineering School Department of Electrical
University of Sfax and Computer Engineering
Sfax, Tunisia College of Engineering
Sultan Quaboos University
Quanmin Zhu Muscat, Oman
University of the West of England
Bristol, UK

ISSN 2198-4182 ISSN 2198-4190 (electronic)

Studies in Systems, Decision and Control
ISBN 978-981-13-2211-2 ISBN 978-981-13-2212-9 (eBook)
https://doi.org/10.1007/978-981-13-2212-9

Library of Congress Control Number: 2018964939

© Springer Nature Singapore Pte Ltd. 2019

This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part of
the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation,
broadcasting, reproduction on microfilms or in any other physical way, and transmission or information
storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology
now known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication
does not imply, even in the absence of a specific statement, that such names are exempt from the relevant
protective laws and regulations and therefore free for general use.
The publisher, the authors, and the editors are safe to assume that the advice and information in this book
are believed to be true and accurate at the date of publication. Neither the publisher nor the authors or
the editors give a warranty, express or implied, with respect to the material contained herein or for any
errors or omissions that may have been made. The publisher remains neutral with regard to jurisdictional
claims in published maps and institutional affiliations.

This Springer imprint is published by the registered company Springer Nature Singapore Pte Ltd.
The registered company address is: 152 Beach Road, #21-01/04 Gateway East, Singapore 189721,
Singapore
Preface

The book New developments and Advances in Robot Control is intended to

encourage technology transfer in the field of autonomous systems and robotics.
The rapid development of new mechatronic systems has been strong propellers for
the advancement of control systems theory and application, as well as the fields of
sensors, industrial processes, and computer methods. Nonetheless, distinct robotic
systems with variety of profiles have still to be developed and controlled in a way
that good performances are guaranteed. This book provides an update on selected
topics in the field of robotics. It also offers an opportunity for researchers to present
an extended exposition of their new works in all aspects of robotic systems for
a wider and rapid dissemination by opening up new challenges to understanding
and developing novel control strategies for such robotic systems. Hence, this
book is fairly balanced between mathematical theory and practical applications of
control engineering aspects for robotic systems ranging from robot manipulators
to aerial vehicles. This makes the book suitable for final year undergraduates and
postgraduates, lecturers, and practitioners in the areas of robotic systems and control
engineering. It is then hoped that the compiled references by each authored chapter
would provide insightful and valuable information about related work in the field
and make up a new starting up for further research studies.
The book consists of 16 contributed chapters by scientific experts who are
specialized in various topics addressed in this book, and it focuses on multidisci-
plinary applications in the field of robotics. This volume has been organized into the
following chapters, which are regrouped into two parts.
• The first part of the book includes nine chapters on controlling robotic manip-
ulators. The first chapter introduces a novel resource allocation framework for
cloud robotics and explores the trends of robotic technology in Industry 4.0.
The second chapter proposes a novel approach based on 4 × 2 visual approach
to control a robotic manipulator. In Chap. 3, the parallel robotic machine is
extensively studied and controlled with different control approaches. Real-time
experimentation has been conducted to demonstrate the effectiveness of the
proposed approaches. Chapter 4 addresses the modeling problem of a cable-

v
vi Preface

driven parallel robot. Chapters 5 and 6 employ the backstepping techniques

along with advanced robust control schemes to control the 7-DOF Annat robot
manipulator. In Chap. 7, the nonlinear observer is employed to tackle the problem
of fault detection and isolation for robotic manipulator. Chapter 8 proposes a
novel robust technique for controlling a robotic manipulator using the Crone
control formalism. Finally, Chap. 9 applies the sliding mode control technique
to control the upper-extremity exoskeleton robot for rehabilitation purposes.
• The second part, containing seven chapters, covers the control of different types
of vehicles and discusses in separate chapters different adopted control strategies.
Chapter 10 applies Particle Swarm to design an optimal PID 2-DOF controller
for a hybrid vehicle. Chapter 11 describes a digital stabilizing controller for two-
wheeled robot, where the matrix fraction description theory is fully employed
along with the digital PID controller for the tracking and stabilization of the
two-wheeled robot. Chapter 12 proposes a path-planning algorithm based on
optimized fuzzy logic controller to steer a mobile robot in its environment. Along
the same spirit of path-planning, Chap. 13 offers a strategy for the navigation of
mobile robots in a cluttered environment, using artificial potential function based
on fuzzy logic design. Chapter 14 extends the navigation of single mobile robot
to multiples agents moving in a coordinated fashion as they try to cover a spatial
region in an optimal way. New sets of flocking-based algorithms are proposed
to illustrate the efficacy of the strategy in covering wide area in an optimal way.
Finally, the last two chapters expose the reader to new challenge in the motion
control of unmanned aerial vehicle, in particular, Chap. 15 solves the tracking
control problem of quadrotor-type UAV. A combination of the backstepping and
the RISE techniques is employed to guarantee robust three-dimensional tracking
algorithm for quadrotor-UAV. Chapter 16 addresses the problem of designing a
model-based estimation algorithm for the unobservable flapping angles while a
near-hover flight is considered.
We would like to acknowledge the efforts of many reviewers who devoted some
of their time in providing feedback to the chapter’s contributors; thus, all chapters
were meticulously peer reviewed. We would also acknowledge the efforts of all
authors who contributed to this book. They are among the researchers who keep
enlarging the envelope on the state of the art in the field of robotics. Lastly, we would
like to sincerely express our deepest gratitude to the Springer editorial staff for
their continuous support, assistance, and significant improvement in the manuscript.
Without their help, the book would not be published as scheduled.

Sfax, Tunisia Nabil Derbel

Muscat, Oman Jawhar Ghommam
Bristol, UK Quanmin Zhu
June 2018
Contents

1 Cloud Robotic: Opening a New Road to the Industry 4.0 . . . . . . . . . . . . . 1

Manal Aissam, Mohammed Benbrahim,
and Mohammed Nabil Kabbaj
2 4 × 2D Visual Servoing Approach for Advanced Robot
Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
Mohamad Bdiwi, Jozef Suchý, and Matthias Putz
3 A Redundant Parallel Robotic Machining Tool: Design,
Control and Real-Time Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
Hussein Saied, Ahmed Chemori, Micael Michelin, Maher El-Rafei,
Clovis Francis, and Francois Pierrot
4 Cable-Driven Parallel Robot Modelling for Rehabilitation Use . . . . . . . 81
Hachmia Faqihi, Maarouf Saad, Khalid Benjelloun,
Mohammed Benbrahim, and M. Nabil Kabbaj
5 Control of Robot Manipulators Using Modified Backstepping
Sliding Mode . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
Yassine Kali, Maarouf Saad, and Khalid Benjelloun
6 Robot Manipulator Control Using Backstepping
with Lagrange’s Extrapolation and PI Compensator . . . . . . . . . . . . . . . . . . 137
Yassine Kali, Maarouf Saad, Jean-Pierre Kenné,
and Khalid Benjelloun
7 Nonlinear Observer-Based Fault Detection and Isolation
for a Manipulator Robot . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 163
Khaoula Oulidi Omali, M. Nabil Kabbaj, and Mohammed Benbrahim
8 Crone Controller Design for a Robot Arm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 187
Ahmed Abid, Rim Jallouli–Khlif, Nabil Derbel, and Pierre Melchior

vii
viii Contents

9 Cartesian Sliding Mode Control of an Upper Extremity

Exoskeleton Robot for Rehabilitation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 201
Brahim Brahmi, Maarouf Saad, Cristobal Ochoa-Luna,
Mohammad H. Rahman, and Abdelkrim Brahmi
10 Canonical Particle Swarm Optimization Algorithm
Based a Hybrid Vehicle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 221
Mohamed Elhedi Hmidi, Ines Ben Salem, and Lilia El Amraoui
11 Digital Stabilizing and Control for Two-Wheeled Robot. . . . . . . . . . . . . . . 237
Bachir Nail, Abdellah Kouzou and Ahmed Hafaifa
12 Mobile Robot Path Planning Based on Optimized Fuzzy Logic
Controllers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 255
L. Cherroun, M. Boumehraz, and A. Kouzou
13 Design and Implementation of a Reactive Navigation System
for a Smart Robot Using Udoo Quad . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 285
Mohamed Njah and Ridha El-Hamdi
14 Area Coverage Algorithms for Networked Multi-robot Systems . . . . . 301
Lamia Iftekhar, H. M. Nafid Rahman, and Imran Rahman
15 Backstepping-Based Nonlinear RISE Feedback Control
for an Underactuated Quadrotor UAV Without Linear
Velocity Measurements. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 321
Jawhar Ghommam, Luis F. Luque-Vega, and Maarouf Saad
16 On the Tip–Path-Plane Flapping Angles Estimation
for Small–Scale Flybarless Helicopters at Near–Hover . . . . . . . . . . . . . . . . 343
Mohammad K. Al–Sharman and Mamoun F. Abdel–Hafez
Chapter 1
Cloud Robotic: Opening a New Road
to the Industry 4.0

Manal Aissam, Mohammed Benbrahim, and Mohammed Nabil Kabbaj

Abstract Cloud Robotics (CR) is a rising field of robotics rooted in cloud com-
puting, cloud storage, and other Internet technologies centered around the benefits
of converged infrastructure and shared services. It allows robots to benefit from
the powerful computational, storage, and communications resources of modern data
centers. In addition, it removes overheads for maintenance and updates, and reduces
dependence on custom middleware. This chapter reviews the concept of cloud-
enabled robotics with some of the most applications related to it. Also, it explores
the trends of robotic technology in Industry 4.0. Since the advent of Information
and Communication Technologies (ICT), economies around the world have ground
dramatically as companies can compete on a global scale. The Fourth Industrial
Revolution will be based on cyber-physical systems, the Internet of Things and
Internet of Services.

Keywords Cloud robotics · Cloud computing · Robotization · Industry 4.0

1.1 Introduction

The principle of Cloud Computing (CC) dates back to the 1990s, where the
foundations for the distribution of computing power as well as distributed data
storage in networked computer systems were laid. CC is one of the dominant
computing paradigm, Since past few years Robotics applications have also started to
build around these paradigms. With using cloud infrastructure in conjunction with
robotics, the development of the term CR from Kuffner (2010) has led to a wide
array of new research interests.

M. Aissam () · M. Benbrahim · M. N. Kabbaj

Faculty of Science Dhar El Marhraz, University of Sidi Mohamed Ben Abdellah, Fez, Morocco
e-mail: [email protected]; [email protected];
[email protected]

© Springer Nature Singapore Pte Ltd. 2019 1

N. Derbel et al. (eds.), New Developments and Advances
in Robot Control, Studies in Systems, Decision and Control 175,
https://doi.org/10.1007/978-981-13-2212-9_1
2 M. Aissam et al.

CR is a rapidly evolving field that allows robots to offload computation-intensive

and storage-intensive jobs into the cloud. Robots are limited in terms of computa-
tional capacity, memory and storage. Cloud provides unlimited computation power,
memory, storage and especially collaboration opportunity.
Robotic systems have brought significant economic and social impacts to human
lives over the past few decades. For example, industrial robots (especially robot
manipulators) have been widely deployed in factories to do tedious, repetitive,
or dangerous tasks, such as assembly, painting, packaging, and welding. These
preprogrammed robots have been very successful in industrial applications due
to their high endurance, speed, and precision in structured factory environments.
To enlarge the functional range of these robots or to deploy them in unstructured
environments, robotic technologies are integrated with network technologies to
foster the emergence of networked robotics.
According to current research linking robots to the Internet, CR and Automation
build on emerging research in cloud computing, machine learning, big data, open
source software, and major industry initiatives in the “Internet of Things”, “Smarter
Planet”, “Industrial Internet”, and “Industry 4.0”.
Industry 4.0 is the current trend of automation and data exchange in manufac-
turing technologies. It includes cyber-physical systems, the Internet of things and
Cloud Computing. So the CR will revolutionize the industry 4.0.
This survey is organized around the following essential concepts:
1. Cloud-Enabled Robotics
2. Cloud Computing
3. Industry 4.0

1.2 Literature Review

Networked Robotics has a long history but CR is a relatively recent term. “Robotics
on Cloud” is a concept that utilizes data collected from robots and sensor networks
by creating cloud storage to save the information and reusing it at the time of need.
It uses the help of the Internet to increase a robot’s capabilities by reducing on-board
computation and providing services on demand. Since robots will be connected to
the cloud, they can communicate with each other and exchange useful information
among each other. This off board computing is called Cloud Computing (see
Sect. 1.3.1). Massive parallel computation and real-time sharing of data resources
are the basic services provided by Cloud Computing.
The term “cloud-enabled robotics” was presented by James Kuffner for the first
time at the IEEE RAS International Conference on Humanoid Robotics in 2010
(Kuffner 2010). He was first to point out the potential of distributed networks
combined with robotics, primarily to enhance the robot agents limited capabilities.
CR presented a need of fast, reliable and ubiquitous network connections. The
Internet infrastructure has gone through a radical change in the past 10 years in
terms of bandwidth and Quality of Service, and is now able to enable to support the
cloud reliably.
1 Cloud Robotic: Opening a New Road to the Industry 4.0 3

Fig. 1.1 The PR2 research robot

At the 2011 Google I/O developer’s conference, Google announced a new

initiative called “cloud robotics” in conjunction with robot manufacturer Willow
Garage and introduced their theory and foreseen application of CR. Google has
developed an open source (free) operating system for robots, with the unsurprising
name “ROS” – or Robot Operating System. In other words, Google is trying to
create the MS-DOS (or MS Windows) of robotics (Kohler et al. 2011).
The PR2 was the first demonstration about how to make robots smarter and
more energy-efficient using these techniques. PR2 is a generic research platform
widely used for cloud applications (see Fig. 1.1, Kharel et al. 2014). The PR2 has
demonstrated various capabilities relying on the Robot Operating System (ROS)
(Quigley et al. 2009).
In a human-centered, real world environment, the robot can encounter a lot of
computation-intensive tasks. One of the most common problems is the recognition
and transfer of an unknown object. This is a basic human action, yet– even
nowadays–it is a complex and difficult task to be carried out by a home robot.
Willow Garage’s PR2 performed this task employing cloud-based technologies in
the following sequence (Jordan et al. 2013):
1. Capture data with robot: detecting objects with a 2D camera and 3D scanner.
2. The processing of data on cloud servers: analysing the 2D image.
3. The robot applies the processed data: after proceeding, the robot applies the
grasping movement.
4. Feedback to the server: Finally, the robot sends the results of the grasping
action, so that train the server’s database for further uses.
4 M. Aissam et al.

Fig. 1.2 The suckerbot

The process has online and offline elements, and it is based on both models.
The Cloud can also be used to facilitate open challenges and design competitions.
For example, the African Robotics Network with support from IEEE Robotics and
Automation Society hosted the “$10 Robot” Design Challenge in the summer of
2012. This open competition attracted 28 designs from around the world including a
winning entry from Thailand that modified a surplus Sony game controller, adapting
its embedded vibration motors to drive wheels and adding lollipops to the thumb
switches as inertial counterweights for contact sensing, which can be built from
surplus parts for US $8.96.
Suckerbot, designed by Tom Tilley of Thailand, a winner of the $10 Robot
Design Challenge (see Fig. 1.2, Kehoe et al. 2015).
Thus, several research groups continue to explore the use of cloud technologies
in robotic applications. For example, research groups Scientists in The Netherlands
have created a new World Wide Web for robots. It’s called RoboEarth (Waibel
et al. 2011), and the corporate-sponsored collaborators who developed it describe
it simply as “a World Wide Web for Robots.” After four years of research, the team
has arranged to host its first public demonstration of the technology: in 2014 in The
Netherlands, they set four robots loose in a hospital under staged conditions. These
robots became “collaboratively working together to help patients in a hospital”.
These robots use RoboEarth as a knowledge base, communication medium, and
computational resource to offload some of their heavy computation.
1 Cloud Robotic: Opening a New Road to the Industry 4.0 5

Fig. 1.3 RoboEarth’s architecture

RoboEarth is implemented based on a three-layered architecture (see Fig. 1.3,

Waibel et al. 2011). The core of this architecture is a server layer that holds the
RoboEarth database.
In the robot Cloud centre (Du et al. 2011) designed a framework following the
general Cloud computing paradigm, in order to address the current limitations in
capacity and versatility of robotic applications.
The work of Mouradian et al. (2014) concentrated on the IaaS aspects of
robotic applications as Cloud computing services. It suggested an architecture that
enabled cost efficiency through virtualization and dynamic task delegation to robots,
including the robots that might belong to other Clouds.
Gherardi et al. (2014) introduced a PaaS approach for configurable product
line based-on Cloud robotics applications. With this method, robotics developers
eliminated the need for low-level decision making by end users required to configure
the complex architecture of distributed systems on the robot and the Cloud.
Mohanarajah et al. (2014) presented the design and implementation of Rapyuta.
Rapyuta was an open source cloud robotics platform. It allowed robots to distribute
complex and heavy computation overhead over secured customizable computing
environments in the Cloud.
6 M. Aissam et al.

Using the Artificial neural network for the training of locations. The idea of
Ansari et al. (2012) was to establish the communication between the Cloud and
robot over a large environment and identify the location from the images sent by the
robot at the SaaS level.
Chen and Hu (2013) debated Internet of intelligent things and Robot-as-a-Service
(RaaS). The idea of attaining RaaS is through autonomous and intelligent mobile
physical services or robots to form a local pool of intelligent devices and that could
make local decisions without communications with the Cloud.
At the application level, Kamei et al. (2012) argue that an increase in industry
adoption of cloud connected networked robotics will encourage the development of
Cloud-related robotic applications. This in turn will create a need for new services
such as daily activity and accessible support. Some of the issues were identified
as future challenges include multirobot management, multi-area management, user
attribute management, and service coordination management.
Additionally, a Human-Robot Cloud (HRC) element was suggested, as part of the
Cloud, as an extension to its support of the physical and cognitive roles of humans
in Cloud computing which were neither expected to be experts nor to be engaged
with the Cloud full-time (Mavridis et al. 2012).
From a machine learning perception, Ren (2011) described the concept of an
advanced intelligence machine, which is a device that used both natural and artificial
intelligence and is capable of affective recognition and generation of affective
speech and behavior.
Nakagawa et al. (2012) proposed a distributed service framework using Robot
Service Network Protocol (RSNP) to integrate various devices including robots with
internet services.
A robot developer should study many languages for development of firmware,
to solve this problem, Jang and Kim (2013) developed a script language-based
template for the source code generation and exchange.
The applications that will emerge for cloud robots are of many kinds; some are
emerging now, others are at an early stage of development.

1.3 From Cloud Computing to Cloud Robotics

1.3.1 Cloud Computing

According to the National Institute of Standards and Technology (NIST), Cloud

computing is a model for enabling-ubiquitous, convenient, on-demand network
access to a shared pool of configurable computing resources (e.g., networks, servers,
storage, applications, and services) that can be rapidly provisioned and released with
minimal management effort or service provider interaction (Fig. 1.4).
This cloud model is composed of five essential characteristics, three service
models, and four deployment models (Fig. 1.5).
1 Cloud Robotic: Opening a New Road to the Industry 4.0 7

Fig. 1.4 Cloud computing

Fig. 1.5 Cloud computing infrastructure

1.3.1.1 Cloud Computing Deployment Models

Cloud computing services can be private, public or hybrid. Private cloud services are
delivered from a business’ data center to internal users. This model offers versatility
and convenience, while preserving the management, control and security common
to local data centers. Internal users may or may not be billed for services through
IT charge back. In the public cloud model, a third-party provider delivers the cloud
service over the Internet. Public cloud services are sold on demand, typically by
the minute or hour. Customers only pay for the CPU cycles, storage or bandwidth
they consume. Leading public cloud providers include Amazon Web Services
(AWS), Microsoft Azure, IBM Soft Layer and Google Compute Engine. Hybrid
cloud is a combination of public cloud services and on-premises private cloud with
orchestration and automation between the two. Companies can run mission-critical
workloads or sensitive applications on the private cloud while using the public cloud
for bursting workloads that must scale on demand. The goal of hybrid cloud is to
create a unified, automated, scalable environment that takes advantage of all that a
public cloud infrastructure can provide while still maintaining control over mission-
critical data.
8 M. Aissam et al.

Fig. 1.6 Cloud computing

services

1.3.1.2 Cloud Computing Service Categories

Cloud computing consists of three fundamental models as Software as a Service

(SaaS), Platform as a Service (PaaS) and Infrastructure as a Service (IaaS) (Koken
2015) as shown in Fig. 1.6:
• SaaS applications are served over the Internet, thus eliminating the need to install
and run the application on the users system (Mathur and Nishchal 2010). They
are managed from a centralized location and accessed remotely by a web browser
or a mobile client. Google Apps is the most widely used SaaS application suit.
• PaaS refers to a computing platform served by cloud infrastructure. PaaS offers
developers to get a hold of all the systems and environments required for the
life cycle of software, be it developing, testing, deploying and hosting of web
applications. Some examples are Amazon Web Services (AWS) and Microsoft’s
Azure (Rimal et al. 2009).
• IaaS provides the required infrastructure as a service. The client doesn’t need
purchase the required servers, data center or the network resources. The essence
of IaaS model is a pay-as-you-go financial model. Amazon and Microsoft are
also IaaS providers.

1.3.1.3 Cloud Architecture

When talking about a cloud computing system, it’s helpful to divide it into two
sections: the front end and the back end. They connect to each other through a
network, usually the Internet. The front end is the side the computer user, or client,
sees. The back end is the “Cloud” section of the system. The front end includes
the client’s computer (or computer network) and the application required to access
the cloud computing system. Not all cloud computing systems have the same user
interface. Services like Web-based e-mail programs leverage existing Web browsers
like Internet Explorer or Firefox. Other systems have unique applications that
provide network access to clients (see Fig. 1.7).
1 Cloud Robotic: Opening a New Road to the Industry 4.0 9

Fig. 1.7 Cloud computing architecture

Fig. 1.8 Layers of a cloud architecture

A central server administers the system, monitoring traffic and client demands
to ensure everything runs smoothly. It follows a set of rules called protocols and
uses a special kind of software called middleware. Middleware allows networked
computers to communicate with each other (see Fig. 1.8). A Cloud architecture can
be divided into different layers (Bakshi 2011):
• WAN edge: are edge routers act as DC/WAN;
• Core/Agg: are switches that provide OSI (Open Systems Interconnection) layer
2 between compute nodes and access points;
• Network Services: Provides intelligence to the network via capabilities such as
security firewalls;
• Access: This is where the network can be virtualized via VLANs and new policies
and an access control list can be applied;
• Compute: runs the hypervisor to support virtual machines;
• Storage: consists of storage elements that are Storage Disk arrays, Storage Tape,
Fiber Channel and Ether switches;
• Virtual Network Access: It acts as a virtual access layer for MVs.
10 M. Aissam et al.

1.3.1.4 Service Level Agreement

A Service-Level Agreement (SLA) is a contract between a service provider and its

internal or external customers that documents what services the provider will furnish
and defines the performance standards the provider is obligated to meet. SLAs
establish customer expectations with regard to the service provider’s performance
and quality in a number of ways. Some metrics that SLAs may specify include:
• Availability and uptime, the percentage of the time services will be available.
• Specific performance benchmarks to which actual performance will be periodi-
cally compared.
• Application response time.
• The schedule for notification in advance of network changes that may affect
users.
• Help desk response time for various classes of problems.
• Usage statistics that will be provided.
An SLA may specify availability, performance and other parameters for different
types of customer infrastructure, internal networks, servers and infrastructure
components such as uninterruptable power supplies, for example.
In addition to establishing performance metrics, an SLA may include a plan
for addressing downtime and documentation for how the service provider will
compensate customers in the event of a contract breach. Service credits are a
typical remedy. Here, the service provider issues credits to the customer based
on an SLA-specified calculation. Service providers, for example, might provide
credits commensurate with the amount of time it exceeded the SLA’s performance
guarantee.
The SLA will also include a section detailing exclusions, that is, situations in
which an SLA’s guarantees and penalties for failing to meet them don’t apply. The
list might include events such as natural disasters or terrorist acts. This section is
sometimes referred to as a force majeure clause, which aims to excuse the service
provider from events beyond its control.

1.3.2 Cloud Robotics

1.3.2.1 Definition

Cloud Robotics (CR) is an emerging field within robotics, currently covering various
application domains and robot network paradigms. CR was born from the merger
of cloud technologies and service robotics, which was preceded by a change in
paradigm in both domains. Cloud technology-based computing–or simply Cloud
Computing–is one of the most dynamically growing areas of Info-Communication
Technologies (ICT) (Jordan et al. 2013).
1 Cloud Robotic: Opening a New Road to the Industry 4.0 11

CR allows robots to take advantage of the rapid increase in data transfer rates to
offload tasks without hard real time requirements. This is of particular interest for
mobile robots, where on-board computation entails additional power requirements
which may reduce operating duration and constrain robot mobility as well as
increase costs.

1.3.2.2 Robots Classification

Robots have some constraints in terms of computational capacity, memory and

storage. CR helps them to overcome these challenges. Opportunity to use cloud
allows cost effective robots to be produced. Robots can be classified as traditional
robots and cloud-enabled robots. Cloud technologies not only empower robots but
also it allows them to network each other regardless of distance. Cloud-enabled
robots are divided into two categories as standalone robots and networked robots.
Classification of robots is shown in Fig. 1.9 (Koken 2015).
Robots can do a wide variety of works such as grasping, identifying objects,
SLAM (Simultaneous Localization and Mapping) (Durrant-Whyte and Bailey
2006), monitoring, networking and some other actuating works. Robots can grasp
formerly known objects easily. They can also grasp novel objects with the help of
cloud. In Hu et al. (2012), a study about grasp planning in the presence of shape
uncertainty and how cloud computing can facilitate parallel Monte Carlo sampling
is presented. Standalone robots can benefit from cloud in terms of computation
power, storage capacity and memory. However, networked robots can make net-
works, share their information through cloud and can perform collaborative works.
CR infrastructure with standalone robots and networked robots is presented in
Fig. 1.10.

Fig. 1.9 Classification of robots

12 M. Aissam et al.

Fig. 1.10 Standalone and networked robots

1.3.2.3 Benefits and Risks

For us humans, with our non-upgradeable, offline brains, the possibility of acquiring
new skills by connecting our heads to a computer network is still science fiction. It
is a reality for robots.
Cloud Robotics can allow the robot to access vast amounts of processing
power, data and offload compute-intensive tasks like image processing and voice
recognition and even download new skills instantly, Matrix-style.
Future robotic applications will benefit from cloud robotics, which provides the
following advantages over traditional networked robots (Hu et al. 2012).
• Ability to offload computation-intensive tasks to the cloud. The robots only have
to keep necessary sensors, actuators, and basic processing power to enable real-
time actions (e.g., real-time control). The battery life is extended, and the robotic
platform becomes lighter and less expensive with easier to maintain hardware.
The maintenance of software onboard with the robots also becomes simpler,
with less need for regular updates. As the cloud hardware can be upgraded
independently from the robotic network, the operational life and usefulness of
the robotic network can be easily extended.
• Access to vast amounts of data. The robots can acquire information and
knowledge to execute tasks through databases in the cloud. They do not have
to deal with the creation and maintenance of such data.
• Access to shared knowledge and new skills. The cloud provides a medium for
the robots to share information and learn new skills and knowledge from each
other. The cloud can host a database or library of skills or behaviors that map
to different task requirements and environmental complexities. The RoboEarth
project (Waibel et al. 2011) is trying to turn this into a reality.
Any old robot full of dust, stored in the closets of the research labs of the world
can be reused again by using the hardware and software infrastructure of the cloud.
1 Cloud Robotic: Opening a New Road to the Industry 4.0 13

CR has certainly benefited many enterprises by reducing costs and allowing them
to concentrate on their core business competence rather than IT and infrastructure
issues. But, there are still distinct disadvantages of CR; especially relating to
security. Here are the top CR risks we identified:
• Environmental security: The concentration of computing resources and users
in a cloud computing environment also represents a concentration of security
threats. Because of their size and significance, cloud environments are often
targeted by virtual machines and bot malware, brute force attacks, and other
attacks.
• Data privacy and security: Hosting confidential data with cloud service
providers involves the transfer of a considerable amount of an organization’s
control over data security to the provider. For example, every cloud contains a
huge information from the clients include personal data. If a household robot is
hacked, users could have risk of their personal privacy and security, like house
layout, life snapshot, home-view, etc. It may be accessed and leaked to the world
around by criminals. Another problem is once a robot is hacked and controlled
by someone else, which may put the user in danger.
• Ethical problems: Some ethics of robotics, especially for Cloud based robotics
must be considered. Since a robot is connected via networks, it has risk to be
accessed by other people. If a robot is out of control and carries out illegal
activities, which should be responsible for it.

1.3.3 Applications by Domain

Unquestionably, CR is a fascinating research topic, yet it has to identify its target

applications and finds its particular domains. It is believed that CR is a way to
introduce cognition into the domain of robotics, which is envisioned and strongly
supported by the European Union through the FP calls.
The CR is used in several fields, The most important applications are in the
following fields:
• Industrial robots: As highlighted by the Germany Industry 4.0 (MacDougall
2014) Plan “Industry is on the threshold of the fourth industrial revolution.
Driven by the Internet, the real and virtual worlds are growing closer and closer
together to form the Internet of Things. Industrial production of the future will be
characterized by the strong individualization of products under the conditions of
highly flexible (large series) production, the extensive integration of customers
and business partners in business and value-added processes, and the linking of
production and high-quality services leading to so-called hybrid products.” In
manufacturing, such cloud based robot systems could learn to handle tasks such
as threading wires or cables, or aligning gaskets from professional knowledge
base. A group of robots can share information for some collaborative tasks.
Even more, a consumer is able to order customized product to manufacturing
14 M. Aissam et al.

robots directly with online order system. Another potential paradigm is shopping-
delivery robot system- once an order is placed; a warehouse robot dispatches the
item to an autonomous car or autonomous drone to deliver it to its recipient.
• Cloud medical robots: a medical cloud (also called a healthcare cluster) consists
of various services such as a disease archive, electronic medical records, a patient
health management system, practice services, analytics services, clinic solutions,
expert systems, etc. Health care system is a very complicated system and involved
in a variety of related technologies (Ma et al. 2015). A robot can connect to
the cloud to provide clinical service to patients, as well as deliver assistance
to doctors (e.g. a co-surgery robot). Moreover, it also provides a collaboration
service by sharing information between doctors and care givers about clinical
treatment. Assistive robots: A domestic robot can be employed for healthcare
and life monitoring for elderly people. The system collects the health status of
users and exchange information with cloud expert system or doctors to facilitate
elderly people’s life, especially for those with chronic diseases. For example,
the robots are able to provide support to prevent the elderly from falling down,
emergency healthy support such as heart disease, blooding disease. Care givers
of elderly people can also get notification when in emergency from the robot
through network.
• Autonomous mobile robots: self-driving cars are cloud robots (Boeglin 2015).
The cars use the network to access Google’s enormous database of maps and
satellite and environment model (like Street view) and combines it with stream-
ing data from GPS, cameras, and 3D sensors to monitor its own position within
centimeters, and with past and current traffic patterns to avoid collisions. Each
car can learn something about environments, roads, or driving, or conditions, and
it sends the information to the Google cloud, where it can be used to improve the
performance of other cars.

1.3.4 Case Study Example

To apply CR functions in a real word scenario, the team of Sherpa project (Marconi
et al. 2012) has developed a robotics platform to support search and rescue
activities in hostile environments such as the alpine project. In this approach,
to rescue survivor victims after avalanches, a human operator cooperates with
a heterogeneous robotic system. The robotic team is principally composed of
unmanned aerial vehicles (UAVs) with different characteristics and equipped with
different types of sensors in order to retrieve information from the rescue scene and
assist the rescuer during a mission. The Sherpa team is composed of:
• A human rescuer: who is a professional of the rescuing and surveillance
mission.
• Small scale rotary-wing Unmanned Aerial Vehicles: share with rescue mem-
bers visuals information of victims detected with cameras.
1 Cloud Robotic: Opening a New Road to the Industry 4.0 15

• A ground rover: is used as a transportation module for the rescuer equipment

and as a hardware station with computational capabilities.
• A fixed-wing UAV and an unmanned helicopter: are used for constructing
a 3D map of the rescuing area, as communication hub between the platforms
in presence of critical terrain morphologies, for patrolling large areas not
necessarily confined in the neighborhood of the rescuer.
The unmanned aerial vehicles (UAVs), have been widely used in several appli-
cations such as industrial building inspection and surveillance (Cacace et al. 2015),
photogrammetry (Colomina and Molina 2014), remote sensing (Mitchell et al. 2012)
and many others. In the same philosophy as Sherpa, similar works (Cacace et al.
2016) demonstrate the ability of a human rescuer to orchestrate a heterogeneous
multi-robot system.

1.4 Cloud Robotics and Industry 4.0

1.4.1 Industry 4.0

The Industrial Revolution was a time in the eighteenth century when many important
inventions were made. Many of these inventions made work easier and cheaper. As
these inventions created new manufacturing and industry, many people also moved
away from farms into cities. It was a time of very rapid change in the world.
Over the course of history, mankind has perfected its industry by not only relying
on technical evolution but also by reinventing it as new resources have created new
technical means. Therefore, industry has benefited from qualitative advancements
which have sometimes been so ingrained in a certain time period and have had such
an overwhelming impact that we have dubbed them “revolutions”.Here is a quick
look back in time at these first three industrial revolutions to define the contours of
a fourth revolution which is taking shape right before our very eyes (Fig. 1.11).

Fig. 1.11 Industrial revolutions timeline

16 M. Aissam et al.

Today we stand on the cups of a fourth industrial revolution; one which promises
to marry the worlds of production and network connectivity in an “Internet of
Things” which makes “Industry 4.0” a reality.
The term “Industry 4.0” originates from a project in the high-tech strategy of
the German government, which promotes the computerization of manufacturing.
Exactly, this term was revived in 2011 at the Hannover Fair. In October 2012 the
Working Group on Industry 4.0 presented a set of Industry 4.0 implementation
recommendations to the German federal government.
Therefore, some have compared Industry 4.0 with the Fourth Industrial Revo-
lution. However, the latter refers to a systemic transformation that includes impact
on civil society, governance structures, and human identity in addition to solely
economic/manufacturing ramifications. The first industrial revolution mobilized the
mechanization of production using water and steam power; the second industrial
revolution then introduced mass production with the help of electric power, followed
by the digital revolution and the use of electronics and IT to further automate
production. The term “fourth industrial revolution” has been applied to significant
technological developments several times over the last 75 years, and is up for
academic debate. Industry 4.0, on the other hand, focuses on manufacturing
specifically in the current context, and thus is separate from the fourth industrial
revolution in terms of scope.
Industry 4.0 emphasizes the idea of consistent digitization and linking of
all productive units in an economy. There are several technological areas that
underpin Industry 4.0, which are horizontal and vertical system integration, the
internet of things, cybersecurity, the cloud, big data analytics, simulation, additive
manufacturing (3d printing), augmented reality, and robot (Rüssmann et al. 2015).
The figure below shows the technologies related to Industry 4.0 (Fig. 1.12).

1.4.2 Robot in Industry 4.0

Manufacturers in many industries have long used robots to tackle complex assign-
ments, but robots are evolving for even greater utility. They are becoming more
autonomous, flexible, and cooperative. Eventually, they will interact with one
another and work safely side by side with humans and learn from them. These
robots will cost less and have a greater range of capabilities than those used
in manufacturing today. For example, Kuka, a European manufacturer of robotic
equipment, offers autonomous robots that interact with one another. These robots
are interconnected so that they can work together and automatically adjust their
actions to fit the next unfinished product in line. High-end sensors and control units
enable close collaboration with humans. Similarly, industrial-robot supplier ABB is
launching a two-armed robot called YuMi that is specifically designed to assemble
products (such as consumer electronics) alongside humans. Two padded arms and
computer vision allow for safe interaction and parts recognition.
1 Cloud Robotic: Opening a New Road to the Industry 4.0 17

Fig. 1.12 Technologies related to industry 4.0

Industrial robots, which are one of the key drivers in Industry 4.0, have pro-
gressed greatly since the last 10 years of the twentieth century. They are becoming
more productive, flexible, versatile, safer, and collaborative and are thus creating
an unprecedented level of value in the entire ecosystem. Smart factories, which
will be at the heart of Industry 4.0, will embark information and communication
technology for an evolution in the supply chain and production line that brings a
much higher level of both automation and digitization. It means machines using self-
optimization, self-configuration and even artificial intelligence to complete complex
tasks in order to deliver vastly superior cost efficiencies and better quality goods or
services (Fig. 1.13).
With all this evolution, various companies have focused on the development of
robots, such as Fanuc, Universal Robots, Gomtec, . . . (Bahrin et al. 2016).
Also, we can notice that various organizations and key players have substan-
tial research departments which support product development and automation of
manufacturing processes, while also marketing their know-how to other industrial
enterprises or providing platforms for ‘smart’ services for third parties. This can
further stimulate the implementation of Industry 4.0. And in this case, there are
many organizations that are interested in the robotic industry of Industry 4.0 (Bahrin
et al. 2016).
18 M. Aissam et al.

Fig. 1.13 Robotics in the

industry 4.0

1.4.3 Expectations Regarding Industry 4.0

Industry 4.0 raised high expectations, but not all have been met. Nonetheless, a
handful of manufacturers have discovered numerous applications for Industry 4.0
and are reaping the benefits. Implementing Industry 4.0 is a process that could
take years, and more applications will develop as technologies mature further. It
is imperative that manufacturers in all countries start now with a set of concrete
applications. This will build the organizational and technical muscle to tackle more
ambitious projects in the future, such as the complete integration of data throughout
the product life cycle. Here are some expectations regarding Industry 4.0:
• More Flexibility and Adaptability: Fixed structures will be replaced by
adaptable networks and Local Intelligence will help to handle complexity.
• More Modularity and Autonomy: Cyber Physical Products will be introduced
in all levels of Automation also Granularity will adjust according to market
needs.
• Highest Productivity: Resource optimization will be done in the network based
swarm optimization, Plug & Produce features will bring down development cost
significantly and Management will do more optimization than fire fighting.
• New Business Models: “app-store” and “cloud” will provide the new knowledge
management, (Big Data) shall become a new driver and Open Source Approaches
shall provide new opportunities.

1.5 Conclusion

Overall, the exciting field of cloud robotics has only just begun to unravel paving
the way to new and bold experiments within automation and control. Having the
capacity to exploit such large computing resources and allocate them to robots
is an advancement. Using a cloud infrastructure to handle an ever growing and
1 Cloud Robotic: Opening a New Road to the Industry 4.0 19

increasingly complicated set of algorithms, utilized in vision acquisition systems

and formation control, is a more logical approach. In this chapter, we have also
shown an approach to introduce robotic services in the cloud and in Industry 4.0,
linking a real-life factory with virtual reality, will play an increasingly important
role in global manufacturing. Human-robot collaboration will have a breakthrough
in this period.

References

Ansari, F., Pal, J., Shukla, J., Nandi, G., & Chakraborty, P. (2012). A cloud based robot localization
technique. In Internaional Conference on Contemporary Computing, Noida, India (pp. 347–
357).
Bahrin, M. A. K., Othman, F., Azli, N. H. N., & Talib, M. F. (2016). Industry 4.0: A review on
industrial automation and robotic. Jurnal Teknologi, 78(6–13), 137–143.
Bakshi, K. (2011). Considerations for cloud data centers: Framework, architecture and adoption.
In IEEE Aerospace Conference, Big Sky, MT, USA (pp. 1–7).
Boeglin, J. (2015). The costs of self-driving cars: Reconciling freedom and privacy with tort
liability in autonomous vehicle regulation. Yale Journal of Law and Technology, 17, 172–203.
Cacace, J., Finzi, A., Lippiello, V., Loianno, G., & Sanzone, D. (2015). Aerial service vehicles for
industrial inspection: Task decomposition and plan execution. Applied Intelligence, 42, 49–62.
Cacace, J., Finzi, A., Lippiello, V., Furci, M., Mimmo, N., & Marconi, L. (2016). A control
architecture for multiple drones operated via multimodal interaction in search & rescue
mission. In IEEE International Symposium on Safety, Security, and Rescue Robotics, Lausanne,
Switzerland (pp. 233–239).
Chen, Y., & Hu, H. (2013). Internet of intelligent things and robot as a service. Simulation
Modelling Practice and Theory, 34, 159–171.
Colomina, I., & Molina, P. (2014). Unmanned aerial systems for photogrammetry and remote
sensing: A review. ISPRS Journal of Photogrammetry and Remote Sensing, 92, 79–97.
Du, Z., Yang, W., Chen, Y., Sun, X., Wang, X., & Xu, C. (2011). Design of a robot cloud
center. In Tenth International Symposium on Autonomous Decentralized Systems (ISADS),
Tokyo/Hiroshima, Japan (pp. 269–275).
Durrant-Whyte, H., & Bailey, T. (2006). Simultaneous localization and mapping: Part I. IEEE
Robotics and Automation Magazine, 13, 99–110.
Gherardi, L., Hunziker, D., & Mohanarajah, G. (2014). A software product line approach for
configuring cloud robotics applications. In IEEE 7th International Conference on Cloud
Computing (CLOUD), Anchorage, AK, USA (pp. 745–752).
Hu, G., Tay, W., & Wen, Y. (2012). Cloud robotics: Architecture, challenges and applications.
IEEE Network, 26, 21–28.
Jang, W.-S., & Kim, R. (2013). Template design of automatic source code generation based on
script language used in cloud robot compiling environment. CST, 27, 184–185.
Jordan, S., Haidegger, T., Kovs, L., Felde, I., & Rudas, I. (2013). The rising prospects of cloud
robotic applications. In IEEE 9th International Conference on Computational Cybernetics,
Tihany, Hungary (pp. 327–332).
Kamei, K., Nishio, S., Hagita, N., & Sato, M. (2012). Cloud networked robotics. IEEE Network,
26, 28–34.
Kehoe, B., Patil, S., Abbeel, P., & Goldberg, K. (2015). A survey of research on cloud robotics and
automation. IEEE Transactions on Automation Science and Engineering, 12, 398–409.
Kharel, A., Bhutia, D., Rai, S., & Ningombam, D. (2014). Cloud robotics using ROS. International
Journal of Computer Applications (pp. 18–21).
20 M. Aissam et al.

Kohler, D., Hickman, R., Conley, K., & Gerkey, B. (2011). Cloud robotics. In Google I/O 2011
Developer Conference, Mountain View, CA, USA.
Koken, B. (2015). Cloud robotics platforms. Interdisciplinary Description of Complex Systems, 13,
26–33.
Kuffner, J. J. (2010). Cloud-enabled robots. In IEEE-RAS International Conference on Humanoid
Robots, Nashville, TN, USA (pp. 1–9).
Ma, Y., Zhang, Y., Wan, J., Zhang, D., & Pan, N. (2015). Robot and cloud-assisted multi-modal
healthcare system. New York: Springer.
MacDougall, W. (2014). Industrie 4.0: Smart manufacturing for the future. Berlin: Germany Trade
and Invest.
Marconi, L., Melchiorri, C., Beetz, M., Pangercic, D., Siegwart, R., Leutenegger, S., Carloni, R.,
Stramigioli, S., Bruyninckx, H., Doherty, P., Kleiner, A., Lippiello, V., Finzi, A., Siciliano,
B., Sala, A., & Tomatis, N. (2012). The sherpa project: Smart collaboration between humans
and ground-aerial robots for improving rescuing activities in alpine environments. In IEEE
International Symposium on Safety, Security, and Rescue Robotics, College Station, TN, USA.
Mathur, P., & Nishchal, N. (2010). Cloud computing: New challenge to the entire computer
industry. In First International Conference on Parallel, Distributed and Grid Computing, Solan,
India (pp. 223–228).
Mavridis, N., Bourlai, T., & Ognibene, D. (2012). The human-robot cloud: Situated collective
intelligence on demand. In IEEE International Conference on Cyber Technology in Automation,
Control, and Intelligent Systems (CYBER), Bangkok, Thailand (pp. 360–365).
Mitchell, J. J., Glenn, N. F., Anderson, M. O., Hruska, R. C., Halford, A., Baun, C., & Nydegger, N.
(2012). Unmanned aerial vehicle (UAV) hyperspectral remote sensing for dryland vegetation
monitoring. In 4th Workshop on Hyperspectral Image and Signal Processing, Shanghai, China.
Mohanarajah, G., Hunziker, D., D’Andrea, R., & Waibel, M. (2014). Rapyuta: A cloud robotics
platform. IEEE Transactions on Automation Science and Engineering, 12, 481–493.
Mouradian, C., Errounda, F., Belqasmi, F., & Glitho, R. (2014). An infrastructure for robotic
applications as cloud computing services. In IEEE World Forum on Internet of Things, Seoul,
South Korea (pp. 377–382).
Nakagawa, S., Igarashi, N., Tsuchiya, Y., Narita, M., & Kato, Y. (2012). An implementation of a
distributed service framework for cloud-based robot services. In 38th Annual Conference on
IEEE Industrial Electronics Society, Montreal, QC, Canada (pp. 4148–4153).
Quigley, M., Gerkey, B., Conley, K., Faust, J., Foote, T., Leibs, J., Berger, E., Wheeler, R., & Ng,
A. (2009). ROS: An open-source robot operating system. Proceedings of ICRA Workshop on
Open Source Software, Kobe, Japan.
Ren, F. (2011). Robotics cloud and robotics school. In 7th International Conference on Natural
Language Processing and Knowledge Engineering (NLP-KE), Tokushima, Japan (pp. 1–8).
Rimal, B., Eunmi, C., & Lumb, I. (2009). A taxonomy and survey of cloud computing systems.
Fifth International Joint Conference on INC, IMS and IDC, Seoul, South Korea (pp. 44–51).
Rüssmann, M., Lorenz, M., Gerbert, P., Waldner, M., Justus, J., Engel, P., & Harnisch, M. (2015).
Industry 4.0: The future of productivity and growth in manufacturing industries (Vol. 9). Boston
Consulting Group.
Waibel, M., Beetz, M., Civera, J., D’Andrea, R., Elfring, J., Galvez-Lopez, D., Haussermann, K.,
Janssen, R., Montiel, J., Perzylo, A., Schiessle, B., Tenorth, M., Zweigle, O., & Molengraft,
R. V. D. (2011). Roboearth: A world wide web for robots. IEEE Robotics and Automation
Magazine, 18, 69–82.
Chapter 2
4 × 2D Visual Servoing Approach
for Advanced Robot Applications

Mohamad Bdiwi, Jozef Suchý, and Matthias Putz

Abstract Actually, vision information can be used as sensory input in open-loop as

well as in the closed-loop. However, the visual servoing approach is performed only
inside the closed-loop robot control, because in the open-loop the vision sensor will
represent the initial extraction of the features to generate directly the robot motion
sequence and these features and motion could be off-line generated. On the contrast,
closed-loop robot system uses the vision as real time sensor and it consists of two
phases: tracking and control. Tracking provides a continuous estimation and update
of features during the robot/object motion. Based on this information, a real time
control loop will be generated. The main contribution of this work is a proposed
visual servoing approach which will benefit from the images which are obtained by
Kinect camera (RGB-D camera). The proposed visual servoing approach is called
4 × 2D visual servoing which combines the correspondent color and depth images
to build two new images. Using these 4 images the control error signals will be
calculated in order to track the objects. Firstly, this chapter will present all types
of visual servoing then it will introduce the 4 × 2D visual servoing approach and
the visible side coordinate system, after that it will illustrate the concept of the
proposed approach and how the error signal will be calculated. In addition to that,
this approach proposes a coordinate system which is called visible side coordinate
system.

Keywords Visual servoing · Robot vision · RGB-D camera systems

M. Bdiwi () · M. Putz

Fraunhofer IWU, Chemnitz, Germany
e-mail: [email protected]
J. Suchý
Technical University of Chemnitz, Chemnitz, Germany

© Springer Nature Singapore Pte Ltd. 2019 21

N. Derbel et al. (eds.), New Developments and Advances
in Robot Control, Studies in Systems, Decision and Control 175,
https://doi.org/10.1007/978-981-13-2212-9_2
22 M. Bdiwi et al.

2.1 Introduction

The first who used vision feedback was Y. Shirai (Shirai and Inoue 1973), he has
described how a visual feedback loop can be used to correct the position of a robot
to increase task accuracy. Visual servoing term has been first introduced by Hill and
Park (1979). Since that time considerable efforts have been devoted to the visual
control of robot manipulators, e.g. Weiss et al. (1985, 1987). Numerous papers,
e.g. Hutchinson et al. (1996), Hashimoto (2003) and Corke and Hutchinson (2000),
have discussed the principles of the arts of visual servoing for robot manipulators in
details. Figure 2.1 illustrates the main categories of the visual servoing approaches
depending on different issues. These issues are: (1) Controller architecture, (2)
Camera configuration, (3) Observing end-effector, (4) Control law error.
Based on whether the control is applied to the joints directly or as a position
command to a robot controller, visual servoing is divided into two types; (1)
Direct servoing, (2) Dynamic look-and-move. Camera configuration includes two
categories which are: (1) Camera type and (2) Camera location. Regarding the
observing the end-effector, the visual servo control can be also classified whether if
the end-effector is observed or not: (1) End-effector close loop and (2) End-effector
open loop. Chesi and Hashimoto (2002) has presented different types of visual
servoing and it has compared the stability of the visual servoing in respect to the
camera configuration eye-in-hand and eye-to-hand, where Flandin et al. (2000) has

Fig. 2.1 Categories of visual servoing

2 4 × 2D Visual Servoing Approach for Advanced Robot Applications 23

presented a cooperation approach which integrates a fixed camera with one mounted
on the robot. Next section will illustrate the different types of visual servoing based
on control law error in details.

2.2 Related Work and Motivation

The main types of visual servoing which depending on the error used to compute
the control law are: position-based visual servo system, image-based visual servo
system and 2D 12 visual servoing. In the position based control system, the error will
be computed in the Cartesian space. In other words, the features are extracted from
the image and they are used to estimate the pose of the target object with respect to
the camera coordinate system. In general, position-based visual servoing techniques
can be classified into two groups: model-based and model-free visual servoing.
Model-based 3D visual servoing: In this case, 3D model of the target object
is available. The desired camera pose with respect to the object frame could be
estimated from the error between the 3D model of the target object and the current
image features. Figure 2.2 presents scheme of model-based visual servoing. The
extracted features of the object will be compared with its model in order to estimate
the object’s pose respect to the camera coordinate system.
Model-free 3D visual servoing: In this mode, the 3D structure of the object or the
environment is completely unknown. In other words, there is no 3D model of the
target object (Model-free). Model-free 3D visual servoing is based on teaching-by-
showing step. In the teaching phase the robot will be driven to the desired position in
order to store the corresponding reference image features, after that when the object
or the robot have been moved, the control error will drive the robot to the desired
position from the current position using the displacement between the reference
image features and current image features.

Fig. 2.2 Image based visual servoing

24 M. Bdiwi et al.

3D information could be reconstructed by fusing many 2D features, such as

brightness, texture, color, size, and transparentness of the object. Furthermore
photographic measurement or stereo camera could be used to reconstruct the 3D
information. Photographic measurement uses the object shape and size to estimate
the object position and orientation.
The image based visual servoing is a special case from model-free control
because it does not need the knowledge of the 3D model of the target object and
the control error will be formulated directly in 2D image space, so it will be called
image error function. In the image-based visual servoing, the interaction matrix
(image Jacobian matrix) will be used. This matrix describes how image feature
parameters will change with respect to the change of the object or manipulator
pose. This approach works with single camera and there is no need to the stereo
computations.
In general, both methods position-based and image-based visual servoing have
some drawbacks, e.g. in the position-based approach there is no performed control
in the image space, which could imply that the target object may get out of the
camera view during the servoing. Furthermore, a model of the target or teaching
phase is needed to estimate the pose of the camera with respect to the target. In
the image-based approach the depth estimation is needed to compute the interaction
matrix and to design the control law. In addition to that, the matching between the
desired and the current position will be insured only when the current position is in
a neighborhood of the desired position.
A new approach was suggested by Malis et al. (1999) in order to reduce these
drawbacks. This approach is called 2 12 D visual servoing.
2 12 D visual servoing (Malis et al. 1999) is based on the estimation of the partial
camera displacement from the current to the desired camera poses at each iteration
of the control law. The extracted data from the partial camera displacement allow
designing a decoupled control law which controls the six camera degree of freedom.
In other words, 2 12 D visual servoing can be used to decouple the translational and the
rotational control loop. Since the position and orientation errors is controlled by two
independent control loops, there is possibility to combine the advantages of IBVS
and PBVS by selecting adequate visual features which part of them are defined in
2D and the other part in the 3D. Hence, the control scheme will always converge
and avoid the singularities. Malis (2002) and Kragic and Christensen (2002) have
described the advantages and disadvantages of IBVS, PBVS and 2 12 D approaches
with comparison between them. Goncalves et al. (2002) has compared the behavior
of 2D and 2 12 D visual servoing in achieving a desired goal by planar robot. The
overview of 2 12 D visual servoing structure is illustrated in Fig. 2.3, where ue and
pe are estimated rotation and position control vector, ud and pd are desired rotation
and position control vector. In this thesis a new visual servoing approach will be
introduced in Chap. 3 which is called 4 × 2D visual servoing approach aimed to
perform visual servoing using RGBD camera.
2 4 × 2D Visual Servoing Approach for Advanced Robot Applications 25

Fig. 2.3 2 12 D visual servoing

2.3 Concept of 4 × 2D Visual Servoing Approach

All the previous visual servoing approaches are proposed with consideration of
using 2D camera, e.g.: (1) A single 2D camera will be used in case of IBVS.
(2) In the case of PBVS stereo 2D camera system or single 2D camera with
multiple viewpoints or multiple focusing factors could be used. In the last years the
RGB-D cameras (such as Kinect camera) are spread widely everywhere. RGB-D
cameras deliver depth and color images simultaneously and give us the opportunity
to use this information in the visual servoing tasks. To exploit all the capabilities
and advantages of the RGB-D cameras, this work will suggest a new approach
of visual servoing. This approach will be called 4 × 2D visual servoing. 4 × 2D
visual servoing combines the correspondence of color and depth images to build
(u, w) and (w, v) images. By using these 4 images [RGB, depth, (u, w) and (w, v)]
the system can calculate the control error signals at every subspace directly from
the obtained images as is shown in Fig. 2.4. In PBVS approach many methods
are used to estimate the 3D pose of the object from 2D images such as stereo
vision, photometric stereo, shape from shadow etc. These approaches have many
problems and disadvantages, for example the stereo vision has two main problems
(1) Correspondence problem: determining which pixel on the left image corresponds
to which pixel on the right image. (2) Reconstruction problem: define a number
of correspondence pairs to find location and 3D structure of the target object.
Furthermore, the using of IBVS approach in the 3D visual servoing applications
is limited.
Otherwise 4 × 2D visual servoing approach is a simple approach which allows
to use IBVS in 3D space and to estimate the pose of the target object when PBVS
is needed. Next sections will discuss some suggested concepts and notations to
26 M. Bdiwi et al.

Fig. 2.4 4 × 2D visual servoing approach

Fig. 2.5 Coordinate systems

perform the 4 × 2D visual servoing approach, such as mixing of RGB/depth images,

visible side coordinate system, projections in the 3D image space etc.
Figure 2.5 shows three different coordinate systems which are camera (C), tool
(T) and object (O) coordinate systems. The (T) coordinate frame is attached to
the tool of the robot which could be 2 or 3 fingers gripper. There are two other
coordinate systems, which are world coordinate system (W) and a new suggested
coordinate system (S). The (S) coordinate frame is attached to the visible side of the
target object which will be illustrated in the next section.
2 4 × 2D Visual Servoing Approach for Advanced Robot Applications 27

Fig. 2.6 Examples of visible side coordinate system

Figure 2.6 shows some examples with different poses of the visible side planes
according to the camera coordinate system. In the first case the visible side (S)
coordinate system is parallel to the camera (C) coordinate system. In the second
case there is the rotation about xs , whereas in the third case the rotation is about ys
direction.
The transformation from S coordinate system to the image plane coordinate
system will be performed depending on perspective projection matrix as follows:
In the first case of Fig. 2.6, where the S coordinate system is parallel to the camera
coordinate system:
⎡ ⎤ ⎡ ⎤
⎡ ⎤ ⎡ ⎤ 1 0 0 tx xs
u λx Sc u0 0 ⎢ 0 1 0 ty ⎥ ⎢ ys ⎥
S × ⎣ v ⎦ = ⎣ 0 λy v0 0 ⎦ × ⎢ ⎥ ⎢ ⎥
⎣ 0 0 1 tz ⎦ × ⎣ 0 ⎦ (2.1)
1 0 0 1 0
000 1 1

In the second case of Fig. 2.6, where the rotation is about xs direction:
⎡ ⎤ ⎡ ⎤
⎡ ⎤ ⎡ ⎤ 1 0 0 tx xs
u λx Sc u0 0 ⎢ 0 cos θx − sin θx ty ⎥ ⎢ ys ⎥
S × ⎣ v ⎦ = ⎣ 0 λy v0 0 ⎦ × ⎢
⎣ 0 sin θx cos θx
⎥×⎢ ⎥ (2.2)
tz ⎦ ⎣ 0 ⎦
1 0 0 1 0
0 0 0 1 1

In the third case of Fig. 2.6, where rotation is about direction:

⎡ ⎤ ⎡ ⎤
⎡ ⎤ ⎡ ⎤ cos θx 0 sin θx tx xs
u λx Sc u0 0 ⎢ 0 1 0 ty ⎥ ⎢ ys ⎥
S × ⎣ v ⎦ = ⎣ 0 λy v0 0 ⎦ × ⎢
⎣ − sin θx
⎥×⎢ ⎥ (2.3)
0 cos θx tz ⎦ ⎣ 0 ⎦
1 0 0 1 0
0 0 0 1 1
28 M. Bdiwi et al.

2.4 Transformations in 4 × 2D Visual Servoing Approach

2.4.1 Relative Position Between Visible Side and Camera

Coordinate Systems

Figure 2.7 shows the relative position of the origin of the visible side frame
S0 with respect to the camera coordinate system. Kinect camera will deliver the
projected position of S0 into the image plane (u, v) and the distance D between
both coordinate systems. As previously, it can be written:
xc
u=λ (2.4)
zc
yc
v=λ (2.5)
zc

where (xc , yc , zc ) coordinate of point S0 relative to camera coordinate system, u and

v can be calculated from the image plane and λ is camera focal length.
From the depth image the distance D between the origin S0 and the camera
coordinate system, it can be written:

D 2 = xc2 + yc2 + zc2 (2.6)

In Eqs. (2.4), (2.5) and (2.6) four values are known u, v, D, λ and three others are
unknown xc , yc and zc . If we could calculate these three values, we can define the
relative position between the visible side and the camera coordinate systems. The
previous equations could be rewritten as follows:
u
xc = zc (2.7)
λ

Fig. 2.7 Visible side and camera coordinate system

2 4 × 2D Visual Servoing Approach for Advanced Robot Applications 29

v
yc = zc (2.8)
λ
zc2 = D 2 − xc2 − yc2 (2.9)

By substituting (2.7) and (2.8) in (2.9):

 u 2  v 2
zc2 = D 2 − zc − zc (2.10)
λ λ
This gives:

λD
zc = ± √ (2.11)
λ2 + u2 + v 2

By substituting (2.11) in (2.7), xc can be calculated:

uD
xc = ± √ (2.12)
λ2 + u2 + v 2

By compensation (2.11) in (2.8), yc can be calculated:

vD
yc = ± √ (2.13)
λ2 + u2 + v 2

From (2.11), (2.12) and (2.13), the relative position of the origin S0 with respect
to the camera coordinate system (xc , yc , zc ) can be calculated.

2.5 Relative Orientation Between Visible Side and Camera

Coordinate Systems

This section presents how to calculate the orientation of the object with respect to
the camera coordinate system. According to the approach suggested in this work, the
orientation of the object will be calculated from three 2D images (u, v), (u, w), and
(w, v) with the help of image moments. First of all we will explain how to calculate
the 2D orientation of an object with the help of image moments after that we will
generalize this concept to be used in the three images (3D spaces of images).

2.5.1 Object’s Orientation in Image Plane

The orientation of the object in 2D image plane is defined as the angle between the
horizontal axis and the axis, where the object can be rotated with minimal inertia
30 M. Bdiwi et al.

Fig. 2.8 Orientation of the

target object

(i.e. the direction of the major semi-axis). In this direction the object has its longest
extension. Figure 2.8 presents the projection of an object in the (u, v) image plane.
Let’s assume that the target object found in the current image I (u, v) which is
shown in Fig. 2.8 is represented by the indicate function S(u, v).S(u, v), when pixel
belongs to the target object, otherwise S(u, v) = 0. There are two types of moments:
non-central moments mpq and central moments μpq :

V −1 U −1
mpq = up v q S(u, v) (2.14)
v=0 u=0

where the order of the moment is (p + q), so when p = 0 and q = 0 the m00
moment can be calculated as follows:
V −1 U −1 V −1 U −1
mpq = u0 v 0 S(u, v) = S(u, v) (2.15)
v=0 u=0 v=0 u=0

where m00 is the total number of pixels belonging to the object. In the same way the
centroid point of the object can be calculated:
m10
u= (2.16)
m00
m01
v= (2.17)
m00

The central moments of the object S(u, v) can be calculated:

V −1 U −1
μpq = (u − u)p (v − v)q S(u, v) (2.18)
v=0 u=0
2 4 × 2D Visual Servoing Approach for Advanced Robot Applications 31

The main advantage of central moments is that they are invariant to translations
of the object. The disadvantage of the central moments is their dependency on the
size of the object (scale). Usually, the area of the object could be used as a scaling
factor. Hence, by dividing the central moments with power of the area of the object,
we get the central normalized moments νpq :

μpq
νpq = p+q (2.19)
+1
m002

The main advantage of this kind of moments is their invariancy to the scale and
translations. Using the central moments, the orientation of the 2D object can be
calculated as follows:
1 2μ11
θ= arctan (2.20)
2 μ20 − μ02

2.5.2 Object’s Orientation in the 3D Image Space

As shown previously, in our experiment we have supposed that the camera

coordinate system is parallel to the tool coordinate system, which means the object’s
visible side seen by the camera will take the main role to specify the strategies of
performing the grasping or contacting tasks. Therefore, this side will be projected
on the two other planes (u, w) and (w, v). The projections of the visible side to the
two other planes will be calculated by combining the RGB and depth images. These
projections will give us the opportunity to achieve the contacting or grasping tasks
with easier way to extract the vision error between the actual and target position of
the object. Three different examples will be presented in Figs. 2.9, 2.10 and 2.11 to
get better overview of visible side’s projection.
As shown in the previous figures, there are three 2D objects which are the
projections of the visible side of the object in the 3D image space. Hence, there
are three angles Δθ1 , Δθ2 and Δθ3 that can be calculated easily from three image
planes. Δθ1 is the rotation angle between the projection of visible side coordinate

Fig. 2.9 Projections of the visible side (Example 1: Rotation about)

32 M. Bdiwi et al.

Fig. 2.10 Projections of the visible side (Example 2: Rotation about)

Fig. 2.11 Projections of the visible side (Example 3: Rotation about)

system (xs , ys ) and the image plane (u, v). Δθ2 is the rotation angle between the
projection of visible side coordinate system (xs , zs ) and the image plane (u, w). Δθ3
is the rotation angle between the projection of visible side coordinate system (zs , ys )
and the image plane (w, v). As discussed previously, using the central moments of
the image (see (2.20)) and depending on the definition of the 2D object’s orientation
(the angle between the horizontal axes and the direction of the major semi-axis), the
orientation of the 2D objects in three image planes can be easily calculated.

2.6 Control Error of 4 × 2D Visual Servoing Approach

The first step of the visual servoing is defining in any image a particular set of
features that describes the goal which should be reached. In the image-based visual
servoing approach, the determination of the visual features is still considered as
open problem. In other words, how does the system choose the most appropriated
visual features in the control scheme in order to obtain an optimal behavior of the
system? Previously, many papers have used the coordinates of points or parameters
which describe the segments of the image, e.g. straight line, ellipses (Espiau et
al. 1992; Hutchinson et al. 1996). In other works (Chaumette 2002, 2004), it has
been attempted to use moments in image-based visual servoing. In Bdiwi (2014),
the calculation of the general interaction matrix of image moments is illustrated.
Furthermore, the interaction matrix of coordinates of the center of gravity (xg , yg ),
the area (a) and the orientation (θ ) of the target object are in details calculated.
2 4 × 2D Visual Servoing Approach for Advanced Robot Applications 33

Actually, the image moments have been widely used in pattern recognition and
visual servoing for a long time (Prokop and Reeves 1992). However, the main
problem of using image moments in visual servoing loop was that the analytical
form of the interaction matrix related to the image moments was unavailable. In
Bien et al. (1993) the image moments implemented to control only four degrees of
freedom of a robot (x, y, z, θz ) were: the area, the centroid and the main orientation
of an object in the image. These four visual features (coordinates xg and yg of the
center of gravity, area a and orientation θ ) can be easily obtained from moments
of order less than 3 which are closely related to the velocities of the camera vx and
wy , vy and wx , vz and wz respectively. However, using the previous image moments
leads to the situation that the robot will execute trajectories which are desirable in
the image and which can be indirectly and seemingly contorted in Cartesian space.
This potential problem appears when redundant image points coordinates are used,
e.g. coupled features.
The work Corke and Hutchinson (2001) has overcome this problem by decou-
pling the z-axis rotational and translational components of the control from the
remaining degrees of freedom, as follows:

−̇
→
f = Jxy −̇
→r xy + Jz −̇
→rz (2.21)

where −̇
→
r xy = [vx vy wx wy ], −̇
→r = [vz wz ], Jxy , Jz are respectively components
−̇
→
or columns 1, 2, 4, 5 and 3, 6 of Jacobian matrix J and f is the feature point
coordinate error. Hence,

−̇
→ + −̇
→
r xy = Jxy f − Jz −̇
→rz (2.22)

The z-axis velocity will be based directly on two features (orientation and area
of the object).

vz = γvz (a  − a) , wz = γwz (θz − θz ) (2.23)

where γvz , γwz are the scalar gain coefficient. Another approach Chaumette (2002)
has used two supplementary visual features in order to decouple wy from vx and wx
μ30
from vy . These features are based on classical skewness moments sx = 3 and
2
μ20
μ03
sy = 3 . For the symmetrical objects:
2
μ03
⎧
⎪ 1√
⎨ sx = a(s1 t1 + s2 t2 )
K (2.24)
⎪ 1 √
⎩ sy = a(s2 t1 + s1 t2 )
K
34 M. Bdiwi et al.

where:
⎧
⎪
⎨ s1 = μ03 − 3μ21 , t1 = (μ20 − μ02 ) − 4μ12
2 2

s2 = μ30 − 3μ12 , t2 = 4μ11 (μ20 − μ02 ) (2.25)

⎪
⎩ K = Δ(μ + μ ) 32
20 02

For non symmetrical objects:

⎧
⎨p = Δ
x
(μ20 + μ02 )2 (2.26)
⎩
py = a(s12 + s22 )(μ20 + μ02 )3

Hence, the vector of visual features used for tracking symmetrical objects is as
follows:

s = (xg , yg , a, sx , sy , θz ) (2.27)

whereas the vector of visual features used for tracking the nonsymmetrical
objects is:

s = (xg , yg , a, px , py , θz ) (2.28)

For the proposed features and for the configurations where the object is parallel
to the image plane at the desired position, the interaction matrix has shown nice
decoupling feature and the results are satisfactory even if the object is not parallel
to the image plane at the beginning of the positioning task. In this approach, the
maximum order of the implemented moments is the third order. However, whenever
the used moments have less order then the system will be more robust and it will
have more numerical stability. Because of that, one advantage of the proposed
approach in this chapter that the control loop can use the calculated orientations
from (u, w) and (w, v) images instead of sx and sy which can be easily calculated
from the second order moments.
Figures 2.12, 2.13 and 2.14 illustrate some experimental results during the
rotation of the camera about different axes. The target object and its projections
are of green color. Figure 2.12 shows the proportional changing of the orientation
in (u, v) image during the rotation of the camera about axis zc . As shown, there
is nice decoupling between the rotation about zc and the orientation in (w, v) and
(u, w) images, they have constant values. In the same way, the experiment will be
repeated during the rotation of the camera about axis yc , as shown in Fig. 2.13. Only
the orientation (u, w) in image will be changed, whereas orientations in (u, v) and
(w, v) images are constant and they are decoupled from the rotating about yc . In
Fig. 2.14, the experiment will be performed during the rotation of the camera in xc
and only the orientation in (w, v) image will be changed, whereas the orientations
in (u, v) and (u, w) images are constant and they are decoupled from the rotating
in xc .
2 4 × 2D Visual Servoing Approach for Advanced Robot Applications 35

Fig. 2.12 Relation between rotation about zc and orientation in (u, v) image

Fig. 2.13 Relation between rotation about yc and orientation in (u, w) image

Fig. 2.14 Relation between rotation about xc and orientation in (w, v) image
36 M. Bdiwi et al.

Fig. 2.15 Different poses of the target object

Figure 2.15 presents the projections of the target object in different poses on the
four images: (depth, (u, v), (u, w) and (w, v)). The target object is green, whereas
the blue part indicates that this side of the object has higher position (object is diag-
onally positioned). This part of the target object is the position where the robot will
grasp the object, because it is the nearest part of the object from the camera (robot
tool). It is clear that one can directly calculate the inclination angle of the object
either in the first case from (w, v) image or in the second case from (u, w) image.

2.7 Conclusions

To conclude, this chapter has illustrated some previous work of the visual servoing
and it has proposed a new visual servoing approach which is suitable for the new
generation of the RGBD camera (Kinect camera) and advanced robot application.
4 × 2D visual servoing has combined the correspondence of color and depth images
to build two new images. Depending on measured orientations in (u, v), (u, w) and
(w, v) images, the control error signals will be calculated in order to track any flat
surface. Some advanced robot applications based on the proposed approach could
be found in following references (Bdiwi 2014; Bdiwi and Suchý 2012, 2014; Bdiwi
et al. 2013; Kolker et al. 2013).

References

Bdiwi, M. (2014). Development of integration algorithms for vision/force robot control with
automatic decision system. Dissertation, Technical University of Chemnitz, Germany.
Bdiwi, M., & Suchý, J. (2012). Library automation using different structures of vision–force robot
control and automatic decision system. In IEEE/RSJ International Conference on Intelligent
Robots and Systems (IROS), Portugal.
2 4 × 2D Visual Servoing Approach for Advanced Robot Applications 37

Bdiwi, M., & Suchý, J. (2014). Integration of vision/force robot control using automatic decision
system for performing different successive tasks. In Processing on 8th German Conference on
Robotics.
Bdiwi, M., Kolker, A., Suchý J., & Winkler A. (2013). Real time visual and force servoing of
human hand for physical human–robot interaction: Handing–over unknown objects. In Work-
shop on Human Robot Interaction for Assistance and Industrial Robots in IEEE International
Conference on Robotics and Automation (ICRA), Karlsruhe, Germany.
Bien, Z., Jang, W., & Park, J. (1993). Characterization and use of feature–Jacobian matrix for
visual servoing. In K. Hashimoto (Ed.), In visual servoing (pp. 317–363). Singapore: World
Scientific.
Chaumette, F. (2002). A first step toward visual servoing using image moments. In IEEE/RSJ
International Conference on Intelligent Robots and Systems (pp. 378–383).
Chaumette, F. (2004). Image moments: A general and useful set of features for visual servoing.
IEEE Transactions on Robotics, 20(4), 713–723.
Chesi, G., & Hashimoto, K. (2002). Static–eye against hand eye visual servoing. In Proceedings of
41st IEEE Conference on Decision and Control (pp. 2854–2859).
Corke, P., & Hutchinson, S. (2000). Real–time vision, tracking and control. In Proceedings of IEEE
International Conference on Robotics and Automation (pp. 622–629).
Corke, P., & Hutchinson, S. (2001). A new partitioned approach to image–based visual servoing
control. IEEE Transactions on Robotics and Automation, 17(4), 507–515.
Espiau, B., Chaumette, F., & Rives, P. (1992). A new approach to visual servoing in robotics. IEEE
Transactions on Robotics and Automation, 8(3), 313–326.
Flandin, G., Chaumette, F., & Marchand, E. (2000). Eye-in-hand/Eye-to-hand cooperation for
visual servoing. In IEEE International Conference on Robotics and Automation (pp. 2741–
2746).
Goncalves, P., Ferreira, P., & Pinto, P. (2002). Comparing visual servoing architectures for a planar
robot. In Proceedings of the 10th Mediterranean Conference on Control and Automation,
Portugal.
Hashimoto, K. (2003). A review on vision–based control of robot manipulators. Advanced
Robotics, 17, 969–991.
Hill, J., & Park, W. (1979). Real time control of a robot with a mobile camera. In Proceedings of
the 9th SRI International, USA (pp. 233–246).
Hutchinson, S., Hager, G., & Corke, P. (1996). A tutorial on visual servo control. IEEE
Transactions on Robotics and Automation, 1, 651–670.
Kolker, A., Winkler, A., Bdiwi, M., & Suchý, J. (2013). Robot visual servoing using the example
of the inverted pendulum. In 10th IEEE International Multi–Conference on Systems, Signals &
Devices (SSD), Hammamet, Tunesien.
Kragic, D., & Christensen, H. (2002). Survey on visual servoing for manipulation. Computational
Vision and Active Perception Laboratory, Sweden.
Malis, E. (2002). Survey of vision–based robot control. In European Naval Ship Design, Captain
Computer IV Forum, France (pp. 1–16).
Malis, E., Chaumette, F., & Boudet, S. (1999). 2 12 D visual servoing. IEEE International Transac-
tions on Robotics and Automation, 15, 238–250.
Prokop, R., & Reeves, A. (1992). A survey of moment–based techniques for unoccluded object
representation and recognition. Proceedings Computer Vision, Graphics, Image Processing
Conference, 54, 438–460.
Shirai, Y., & Inoue, H. (1973). Guiding a robot by visual feedback in assembling tasks. Pattern
Recognition Pergamon Press, 5, 99–108. Great Britain.
Weiss, L., Sanderson, A., & Neuman, C. (1985). Dynamic visual servo control of robots: An
adaptive image–based approach. In Proceedings of IEEE International Conference on Robotics
and Automation (pp. 662–668).
Weiss, L., Sanderson, A., & Neuman, C. (1987). Dynamic sensor based control of robots with
visual feedback. IEEE Journal of Robotics and Automation, 5, 404–417.
Chapter 3
A Redundant Parallel Robotic Machining
Tool: Design, Control and Real-Time
Experiments

Hussein Saied, Ahmed Chemori, Micael Michelin, Maher El-Rafei,

Clovis Francis, and Francois Pierrot

Abstract In this chapter, we present a machining device, named ARROW robot,

designed with the architecture of a redundant parallel manipulator capable of
executing five degrees-of-freedom in a large workspace. Machine-tools based on
parallel robot development are considered a key technology of machining industries
due to their favourable features such as high rigidity, good precision, high payload-
to-weight ratio and high swiftness. The mechanism of ARROW robot isolates its
workspace from any type of inside singularities allowing it to be more flexible and
dynamic. An improved PID with computed feedforward controller is implemented
on ARROW robot to perform real-time experiments of a machining task. The control
system deals with antagonistic internal forces caused by redundancy through a
regularization method, and achieves a stability conservation in case of actuators
saturation. The results are evaluated using the root mean square error criteria over
all the tracking trajectory confirming the high accuracy and good performance of
ARROW robot in machining operations.

H. Saied ()
LIRMM, University of Montpellier, CNRS, Montpellier, France
CRSI, Lebanese University, Beirut, Lebanon
e-mail: [email protected]
A. Chemori · F. Pierrot
LIRMM, University of Montpellier, CNRS, Montpellier, France
e-mail: [email protected]; [email protected]
M. Michelin
Tecnalia France, Centre Spatial Universitaire, Montpellier, France
e-mail: [email protected]
M. El-Rafei · C. Francis
CRSI, Lebanese University, Beirut, Lebanon
e-mail: [email protected]; [email protected]

© Springer Nature Singapore Pte Ltd. 2019 39

N. Derbel et al. (eds.), New Developments and Advances
in Robot Control, Studies in Systems, Decision and Control 175,
https://doi.org/10.1007/978-981-13-2212-9_3
40 H. Saied et al.

Keywords Machine tool · Redundant parallel robot · Kinematics · Dynamics ·

Singularity analysis · Motion planning · Control · PID · Feedforward ·
Anti-windup

3.1 Introduction

Machining is the process of a controlled material-removal that makes a desired

deformation in the shape and size of raw materials. The theme of the material-
removal classifies machining in different subtractive manufacturing classes (Kerbrat
et al. 2011). It is considered an essential part in the manufacturing of many products
based on metal, wood, plastic, ceramic, and composites. The conventional term,
“machining”, refers to a set of processes such as turning, boring, drilling, milling,
broaching, sawing, shaping, planing, reaming, tapping . . . etc (Davim 2008). Nowa-
days, most of the recently mentioned operations are carried out by Computer
Numercial Control (CNC) in which the computers are used to control the movement
and act of shaping.
Three principal operations of machining are classified as turning, drilling and
milling. All other operations are considered as miscellaneous categories described
in what follows:
• Turning operation is performed by rotating the workpiece and facing it over the
cutting tool as shown in Fig. 3.1a.
• Milling operation is the operation in which the cutting tool rotates bringing the
cutting edges to bear against the workpiece, see Fig. 3.1b.
• Drilling operation is the most common machining process. It is the generation
of cylindrical holes by bringing the rotating cutter with cutting edges into the
workpiece as illustrated in Fig. 3.1c.

Fig. 3.1 Principal machining operations. (a) Turning. (b) Milling. (c) Drilling (eFunda 2018)
3 Redundant Parallel Robotic Machining Tool 41

Fig. 3.2 Illustration of

Kinematic manipulators.
(a) Serial manipulators.
(b) Parallel manipulators
(Kucuk 2012)

• Miscellaneous operations are secondary swarf producing operations performed

on the workpiece after the principal machining processes using the typical
machining tools. For example: boring, broaching, burnishing . . . etc.
The successful design of machine-tools is based on several requirements such
as: high accuracy resulting in a best surface integrity, acceptable level of vibration,
high speeds and feeds increasing the productivity, high static stiffness of the
different machine-tool elements such as structure, joints, and spindles, avoidance
of unacceptable natural frequencies that may cause resonance of the machine-tool,
and finally adopting modern control techniques (Youssef and El-Hofy 2008).
Traditional machine-tools are based on serial manipulators designed as a series
of links connected via actuated joints from the base to the end-effector (see Fig. 3.2).
These manipulators, known with their large range of motions, lack to the good
accuracy and low stiffness. Moreover, they have low dynamic response and valuable
level of vibration (Kucuk 2012; Merlet 2006).
Parallel Kinematic Manipulators (PKMs) offering many advantages over their
serial counterparts are welcomed candidates to be the future machining robots.
According to Merlet (2006), any structure made up of a fixed base and traveling plate
linked together with at least two independent kinematic chains is called a PKM (see
Fig. 3.2). PKMs provide large accuracy, low vibration, high acceleration capabilities
thanks to the light moving parts, and high stiffness due to the closed-chains
structure. However, these manipulators are characterized by a limited workspace
especially in the rotational motion. In the addition to their complexity in the forward
kinematic solutions, and the considerable number of singularities inside and on the
borders of their workspace (Kucuk 2012; Merlet 2006).
Due to the satisfying features of PKMs, they have achieved a real success
in various applications other than machining. PKMs are considered the leading
robots in the automated industrial applications, particularly in the pick-and-place
food packaging tasks which require rapid execution maintaining an acceptable
precision (Natal et al. 2015). Also they are used in the field of laser cutting
42 H. Saied et al.

operations (Bruzzone et al. 2002), medical applications (Robot-Assisted Surgery

(RAS) (Shoham et al. 2003), CardioPulmonary Resuscitation (CPR) (Li and Xu
2007)), flight simulators (Stewart platforms (Stewart 1965)), Haptic devices (delta
robot (Grange et al. 2001)).
Beside their closed-chains structure which increases the stiffness, decreases
the level of vibration and enhances the precision, PKMs shall be equipped with
advanced control techniques to have the best performance of automation. Control of
PKMs is considered as a very challenging task in the literature due to a set of factors
including:
• PKMs are known by their higly nonlinear dynamics which may increase
considerably when operating at high speeds.
• Actuation redundancy which means that the number of actuators is greater than
the number of degrees of freedom (Merlet 2006) exists in some parallel robots.
• Uncertainties and variations of the dynamics and operating conditions are often
present in PKMs.
Several control techniques have been proposed and studied in the literature starting
from the non-model-based non-adaptive and adaptive controllers, ending up with
model-based non-adaptive and adaptive controllers. One can mention few examples
from the aforementioned classes of control implemented on PKMs respectively:
Proportional-Integral-Derivative (PID) (Natal et al. 2015; Ziegler and Nichols
1942), nonlinear PID (Su et al. 2004), Augmented PD (APD) (Zhang et al. 2007),
Adaptive Feedforward plus PD (Codourey et al. 1997; Natal et al. 2012), Desired
Compensation Adaptive Control Law (DCAL) (Bennehar et al. 2014a), Adaptive
controller based on the Robust Integral of the Sign of the Error (RISE) (Bennehar
et al. 2014b), Feedforward Compensation in L1 Adaptive Control (Bennehar et al.
2015).
This chapter introduces a five-degree-of-freedom redundantly actuated PKM
named ARROW1 robot characterized by the good compromise between rapidity
and precision developed for machining tasks. In Sect. 3.2, a general overview on
ARROW PKM structure is stated, in addition to the kinematic and dynamic model-
ing with a brief presentation of the analyzed singularities. Section 3.3 describes the
machining task to be tested, the required trajectory generation, and the proposed
control solution. Experimental results are illustrated, discussed and interpreted
in Sect. 3.4. Finally, Sect. 3.5 concludes the chapter and provides future work
suggestions.

1 ARROW project: a national french research project financed by the National Research Agency
(ANR). Its main objectives can be summarized in the design of Accurate and Rapid Robots with
large Operational Workspace, from which the acronym “ARROW” has been derived. The project
embraces three partners: IRCCyN (Institut de Recherche en Communication et Cybernétique de
Nantes), LIRMM (Laboratory of Informatics, Robotics and Microelectroncs of Montpellier) and
Tecnalia France.
3 Redundant Parallel Robotic Machining Tool 43

3.2 ARROW PKM: General Overview, Modeling

and Singularity Analysis

The ARROW robot is a redundantly actuated PKM has been designed and developed
at LIRMM within the ARROW project. This section covers the general description
of ARROW PKM, the detailed inverse kinematic model, the differential kinematics
and dynamics associated to ARROW robot, and a brief analysis for the singularities
related to the operational workspace. Note that the reader can refer to (Shayya et al.
2014), (Shayya 2015) for further information.

3.2.1 General Overview of ARROW PKM

The structure of ARROW PKM consists of two separate modules as illustrated in

Fig. 3.3: the parallel module and the turntable. The parallel module involves six
linear actuators each pair lays in the same slider track and equidistant from each
other. This configuration facilitates the geometric calibration and reduces the overall
cost from manufacturing point of view. These six actuators are linked by means
of spherical joints to the kinematic chains which are attached to the articulated
traveling plate by revolute joints (see Fig. 3.4). The idea behind using revolute joints
is to avoid some possible internal collisions between the arms. The two intermediate
parallelogram arms are responsible of the rotation of the traveling plate around the
vertical axis parallel to the z-axis, while its position and orientation are originated
from the settle of all the actuators together. Note that the traveling plate can reach

Fig. 3.3 CAD view of the ARROW machining PKM

44 H. Saied et al.

Fig. 3.4 ARROW PKM. (a) 3D CAD view of the parallel module. (b) Graph diagram (Shayya
2015)

Fig. 3.5 Close side view to

the CAD drawing of ARROW
PKM showing the springs
used in the turntable (Shayya
2015)

±45◦ as limits for its allowable range of rotation around z-axis. Thus, the prismatic
motion of the linear actuators sets up the traveling plate in 3T-1R2 dofs.
The turntable is placed facing the parallel module (of Fig. 3.3) and kept outside
the severe zone of collision with the traveling plate. It is actuated with two rotating
motors allowing it to turn around the axis parallel to the x-axis of the fixed frame
shown in Fig. 3.4a. It handles the object to be machined and provides the 5th
rotational dof to the motion range of the traveling plate. It is equipped with a spring
assembly, as illustrated in Fig. 3.5, contributing against gravitational effect and

2 “T” corresponds to translational motion and “R” corresponds to rotational motion.

3 Redundant Parallel Robotic Machining Tool 45

allows maintaining the turntable orientation fixed without any need to the motors
torques. This is beneficial regarding energy saving and reduces control efforts.
From a control point of view, the main part to be controlled is the parallel module.
It holds the machining tool represented as a traveling plate which is a milling
machine in our case. Implementing robust control algorithms to the linear actuators
will set up the machine-tool in a precise tracked trajectory for machining purpose.

3.2.2 Modeling of ARROW PKM

This part explains in details the kinematic, differential kinematic and dynamic
models of ARROW robot. In view of the reported mutations in Shayya (2015) that
have been suggested to implement ARROW PKM avoiding internal arm collisions,
the Forward Kinematic Model (FKM) is difficult to establish if not impossible. Thus,
only the Inverse Kinematic Model (IKM) will be presented and clarified.
Figure 3.6 shows the parallel module with some notations to be used in modeling.
Six arms, simple and parallel, are grouped and numbered from I to VI. The lengths
Li (i = 1, 2, 5, 6) of the simple arms are equal to Ls , while lengths Li (i = 3, 4) of
the parallel arms are equal to Lp .
Two frames are defined for the modeling: Base frame (O, ex , ey , ez ) and
Platform frame (P , exp , eyp , ezp ) (see Figs. 3.7 and 3.9). The following recorded
points are coordinated in the base frame:
• Ai = (xi yi zi )T = (qi yi zi )T is the point attached to the spherical joint for
i = 1, 2, 5, 6 and the center of the linear actuator for i = 3, 4. qi is the position
of the linear actuator in the x-direction ∀i = 1, .., 6 (see Fig. 3.6).
• Bi = (xbi ybi zbi )T ∀i = 1, .., 6 and Di = (xdi ydi zdi )T for i = 1, 2, 5, 6
are the points attached to the revolute joints at the level of the platform, and that

Fig. 3.6 Facing CAD view of the parallel module of ARROW PKM with some notations (Shayya
2015)
46 H. Saied et al.

Fig. 3.7 Side CAD view of

the traveling plate of
ARROW PKM (Shayya
2015)

Fig. 3.8 Side CAD view of

the parallel module of
ARROW PKM (Shayya
2015)

is shown clearly in the schematic view of Fig. 3.9. For i = 3, 4, Di is centered

between the two revolute joints of the platform on its y-axis (see Fig. 3.7).
• Ci = (xci yci zci )T is the intersection between two lines perpendicular to the
axes of revolution of both revolute joints Bi and Di ∀i = 1, .., 6 (Fig. 3.9).
In Fig. 3.8, we can see the distance separating two points Ai and Ai+2 evaluated
by Lz ∀i = 1, 2, 3, 4. Moreover, Di , for i = 1,2,5,6, are located at distance “a” from
the y-axis in the platform frame.
3 Redundant Parallel Robotic Machining Tool 47

Spatial View Top View

qi
Ai yi
Ai ey
ez
Ci eyp
O ey Bi yr
ji i
Bi P
ex Ci
mi Di
O exp
Di ey

Platform
ex ezp // ez
P ∫ TCP

Fig. 3.9 ARROW PKM: schematic view of a kinematic chain with different notations

Two mandatory assembly conditions for the mechanism to function properly are
given in the following two equalities:

Lz = a, Here Lz is chosen greater than a. (3.1)

qi = xi ≤ xdi ≡ xdi+1 ≤ qi+1 = xi+1 , ∀i = 1, 3, 5 (3.2)

3.2.2.1 Inverse Kinematic Model

Regarding the parallel module of ARROW PKM, the position and orientation of the
moving platform in the 3T-1R DoFs are represented by the 4-dimensional coordinate
vector XP M = (x y z θz )T . The six actuators positions are parametrized by the 6-
dimensional coordinate vector qP M = (q1 q2 q3 q4 q5 q6 )T . IKM represents the
transformation from the cartesian space (XP M ) to the joint space (qP M ) for any
configuration in the workspace.
Knowing the end-effector posture XP M in the base frame and the coordinates of
plat
Di in the platform frame PDi , we get Di as follows:

plat
Di = P + R(PDi ), for i = 1, .., 6 (3.3)

where P = (x y z)T is the 3T coordinates of the end-effector, and R ∈ R3×3 is

the basic rotation matrix of the platform frame around the z-axis. Note that Bi , Ci
and Di are in the same horizontal plane (xy), whereas Ai , Bi and Ci remain in the
same vertical plane as illustrated in Fig. 3.9. Thus, A†i , the projection of Ai on the
xy plane, is collinear with Bi and Ci . Then, a right triangle is formed at Ci in the
horizontal plane and we get the following equation:

||A†i Di ||2 = ||A†i Ci ||2 + ||Ci Di ||2 , f or i = 1, .., 6 (3.4)

48 H. Saied et al.

||A†i Ci || can be calculated using the collinearity property of A†i , Bi and Ci as

follows:
⎧ †
⎪
⎪ ||Ai Ci || = ||A†i Bi || + ||Bi Ci ||, ∀i = 1, .., 6
⎪
⎪
⎪
⎪ ||Bi Ci || = ri , constant value
⎪
⎨ 
†
⎪ ||A Bi || = ||Ai Bi ||2 − ||Ai A†i ||2 , right triangle Ai A†i Bi (3.5)
⎪
⎪
i 
⎪
⎪ = L2i − (zi − zi† )2
⎪
⎪
⎩
zi† = zbi = zdi

Knowing that ||Ci Di || = ri is a constant value, one can substitute (3.5) in (3.4)
and get the length ||A†i Di ||. Then, replace the value of ||A†i Di || in the following
equation:
⎧ 
⎪
⎪ ||A†
D || = (x † − xdi )2 + (yi† − ydi )2 + (zi† − zdi )2 , ∀i = 1, .., 6
⎨ i i  i
⎪ = (qi − xdi )2 + (yi − ydi )2
⎪
⎩ †
zi = zbi = zdi , xi† = xi = qi
(3.6)

Based on the assembly mode condition mentioned in (3.2), the IKM of ARROW
PKM parallel module is derived by solving (3.6) and it can be written as follows:

qP M i = xdi + (−1)i ||A†i Di ||2 − (yi − ydi )2 , ∀i = 1, .., 6 (3.7)

As discussed before, the turntable adds the 5th DoF to ARROW robot allowing
the rotation around x-axis. IKM of the turntable delivers the joints position, knowing
the value of the rotational angle is θx , as follows:

qT 1 = θx
(3.8)
qT 2 = −θx

3.2.2.2 Differential Kinematic Model

The differential Kinematic model of the parallel module can be defined as a

relation between the moving platform’s velocity, ẊP M , and the actuated joints
velocity, q̇P M . The matrix links both velocities known as Jacobian is derived in
this paragraph.
As Ai Bi is supposed to be a rigid body doesn’t change its length whatever is the
configuration of the robot, we have:
3 Redundant Parallel Robotic Machining Tool 49

Ai BTi vAi = Ai BTi vBi ∀i = 1, .., 6 (3.9)

with the velocities of vAi and vBi being respectively:

vAi = q̇i ex ∀i = 1, .., 6 (3.10)


vBi = vDi + whp × Di Bi , ∀i = 1, .., 6
plat (3.11)
vDi = vP + wP lat × PDi

where whp = ψ̇i ez ∀i = 1, .., 6 and wP lat = θ̇i ez ∀i = 1, .., 6 are the angular
velocities of the horizontal plane Bi Ci Di and the moving platform respectively (see
Fig. 3.9), knowing that both axes of rotation are parallel to the z-axis in the base
frame. Substituting (3.10) and (3.11) in (3.9) yields to a relation where q̇P M and
ẊP M are coupled to ψ̇. So, there is need to find another equation that leads to
uncoupled relation between q̇P M and ẊP M . This equation is expressed by writing
vAi in terms of vBi and warm knowing that each arm Ai Bi is performing two
rotations: one around an axis parallel to the z-axis in the base frame and the other is
around the axis μi defined in Fig. 3.9. The expression will be as follows:
⎧
⎪
⎨vAi = vBi + warm × (−Ai Bi ), ∀i = 1, .., 6
⎪
warm = wAi Bi = ϕ̇i μi + ψ̇i ez (3.12)
⎪
⎪
⎩μ = ri ×ez |r × e ||
i | i z

From the relations in (3.9) and (3.12), with some computations and simplifications,
the Inverse Differential Kinematic Model (IDKM) for the parallel module is derived
and the form of the inverse Jacobian matrix, Jm , is obtained:
⎧
⎪
⎪ JqP M q̇P M = JXP M ẊP M ⇒ q̇P M = Jm ẊP M with Jm = J−1 qP M JXP M
⎪
⎪ ⎛ ⎞
⎪
⎪ (m ex )(ez (L1 × μ1 ))
T T
⎪
⎪
⎪
⎪ ⎜ 1 ⎟
⎪Jq = ⎜
⎪ .. ⎟
⎪
⎪ PM ⎝ . ⎠
⎪
⎪
⎪
⎪ (m6 ex )(ez (L6 × μ6 ))
T T
⎪
⎪ ⎧
⎨ ⎪
⎪JX (i, 1) = (ez (Li × μi ))(ex mi )
T T
⎪
⎪ ⎪
⎪
⎪
⎪ ⎪
⎪JX (i, 2) = (eTz (Li × μi ))(eTy mi )
⎪
⎪ ⎨
⎪
⎪
⎪JXP M : ⎪JX (i, 3) = (eTz (ni × μi ))(eTz mi )
⎪
⎪
⎪ ⎪
⎪
⎪
⎪ ⎪
⎪
plat
JX (i, 4) = −(eTz (Li × μi ))(eTz (mi × PDi ))
⎪
⎪ ⎪
⎪
⎪
⎪ ⎩n = A C , ∀i = 1, .., 6
⎪
⎪ i i i
⎪
⎩
mi = Ai Di , ∀i = 1, .., 6
(3.13)
50 H. Saied et al.

Note that Jm is not invertible in the case of redundant actuated PKMs, so the
pseudo-inverse of the inverse Jacobian is used as a solution widely adopted in
robotics. Thus, the Forward Differential kinematic Model (FDKM) of the parallel
module is obtained as follows:

ẊP M = Jm ∗ q̇P M (3.14)

where Jm ∗ ∈ R4×6 is the pseudo-inverse of the inverse Jacobian matrix.

In regard to the turntable, the IDKM can be obtained by differentiating the
equations of (3.8) as the following:

q̇T1 = θ̇x
(3.15)
q̇T2 = −θ̇x

3.2.2.3 Dynamic Model

To establish the dynamics of the parallel module of ARROW PKM, two assumptions
have been taken into consideration, similar for the hypothesis applied usually on
Delta-like robots, are mentioned below:
1. The dry and viscous friction forces are neglected in all passive and active joints.
2. The rotational inertia of the arms is neglected, due to their small masses in
comparison with the other components. The mass of each arm is split up into
two equivalent parts, one part is added to the mass of the corresponding linear
actuator while the other part is considered with the moving platform.
Applying the law of motion on the linear actuators gives the following equation:

Ma q̈P M = ΓP M − JTqP M f
(3.16)
Ma = diag(mas , mas , map , map , mas , mas )

where mas ∈ R is the mass including the actuator, moving cart and the half-mass of
the simple arm, while map ∈ R is the mass including the actuator, moving cart and
the half-mass of the parallelogram arm. Ma is called the actuators’ inertia matrix.
ΓP M ∈ R6 is the vector representing the six actuation forces provided by the linear
motors. JqP M ∈ R6×6 is the joint Jacobian matrix defined in (3.13). f ∈ R6 is the
vector of the applied forces on the arms, at Ai , resulting from the acceleration and
gravitational forces acting on the platform.
The dynamics of the moving platform is described by Newton-Euler’s method as
follows:
3 Redundant Parallel Robotic Machining Tool 51

⎧
⎪
⎪ MP ẌP M + Λc ẊP M = JTXP M f + mP g
⎪
⎪ ⎛ ⎞
⎪
⎪ −b
⎪
⎪ m P 0 0 m P sin θz
⎪
⎪ ⎜ ⎟
⎪
⎪ ⎜ 0 mP 0 b mP cos θz ⎟
⎪MP = ⎜
⎪ ⎟
⎪
⎪ ⎜ ⎟
⎪
⎨ ⎝ 0 0 mP 0 ⎠
−b mP sin θz b mP cos θz 0 Ipzz (3.17)
⎪
⎪ ⎛ ⎞
⎪
⎪ −b
⎪
⎪ 0 0 0 m P cos θ z
⎪
⎪ ⎜ ⎟
⎪
⎪ ⎜ 0 0 0 −b m sin θz⎟
⎪
⎪ ⎜
Λc = ⎜
P ⎟
⎪
⎪ ⎟
⎪
⎪ ⎝ 0 0 0 0 ⎠
⎪
⎩
0 0 0 0

with MP and Λc are the mass matrix and the centrifugal and Coriolis effects of
the moving platform respectively. JXP M ∈ R6×4 is the Cartesian Jacobian matrix
defined in (3.13). mP ∈ R symbolizes the total mass of the traveling plate including
the added half-masses of the arms, and g = (0 0 − 9.8 m/s2 0)T is the gravity
acceleration acting on it. The term “b” refers to the x-coordinate belonging to the
center of mass of the cluster formed by the platform and the added point masses
on Bi , whereas “Ipzz ” refers to the total moment of inertia about the z-axis through
TCP of the same formed cluster.
The mass mP and the moment of inertia Ipzz are given as follows:

6
1
mP = (mass of platf orm) + (mass of arm)i (3.18)
2
i=1
6
Ipzz = Iplatf orm + Ii (3.19)
i=1

for which Iplatf orm is the moment of inertia about the z-axis passing through TCP
only of the platform’s body, and Ii is that of the point mass formed at Bi after
making assumption (2) around the same axis.
Differentiating the IDKM obtained in (3.13) delivers the acceleration kinematic
relation as: q̈P M = Jm ẌP M + J̇m ẊP M . By the use of this relation and equations
of (3.16) and (3.17), one can formulate the Direct Dynamic Model (DDM) of the
parallel module as follows:

ẌP M = HΓP M − ΛẊP M + AG (3.20)

where H ∈ R4×6 , Λ ∈ R4×4 and AG ∈ R4 are expressed by:

52 H. Saied et al.

⎧  −1
⎪
⎪
⎪
⎪ H = M + JTM J JTm , dim(H) = 4 × 6
⎪
⎪
P m a m
⎪
⎨  −1
Λ = HMa J̇m + MP + Jm Ma Jm
T Λc , dim(Λ) = 4 × 4 (3.21)
⎪
⎪
⎪
⎪  −1
⎪
⎪
⎪
⎩AG = MP + JTm Ma Jm mP g, dim(AG ) = 4 × 1

The Inverse Dynamic Model (IDM) of the parallel module can be reformulated
from (3.20) by getting the actuator forces vector ΓP M as function of the moving
platform’s position XP M and its derivatives. In case of actuation redundancy, there is
no unique Γ P M vector for a specific desired acceleration. Considering the minimum
norm solution of ΓP M and maintaining the condition of |Γ P M i | ≤ Γ P M max ∀i =
1, .., 6 (all actuators are similar) (Shayya et al. 2014), the IDM of the parallel module
is then arranged in joint space as follows:
 
∗
Γ P M = H ẌP M + ΛẊP M − AG (3.22)

with H∗ ∈ R6×4 being the pseudo-inverse of H.

Considering the dynamics of the turntable part, the varying inertia of the object
to be machined is neglected due to the low acceleration needed for the turntable
rotation. The total moment of inertia of turntable with respect to the axis of rotation
is given as follows:

IT = 2Iact + Itxx (3.23)

where Iact is the rotative motor’s inertia and Itxx is the inertia of the table rotating
around its proper axis which is x-axis. Assuming that the gravitational effect is
compensated for by the springs (see Fig. 3.5), meaning that the effect of gravity
and spring cancel each other, the IDM of the turntable is given by the following
equations:

ΓT1 = (IT θ¨x )/2
(3.24)
ΓT2 = −ΓT1

with ΓT1 and ΓT2 being the torques provided by the rotative actuators driving the
turntable into its rotational operation.

3.2.3 Singularity Analysis

In this branch, a summarized singularity analysis is stated asserting that within

the feasible workspace of ARROW PKM, there is no existence for all types of
singularities (Shayya 2015). Two cases have been studied: constraint and classical
singularity.
3 Redundant Parallel Robotic Machining Tool 53

At first, let the desired workspace be defined as follows:

⎧
⎪
⎪ (x) : −1.5 m ≤ qi ≤ +1.5 m
⎪
⎪
⎪
⎪
⎨(y) : −0.15 m ≤ y − y0 ≤ +0.15 m
⎪
(z) : −0.15 m ≤ z ≤ +0.15 m (3.25)
⎪
⎪
⎪
⎪ ◦ ◦
⎪(θz ) : −45 ≤ θz ≤ +45
⎪
⎪
⎩(θ ) : −45◦ ≤ θ ≤ +45◦
x x

with y0 being the y offset of the center of the desired workspace knowing that the
axis of rotation of the turntable passes through P0 = (0 y0 0)T and parallel to the x
direction.

3.2.3.1 Constraint Singularities

Constraint singularities occur when the pair of complex chains (III) and (IV)
perform undesired rotation for the platform. Normally, the angular velocity of the
platform is represented by the vector wP = (wx = 0, wy = 0, wz )T . Only the two
complex arms are considered in this study with modifying the position of P ≡ TCP
to be confounded with D3 ≡ D4 for more simplicity. Figure 3.10 shows the virtual
equivalent chains with the modification, and defines new notations and subscripts ij
with i = 3, 4 and j = 1, 2.
Investigating the constraint singularities starts from analysing the following
equation:

vAij = vAi = vBij + waij × (−Lij ), ∀i = 3, 4; ∀j = 1, 2
(3.26)
waij = wP + ψ̇ij ez + ϕ̇ij μij

Recalling that in the absence of any constraint singularity we should have wx =

wy = 0 and normal generation for all necessary velocities (vx , vy , wx , wy ,
ψ̇ij , ϕ̇ij , ∀i = 3, 4; ∀j = 1, 2), (3.26) yields to the condition below:
   
h m3x m4y − m3y m4x L3x μ3y − L3y μ3x L4x μ4y − L4y μ4x = 0 (3.27)

From (3.27), one can distinguish three cases lead to constraint singularity:
1. h = 0. This is not possible because h is restricted to be opposite to zero by
manufacturing
 process.

2. m3x m4y − m3y m4x = 0. This means the xy projections of m3 and m4 are
collinear. This practically is preceded by collision between platform and slider’s
wall, or between the actuators. So it is on the borders or outside the accessible
workspace.
54 H. Saied et al.

Fig. 3.10 The complex chains (III) and (IV) with notations (Shayya 2015)

   
3. L3x μ3y − L3y μ3x = 0 or L4x μ4y − L4y μ4x = 0. These two terms
are simultaneously zero or non-zero due to the structural symmetry of the
mechanism. If both are zero, this means that L3 //L4 //ez which comes after a
collision in the exterior region of the allowed workspace.
Therefore, the analysis outcomes guarantee that there are no constraint singular-
ities within the operational workspace.

3.2.3.2 Classical Singularities

Two cases can be defined for the classical singularities to be existing: First case
is described by knowing q̇P M but we cannot definitely determine ẊP M , ϕ̇ =
(ϕ̇1 , . . . , ϕ̇6 )T or ψ̇ = (ψ̇1 , . . . , ψ̇6 )T , second case is when knowing ẊP M
but we cannot definitely determine q̇P M , ϕ̇ or ψ̇. Mathematically, the matrices
JqP M , JXP M , Jϕ and Jψ should not be rank deficient in order to avoid the afore-
mentioned two cases. JqP M , JXP M were given in (3.13), and Jϕ , Jψ are derived also
from Eqs. (3.9) and (3.12) as follows:
3 Redundant Parallel Robotic Machining Tool 55

⎧
⎪
⎪ Jϕ ϕ̇ = JXP M ϕ ẊP M
⎪
⎪ ⎛ ⎞
⎪
⎪ T (L × μ ))
⎪
⎪ (ez 1 1
⎪
⎪ ⎜ ⎟
⎪
⎪ ⎜ .. ⎟
⎪Jϕ = ⎝
⎪ . ⎠
⎪
⎨
(eTz (L6 × μ6 ))
⎛ ⎞ (3.28)
⎪
⎪ −e T 0 0 0
⎪
⎪ z
⎪
⎪ ⎜ ⎟
⎪
⎪ ⎜−ez T 0 0 0⎟
⎪
⎪ J = ⎜ ⎟
⎪
⎪ XP M ϕ ⎜ T 0 0 0⎟
⎪
⎪ ⎝ −e z ⎠
⎪
⎩
−ez 0 0 0
T

⎧
⎪
⎪Jψ ψ̇ =
⎪
⎪
J Ẋ
⎛ Xψ P M ⎞
⎪
⎪
⎪
⎪ (mT1 ex )(eTz (L1 × μ1 ))
⎪
⎪J = ⎜
⎪
⎪ ⎜ .. ⎟
⎟
⎪
⎪ ψ ⎝ . ⎠
⎪
⎪
⎨ (mT6 ex )(eTz (L6 × μ6 ))
⎛ ⎞
⎪
⎪ 0 ez T (L1 × μ1 ) −(ex T μ1 )(ez T L1 ) (eTx P1 )(eTz (L1 × μ1 ))
⎪
⎪ ⎜. ⎟
⎪
⎪ .. .. ..
⎪
⎪ JXP M ψ = ⎜ ⎝
.. ⎟
⎠
⎪
⎪
. . .
⎪
⎪
⎪
⎪ 0 ez (L6 × μ6 ) −(ex μ6 )(ez L6 ) (ex P6 )(ez (L6 × μ6 ))
T T T T T
⎪
⎪
⎩P = PDplat , ∀i = 1, .., 6
i i
(3.29)

The rank deficiencies of JqP M = Jψ and Jϕ are given by the following

singularities respectively:

det (JqP M ) = det (Jψ ) = 0 ⇐⇒ ∃i0 ∈ {1, .., 6}; eTx mi0 = 0
or eTz (Li0 × μi0 ) = 0 (3.30)

det (Jϕ ) = 0 ⇐⇒ ∃i0 ∈ {1, .., 6}; eTz (Li0 × μi0 ) = 0 (3.31)

Notice that (3.30) covers also the singularity of Jϕ . The case of eTx mi0 = 0 means
that mi0 is perpendicular to the x-axis of the base frame which takes place when mi0
lays in the yz plane. Still this configuration is outside the regional workspace and it
provokes prior collisions. For the other case, when eTz (Li0 × μi0 ) = 0, then Li0 //ez
which is similar to the what have been discussed in the constraint singularities of
collision between sliders and platform.
On the other hand, the rank deficiency of JXP M is studied after performing some
linear operations achieving a simpler form. The changing of P to be confounded
with D1 was for that purpose. Having a deficient rank of JXP M turns mathematically
into the following equations:
56 H. Saied et al.

⎧
⎪
⎪eTz (L3 × μ3 ) = eTz (L4 × μ4 ) = 0
⎪
⎪
⎨ or
(3.32)
⎪
⎪eTz (m3 × rc ) = eTz (m4 × rc ) = 0
⎪
⎪
⎩
with rc = D1 D3 ≡ D1 D4

Or
⎧
⎪
⎨(ez (L1 × μ1 ))(m1 ex ) = 0
⎪ T T

and (3.33)
⎪
⎪
⎩(eT (L × μ ))(mT e ) = 0
z 5 5 5 x

Or
⎧
⎪
⎨ez (L1 × μ1 ) = 0 or ez (n5 × μ5 ) = 0 or L5 ez = 0
⎪ T T T

and (3.34)
⎪
⎪
⎩eT (L × μ ) = 0 or eT (n × μ ) = 0 or LT e = 0
z 5 5 z 1 1 1 z

Regarding the conditions in (3.32), for eTz (L3 × μ3 ) = eTz (L4 × μ4 ) = 0 it

was proved in the previous paragraph that it is not possible to occur without a
prior collision. However, the second case of condition (3.32) (i.e eTz (m3 × rc ) =
eTz (m4 × rc ) = 0) can be interpreted by having the four vectors ez , m3 , m4 and rc
in the same plane which is either xy or yz plane. This can be take into consideration
after a collision between the platform and the slider’s plane or between the actuators
numbered 3 and 4. Thus, it is not possible to have such condition within the
operating workspace. Therefore, condition (3.32) has no concern with singularities
inside ARROW’s workspace.
Furthermore, both cases of condition (3.33) have been interpreted in the previous
parts (particularly condition (3.30)).
Moreover, the possibility of having eTz (L1 × μ1 ) = 0 or eTz (n5 × μ5 ) = 0 in
condition (3.34) has been verified in the precedent parts that it is outside the exe-
cutable range of the platform. For the case of eTz (L5 × μ5 ) = 0 or eTz (n1 × μ1 ) =
0, it can be analysed as having the three corresponding vectors laying in the same
vertical plane. Actually, we have μi ⊥ ni and μiz = 0, so the only incidents lead to
the aforementioned condition is having n1 //ez or n5 //ez which occur after falling
in collisions. Besides, the situations of LT1 ez = 0 and LT5 ez = 0 take place just on
the boundaries of the workspace, but it can’t happen simultaneously because of the
assembly condition (3.1) that states Lz = a. So, as well as for conditions (3.32)
and (3.33), condition (3.34) has no possibility to lead ARROW robot falling in any
singularity in the feasible workspace.
3 Redundant Parallel Robotic Machining Tool 57

Lastly, it was shown that ARROW PKM can not fall into any type of singularities
inside the functional workspace providing more flexibility and less danger of
structural damaging.

3.3 Motion Control

In this section, the proportional-integral-derivative with computed feedforward

control solution implemented on ARROW robot is explained. An improved PID
controller is used in our case to deal with the produced internal forces due to
redundancy. The integral term overshoot and oscillations in case of saturated
actuators have been avoided thanks to the anti-windup strategy added to the adopted
PID controller.

3.3.1 PID Control in Case of Non-redundancy

The Proportional-Integral-Derivative (PID) control law proposed in Ziegler and

Nichols (1942) is the simplest and easiest algorithm that can be adopted for motion
control of parallel robots. Although PID is non-robust against nonlinearity effects,
changing parameters and uncertainties which are abundant in parallel robots, it is
still the most implemented control strategy in the industrial robotized applications.
In a typical non-redundant parallel robot, the PID control law employed in joint
space can be defined in discrete form as follows:
⎧
⎪
⎪ Γ P I D (k) = Γ P (k) + Γ I (k) + Γ D (k)
⎪
⎪
⎪
⎨Γ P (k) = KP  q(k)
Γ I (k) = Γ I (k− 1) + KI Ts (3.35)
⎪
⎪ q(k)
⎪
⎪
⎪Γ (k) = K q(k)−q(k−1)
⎩ D D Ts

where Γ P (k), Γ I (k) and Γ D (k) are the corresponding output torques or forces
of the proportional, the integral and the derivative controllers at time step k, while
KP , KI and KD are their diagonal positive definite matrices gains respectively.

q = qd − q is the joint position error vector defined by the difference between the
desired position trajectory and the measured one. Ts is the sampling period. The
schematic diagram of the PID controller for non-redundant parallel manipulators is
shown in Fig. 3.11.
However, this control scheme doesn’t work correctly for the redundant parallel
manipulators since there are more control inputs than the DoFs of the moving
platform, which will be discussed in the next subsection.
58 H. Saied et al.

Fig. 3.11 Schematic view of the PID control for non-redundant parallel robot in joint space

Fig. 3.12 Actuator measurements and the moving platform’s position. (a) Ideal non-redundant
case. (b) Ideal redundant case (Yang 2012)

3.3.2 PID Control in Case of Redundancy

Beside the various advantages of the actuation redundancy in parallel robots such
as providing more accuracy, improved stiffness and more dexterity, it may declines
the control performance of PKMs in the presence of undesired internal generated
forces. These forces are mainly due to the “fight against” actuators efforts since
there are more inputs than required to accomplish a specific end-effector motion.

3.3.2.1 Actuation Redundancy Effects

Figure 3.12 shows the one DoF of a moving platform which is actuated by one
actuator for the non-redundant case and by two actuators in the redundant case.
In the ideal case shown in Fig. 3.12, the actuators and moving platform can always
track the desired trajectory in both cases of redundancy and non-redundancy. Indeed,
geometric errors always work out from various sources such as inaccuracies in the
3 Redundant Parallel Robotic Machining Tool 59

model geometry, measurement errors, assembly errors, non-synchronized control

of the actuators, backlashes, thermal expansion, etc (Muller 2009). In the shade of
those errors, applying the classical scheme of PID control shown in Fig. 3.11 may
lead to an error with the moving platform position tracking while still able to follow
the desired actuators position signal. On the other hand, the two actuators in the
redundantly actuated systems can hardly reach the desired positions at the same
time conflicting to each other, which may generates antagonistic internal forces that
may destroy the whole mechanism of the robot.
The position vector  q can be considered as a vector from the actuator velocity
space which always be in C(Jm )3 in case of non-redundant actuated mechanisms
even with the existence of geometric errors. On the contrary, in case of redundantly
actuated mechanisms,  q may has some portions from the null space and not only
formed of C(Jm ). These errors may accumulate in the integral term of PID control
generated as an input control torques or forces to the actuators. Thus the internal
forces of the mechanism will be more and more significant becoming destructive to
the mechanisms.

3.3.2.2 Adopted Solutions

Several approaches are proposed in Yang (2012) by following a general idea that is
to get rid of the null space portions of 
q in the control loop. One of the propositions
is to perform regularization for the input of the PID block as well as for its output.

Regularization Before PID Block

The regularization for the input, 

q, consists of two steps described as follows:
1. Transforming the error vector from the joint space to the Cartesian space using
the velocity mapping which guarantees the elimination of any residual parts out
of C(Jm ).


X = J∗m
q (3.36)

2. Calculating again the joint space error from the Cartesian one by the inverse
velocity mapping relationship, which results a regularized joint space error
vector.

qReg = Jm 
 X (3.37)

3 Columns of Jm .
60 H. Saied et al.

The transformations in (3.36) and (3.37) are combined as follows:


qReg = RV 
q
(3.38)
RV = Jm J∗m

with RV being the regularization matrix in the velocity space.

Regularization After PID Block

As it has been explained in Muller and Hufnagel (2011), the internal forces are
caused by the control inputs which are in the null space of the inverse Jacobian
matrix. The regularization matrix projects the generated control inputs into the range
space of Jm in two steps described as follows:
1. Using the force mapping relationship, the output forces in the actuation space are
transformed to the wrench in the Cartesian space.

w = JTm Γ (3.39)

2. Calculating again the actuation space forces from the wrench by the inverse force
mapping relationship, which results a regularized output forces vector.

Γ Reg = (Jm T )∗ w (3.40)

The transformations in (3.39) and (3.40) are combined as follows:


Γ Reg = RΓ Γ
(3.41)
RΓ = (JTm )∗ Jm T

with RΓ being the regularization matrix in the force or torque space. It was shown
in Yang (2012) that the two regularization matrices in velocity and force spaces are
equal, RΓ = RV ≡ Rm .

Regularization for the Integral Term

The output Γ P I D (k) generated from the PID given in (3.35) can be shown as a linear
function of the current input to PID block 
q(k), the previous input q(k − 1), and the
previous output Γ P I D (k − 1). This linear relation is addressed in the following
equation after several manipulations as follows:
3 Redundant Parallel Robotic Machining Tool 61

Fig. 3.13 Regularization algorithm for the integral term of PID controller in discrete form

⎧
⎪
⎪ Γ P I D (k) = α1
q(k) + α2
q(k − 1) + α3 Γ P I D (k − 1)
⎪
⎪
⎨α = K + T K + 1 K
1 P s I Ts P (3.42)
⎪
⎪ 2
α = − 1
K
⎪
⎪ Ts D
⎩
α3 = I

To be sure that both the input and output of PID control remain in the same subspace,
we shall confirm that q(k − 1) and Γ P I D (k − 1) have the same subspace as q(k).
For a PID block having a regularized input position error vector,  qReg , the
C(Jm ) may conserve its dimensions but may not stay in the same subspace during
operation. This may influence the output of PID block, Γ P I D (k) may not stay
always in R(Jm ).4 Moreover, Γ P I D (k − 1) which is calculated in a recursive
way may accumulate undesired quantities in the null spaces. So there is a need
to an independent regularization process on the Γ P I D (k − 1) to ensure the space
conservation of the input and output of PID block.
As well as the PID block with after regularization needs to a regularized
Γ P I D (k − 1) since the generated regularized output Γ Reg eliminates the antago-
nistic forces from the mechanism but doesn’t solve the problem of the increasing
value of the integral term with time.
A solution for the aforementioned problems is proposed in Yang (2012) which
requires a regularization for the integral term at each time step using the algorithm
below (see Fig. 3.13).
At each step time k:
1. Regularize the generated force from the applied integration on the regularized
error joint position vector 
qReg (k).
 
Reg Reg
Γ I (k) = Rm Γ I (k − 1) + KI Ts qReg (k) (3.43)

2. Update the regularized value for the next time step.

Reg Reg
ΓI (k − 1) = Γ I (k) (3.44)

4 Rows of Jm .
62 H. Saied et al.

3.3.2.3 Anti-windup Strategy

To this end, an improved PID controller has been proposed to deal with the
antagonistic forces and integral term accumulation in the null space in case of
redundantly actuated parallel robots. One more modification was employed to PID
controller is the anti-windup strategy.
During the step change in PID controller, the integral windup occurs for
saturation of actuator. Hence the system error decreases more slowly than ideal case
causing the value of integral term to be increased, which may lead to significant
large overshoot and settling time. This loss of performance provoked by the integral
windup in systems of saturated actuators guides mostly to remove the integral term
missing out its advantages in omission of the stable residual errors coming from
friction or external loads.
The aim behind using the anti-windup strategy is to ensure conservation of the
controller stability avoiding the oscillatory behaviour when the actuators saturated.
The back-calculation approach of anti-windup strategy proposed in Kumar and
Negi (2012) is used here for the PID controller block. It requires feeding back the
difference between saturated and unsaturated signals which may reduce the integral
value. The integral output of the PID block with the three regularization treatments
mentioned before and anti-windup strategy is then written as follows:
⎧  
⎨Γ Reg (k) = Rm Γ Reg (k − 1) + KI Ts 
⎪
qReg (k) − KAW P ΔΓReg

I I
(3.45)
⎪
⎩ΔΓReg = Γ sat − ΓReg
Reg

where ΓReg sat and Γ

Reg are the generated output forces or torques from the PID
block after and before saturation block respectively. KAW P is a diagonal positive
definite matrix representing the anti-windup feedback gain. The schematic diagram
illustrating the regularization technique of the integral term and the anti-windup
strategy for the PID block is shown in Fig. 3.14.

3.3.3 Proposed Control Solution for ARROW PKM

PKMs are well known with their highly nonlinear dynamics which increase con-
siderably when operating at high speed leading to instability. These nonlinearities
shall be treated carefully in control algorithms in order to compensate their effects,
knowing that linear controllers fail to attain this mission such as PID controller.
In fact, a closed-loop control algorithm full enough with knowledge about the
dynamics of the PKM brings out good robustness against nonlinearities and distur-
bances. A PID with FeedForward control law (PIDFF) is one of the most efficient
non-adaptive model-based controllers that exist for parallel manipulators. PDFF
control law was firstly suggested in (Liégeois et al. 1980) as an alternative to the
3 Redundant Parallel Robotic Machining Tool 63

Fig. 3.14 PID block schematic diagram with regularized integral term and anti-windup strategy
in the discrete form

on-line computation requirements of others model-based motion control schemes,

particularly the computed torque control method. PIDFF control law consists of
a linear PID feedback plus a feedforward part results from the dynamics of the
robot evaluated using the desired motion trajectory. Consequently, the feedforward
dynamics can be computed offline reducing the control computation efforts and
making this type of model-based controllers easy and simple for implementation
as well as PID controller.
For ARROW PKM, the implemented control law is the PIDFF controller using
the PID block recently obtained in Fig. 3.14, individually for each part of ARROW:
the parallel module and the turntable. The PIDFF control implemented for ARROW
PKM can be written in joint space as follows:
⎧  
⎪
⎪ =  +
Reg
+ ˙ + ΓP M;T

⎪
⎪ UP M;T R m KP q Γ KD q
⎪
⎪ 
I

ff
⎪
⎨
ΓP Mff = H∗d ẌdP M + Λd ẊdP M − AGd (3.46)
⎪
⎪  
⎪
⎪
⎪
⎪ IT /2
⎩ΓTff =
⎪ θx d
−IT /2

Reg
where ΓI is the modified integral term including the regularization and anti-
windup strategy obtained in (3.45). XPd M , ẊPd M and ẌPd M are the desired position,
velocity and acceleration of the moving platform respectively. Hd , Λd and AGd are
the matrices used in the dynamic model of the parallel module in (3.22) calculated
based on the desired generated trajectory. UP M;T is the control input forces vector
applied to the linear actuators or torques vector applied to the rotative motors. θx d
is the desired generated rotation in the Cartesian space around x-axis. The proposed
control solution for ARROW PKM is described as a schematic view in Fig. 3.15.
64 H. Saied et al.

Fig. 3.15 PIDFF block schematic diagram with regularizations and anti-windup strategy applied
on ARROW PKM

It is visible in (3.46) that feedforward dynamics are composed from the desired
trajectories making this controller preferable for real-time implementation. In other
meaning, PIDFF controller enhances the general performance and increases the
accuracy with the same processing time cost of a simple PID. As well as it avoids
the noise estimations of velocity and acceleration for the actuators knowing that
most of the real manipulators are equipped only with position sensors (i.e encoders).
Nevertheless, such kind of model-based controllers needs a precise knowledge about
the dynamics of the parallel robot, which is a difficult development task, in order to
achieve high level of performances.
The stability analysis of this model-based controller has been studied in the
literature. In Reyes and Kelly (2001), it has been reported that the position error
of a PD control with computed feedforward will vanish asymptotically in a local
sense after selecting properly the feedback matrices. Furthermore, it has been proved
in Santibanez and Kelly (2000) that this controller is able to yield a globally
asymptotically stable closed-loop system with an experimental validation.

3.4 Motion Generation

A strong requirement in the motion generation for a machining tool is the constant
feed rate. The tool have to maintain a constant velocity along the path. To ensure
this, the Gcode position data series coming from a CAD/CAM software are used as
input for a Spline based definition of path.

3.4.1 Spline Description

We can extract, from the Gcode instructions, the tool positions along the machining
path. The first idea in creating a trajectory through points is to generate straight
3 Redundant Parallel Robotic Machining Tool 65

Fig. 3.16 Straight lines path

with C0 position continuity
2

1 3

Fig. 3.17 Path with a

tangency continuity C1 2

1 3

Fig. 3.18 Path with a

curvature continuity C2 2

1 3

lines between points such as in Fig. 3.16. This kind of trajectory ensures a position
continuity at each junction point called C0 continuity.
Off course, this kind of trajectories will generate velocity discontinuities along
the path which is not suitable to generate a proper machining motion. A second
manner to generate a trajectory is to create polynomials between points and ensure
that the successive polynomials are tangent at junction points (see Fig. 3.17). This
kind of continuity is called C1 and is characterized by an equality of the first
derivatives of successive polynomials at a junction point.
In the context of high speed machining, the continuity C1 is not sufficient. Indeed,
the path curvature from one side to the other of a point isn’t continuous (ex: point 2
in Fig. 3.17). To ensure the continuity of curvature, called C2 continuity, the second
derivative of successive polynomials have to be equal at the junction points (ex:
point 2 in Fig. 3.18).
This natural cubic spline technique is used to create a path which is optimal for
the machining process. A spline is a function defined piecewise by polynomials
between each successive pair of points describing the path. In the case of natural
cubic spline, those polynomials are of degree three with the following form:

P (t) = a + bt + ct 2 + dt 3 (3.47)
66 H. Saied et al.

Fig. 3.19 Spline passing through four points (defined by four polynomials)

Figure 3.19 shows a spline made of four polynomials (P1 (t), .., P4 (t)) passing
through five points. The position continuity C0 , the tangency continuity C1 and
the curvature continuity C2 are ensured by the use of natural cubic spline curve.
Warning to the parameter t which is not a time variable but a sort of “spline
abscissa”.
In 3D space, the polynomial Pi (t) is represented by a concatenated vector of
three coordinate polynomials xi (t), yi (t) and zi (t) with a similar form to (3.47) as
follows:
⎛ ⎞
xi (t) = aix + bix t + cix t 2 + dix t 3
Pi (t) = ⎝ yi (t) = aiy + biy t + ciy t 2 + diy t 3 ⎠ (3.48)
zi (t) = aiz + biz t + ciz t 2 + diz t 3

The Matlab curve fitting toolbox provides the “cscvn” function allowing to get
the natural cubic spline definition. This function takes as input the n Gcode path
points series and gives as output the spline structure. It is specified by a “break
sequence” along the “spline abscissa” for each Gcode path point and the “coefficient
array” of polynomials composing the spline shown respectively as follows:

breaks = [0 break1 break2 . . . breakn ] (3.49)

⎛ ⎧ ⎞
⎪
⎪ d1x c1x b1x a1x
⎜ ⎨ ⎟
⎜ coef f or P1 = d1y c1y b1y a1y ⎟
⎜ ⎪ ⎟
⎜ ⎪
⎩d c b a ⎟
⎜ ⎟
⎜ ⎧ 1z 1z 1z 1z ⎟
⎜ ⎪d2x c2x b2x a2x ⎟
⎜ ⎪
⎨ ⎟
⎜ ⎟
⎜ coef f or P2 = d2y c2y b2y a2y ⎟
⎜ ⎪ ⎟
coef s = ⎜ ⎪
⎩d c b a ⎟ (3.50)
⎜ ⎟
⎜ 2z 2z 2z 2z ⎟
⎜ .. .. ⎟
⎜ . . ⎟
⎜ ⎧ ⎟
⎜ ⎪ ⎟
⎜ ⎪
⎨ dnx cnx bnx anx ⎟
⎜ ⎟
⎜coef f or Pn = d c b a ⎟
⎝ ⎪ ny ny ny ny ⎠
⎪
⎩d c b a
nz nz nz nz
3 Redundant Parallel Robotic Machining Tool 67

Fig. 3.20 Trajectory re-sampling at each sampling period

Thus, the polynomial P1 (t) gives the position of Gcode point 1 when parameter
t = 0 and gives the position of point 2 when parameter t = break1 . The spline P2 (t)
gives the position of Gcode point 2 when parameter t = 0 and gives the position of
point 3 when parameter t = break2 .

3.4.2 Spline Re-sampling During Constant Velocity Phase

During the real-time execution of the machining motion, the sampling period is
fixed at 0.2 ms (5 KHz). The Gcode path points are generated from CAD/CAM
geometrical specifications and are not distributed in terms of sampling period. The
spline re-sampling will consist in generating the path at each sampling period as
illustrated in Fig. 3.20.
As the velocity has to be constant along the path during a machining process,
we have made the assumption that the distance between two successive re-sampled
points is constant. Indeed, we consider that distance between two points along the
path is very close to the cord distance between those two points. The determination
of re-sampling points position is done by the following iterative process:
1. From Gcode point 1, small iterations δt are done on the parameter t and the
corresponding position on the curve is computed from spline polynomial P1 (t)
as shown in Fig. 3.21.
2. This is repeated till the distance between the Gcode point 1 and the actual
position on the curve exceeds the desired distance between two successive re-
sampling points.
3. The last position reached via this iterative process is considered as the first re-
sampling point.
The process is re-iterated for the following successive re-sampling points (black
points in Fig. 3.21) till the reaching of Gcode points 2, 3 and so on. The more
tiny δt is considered, the more precise process is obtained. During this process,
the increasing of parameter t along the break sequence allows to know which
polynomial has to be used to compute the positions on the curve (ex: (t) is used
between Gcode points 1 and 2, P2 (t) between Gcode points 2 and 3).
68 H. Saied et al.

Fig. 3.21 Iterative process for re-sampling point determination

During the constant velocity phase, the use of spline curve guarantees an optimal
trajectory in terms of velocity and acceleration discontinuities for machining
motion.

3.4.3 Acceleration and Deceleration Phases

The previous section deals with the motion generation along the spline curve when
the velocity is constant. In a fast machining process, the reaching of the constant
velocity phase has to be done in a smooth way aiming at increasing durability of
mechanical parts. The motion has to start with an acceleration phase and to finish
with a deceleration phase. A suitable acceleration procedure is used in integrating
a constant jerk profile to get a proper acceleration and velocity profile. The Jerk
profile (Fig. 3.22a) is introduced as follows:
3 Redundant Parallel Robotic Machining Tool 69

Fig. 3.22 Acceleration phase

⎧
⎪
⎪ 0 < t < t1
⎨J,
J erk = 0, t 1 < t < t2 (3.51)
⎪
⎪
⎩−J, t < t < t
2 3

The integration of Jerk versus time t gives the acceleration profile shown in
Fig. 3.22b such as:
⎧
⎪
⎪ 0 < t < t1
⎨J t,
Acceleration = A, t 1 < t < t2 (3.52)
⎪
⎪
⎩A − J (t − t ), t 2 < t < t3
2

A second integration of acceleration versus time t gives the Velocity profile shown
in Fig. 3.22c such as:
⎧ 2
⎪ Jt
⎪ 2 , 0 < t < t1
⎨
J t12
V elocity = + A(t − t1 ), t 1 < t < t2 (3.53)
⎪
⎪ 2
⎩ J t12 J (t−t2 )2
2 + A(t2 − t1 ) + A(t − t2 ) − 2 , t 2 < t < t3

The deceleration phase is only an inversion of acceleration phase as shown in

Fig. 3.23. Thus, the Velocity profile can be given as follows:
70 H. Saied et al.

Fig. 3.23 Deceleration phase

⎧  J t2 
⎪
⎪ V − 0 < t < t1
⎨ 2 ,
 J t12 
V elocity = V − 2 + A(t − t1 ) , t 1 < t < t2
⎪
⎪  J t2
⎩ J (t−t2 )2 
V − 21 + A(t2 − t1 ) + A(t − t2 ) − 2 , t 2 < t < t3
(3.54)
Those velocity profiles are used in determination of re-sampling points spacing
along the spline curve during acceleration and deceleration phases. Now, we have
gotten a complete motion description at a certain sampling period.

3.5 Real-Time Experiments

This section exposes the real-time evaluation for the machining performance of the
introduced ARROW robot. It covers the description of the experimental testbed
consists of ARROW, the generated machining trajectory and the results concerning
precision and performance.

3.5.1 Experimental Testbed of the ARROW Robot

The mechanical structure and general mechanism of ARROW PKM has been
discussed previously a lot in this chapter. The manufactured ARROW robot is shown
in Fig. 3.24, where you can see the turntable with the object to be machined, the arms
and the actuators.
3 Redundant Parallel Robotic Machining Tool 71

Fig. 3.24 The manufactured

ARROW robot

Fig. 3.25 The experimental setup of ARROW robot. 1: Master computer, 2: xPC Target, 3: Power
supply, 4: Emergency button

The parallel module actuators are all identical of type Ironless ETEL ILM12-060,
providing each one a maximum force of 2500 N and reaching speed up to 15 m/s.
The actuators of the turntable are of type TMB0140-100-3RBS ETEL direct-
drive motors. They can provide a maximum peak torque of 127 Nm and they are
able to reach 550 rpm of maximum speed. Each actuator is equipped with a non-
contact incremental optical encoder providing a total number of 5000 pulses per
revolution.
The experimental setup of the ARROW robot is displayed in Fig. 3.25. Simulink
and Real-Time Workshop from Mathworks Inc. are used to implement the control
scheme and to generate the real-time execution code. The target PC is running under
72 H. Saied et al.

Table 3.1 The main dynamic parameters of ARROW robot

Parallel module Turntable
Parameter Value Parameter Value
Simple slider cart’s mass 11.1 kg
Parallelogram slider cart’s 11.34 kg
mass
Simple arm linear mass 1.744 kg/m Turntable actuator’s inertia 0.004 kg.m2
Parallelogram arm linear mass 3.488 kg/m Total inertia of the turntable 1.204 kg.m2
Simple arms’ length 0.96 m
Parallelograms arms’ length 0.61 m
Platform’s mass 10.2 kg
Platform’s inertia 0.414 kg.m2

Table 3.2 Summary of the controllers’ parameters

Parallel module Turntable
Gain Value Gain Value
kP 5860796.8 kP 5550
kD 8070.24 kD 25.9
kI 20425584.8 kI 7400
kAW P 166×10−5 kAW P 37×10−6

Fig. 3.26 The machining

object before and after
performing the milling
process

5 kHz of frequency (i.e. sample time of 0.2 ms). The set of various geometric and
dynamic parameters of the different parts of the ARROW robot are summarized in
Table 3.1.
The control tuning gains that are specified for both control architectures of
parallel module and turntable are shown in Table 3.2.
The chosen machining trajectory leads to obtain the pyramid squared shape seen
in Fig. 3.26 after the milling operation is performed on a cubed shape. A 3D plot
of the adopted trajectory is illustrated in Fig. 3.27 from the side view of the parallel
module. The chosen machined shape estimates the five DoFs of ARROW robot and
allows us to evaluate the whole performance in the accessible workspace.

3.5.2 Experimental Results

In order to evaluate the performance of the proposed control solution, PIDFF control
law, on the whole trajectory in both Joint and Cartesian spaces, the Root Mean
3 Redundant Parallel Robotic Machining Tool 73

Fig. 3.27 3D plot for the machining trajectory as if seen from the parallel module side view

Square of the tracking Error in Translational (RMSET) and Rotational (RMSER)

motions are calculated as follows:

1 N
! 1/2
RMSETC = 
x 2 (i) + 
y 2 (i) +
z2 (i) (3.55)
N
i=1

1 N
! 1/2
RMSERC = 
θz2 (i) + 
θx2 (i) (3.56)
N
i=1

1 N 6
1/2
RMSETJ = 
qj2 (i) (3.57)
N
i=1 j =1

1 N 2
1/2
RMSERJ = 
qj2 (i) (3.58)
N
i=1 j =1

where N is the number of the time-samples.

The actual operating motion of the moving platform compared to the desired
trajectory shows a high performance in terms of accuracy. Both trajectories are
plotted in Fig. 3.28. The evolution of the tracking Cartesian error is shown in
Fig. 3.29 zoomed to the interval [90,95] seconds for clarification purposes. The
evaluation of the Cartesian tracking error is done by calculating the root mean square
error over the whole trajectory and is shown in Table 3.3. The root mean square error
values show a very small error over the whole trajectory which validates the high
expected precision of the designed ARROW robot.
74 H. Saied et al.

Fig. 3.28 Evolution of the tracking Cartesian coordinates while following the machining trajec-
tory

Fig. 3.29 Evolution of the Cartesian tracking error while following the machining trajectory
3 Redundant Parallel Robotic Machining Tool 75

Table 3.3 Root mean square of Cartesian and joint tracking errors
Cartesian space Joint space
RMSET [m] RMSER [deg] RMSET [m] RMSER [deg]
7.0574×10−6 0.0170 6.8796×10−6 0.0240

Fig. 3.30 Evolution of the joints tracking error of the parallel module while following the
machining trajectory

The tracking joint errors of the parallel module and the turntable are shown
in Figs. 3.30 and 3.31 respectively. Also it verifies a very high performance of
trajectory tracking machining task, and the evaluations of the root mean square error
are mentioned in Table 3.3 approving the good precision. Indeed, the monitoring
of Cartesian coordinates is available through measuring the joints positions using
encoders, and then transforming these measurements into Cartesian space by kine-
matic mapping (Jaccobian). This is used in the most of the parallel manipulators.
As shown, both joint and Cartesian tracking trajectories achieve very high accuracy
which is an extra evidence to the accurate kinematic model designed for ARROW
robot.
The evolution of the control input forces and torques applied to their correspond-
ing linear and rotative motors are shown in Figs. 3.32 and 3.33 respectively. It is
clear that all the control inputs are in the safe range of the allowed capability
for the actuators. Moreover, it seems that at some instants, the actuators reached
the saturation zone, but the control maintained the stability and performance
thanks to the anti-windup strategy that eliminates overshoots and prevent undesired
oscillatory behaviour which might be dangerous for the structure.
76 H. Saied et al.

Fig. 3.31 Evolution of the joints tracking error of the turntable while following the machining
trajectory

Fig. 3.32 Evolution of the control input forces of the parallel module linear motors

3.6 Conclusions and Future Work

In this work, a 5-DoFs (3T-2R) machining tool named ARROW robot based on
redundantly actuated parallel structure was presented. ARROW robot is featured
with a special mechanism providing a combination between rapidity and precision
with large operational workspace without any type of interior singularities. A
detailed explanation for the kinematics and dynamics associated to ARROW PKM
were introduced in this chapter, in addition to the singularity analysis. In order
to follow up accurately a machining trajectory, an improved PID with computed
feedforward control law has been proposed. Regularization techniques were used to
3 Redundant Parallel Robotic Machining Tool 77

Fig. 3.33 Evolution of the control input torques of the turntable rotative motors

eliminate the undesired antagonistic forces resulting from redundancy and to deal
with the increasing value of the integral term in the PID control block. Thanks to the
adopted anti-windup strategy which preserves the stability of the controller in case
of saturation. Real-time experiments were conducted on ARROW PKM making a
certain machining task. The general performance was evaluated using the root mean
square error criteria justifying a good sufficient accuracy for a machining tasks.
For the future work, one can enhance more the control performance in terms
of precision, motion speed and robustness. Implementing some adaptive model-
based controllers based on the online dynamic calibration technique provides more
robustness against parameters variation and disturbances. From a control point of
view, considering more aspects in a PKM to be enclosed in the control-loop system
can improve the general performance, such as motor drives, actuator dynamics,
transmission system and friction in the articulations.

Acknowledgements This work has been supported by the ARPE ARROW project.

References

Bennehar, M., Chemori, A., & Pierrot, F. (2014a). A new extension of desired compensation
adaptive control and its real-time application to redundantly actuated PKMs. In Intelligent
Robots and Systems (IROS), Chicago, IL, USA.
Bennehar, M., Chemori, A., & Pierrot, F. (2014b). A novel RISE-based adaptive feedforward
controller for redundantly actuated parallel manipulators. In Intelligent Robots and Systems
(IROS), Chicago, IL, USA.
78 H. Saied et al.

Bennehar, M., Chemori, A., Pierrot, F., & Creuze, V. (2015). Extended model-based feedforward
compensation in L1 adaptive control for mechanical manipulators: Design and experiments.
Frontiers in Robotics and AI, 2, 32.
Bruzzone, L., Molfino, R., & Razzoli, R. (2002). Modelling and design of a parallel robot for
lasercutting applications. In International Conference on Modeling, Identification and Control
(IASTED’02), Innsbruck, Austria (pp. 518–522).
Codourey, A., Honegger, M., & Burdet, E. (1997). A body-oriented method for dynamic modeling
and adaptive control of fully parallel robots. In 5th Symposium Robot Control (pp. 443–450).
Davim, J. P. (2008). Machining: Fundamentals and recent advances. London: Springer Science &
Business Media.
eFunda. (2018). Machining: An introduction. In: eFunda, processes, machining. Available via
DIALOG. http://www.efunda.com/processes/machining/
Grange, S., Conti, F., & Rouiller, P. (2001). Overview of the delta haptic device. Eurohaptics, 1,
5–7.
Kerbrat, O., Mognol, P., & Hascoët, J. Y. (2011). A new DFM approach to combine machining and
additive manufacturing. Computers in Industry, 62(7), 684–692.
Kucuk, S. (2012). Serial and parallel robot manipulators – Kinematics, dynamics, control and
optimization. Croatia: Intech.
Kumar, S., & Negi, R. (2012). A new DFM approach to combine machining and additive
manufacturing. In 2nd International Conference on Power, Control and Embedded Systems.
Li, Y., & Xu, Q. (2007). Design and development of a medical parallel robot for cardiopulmonary
resuscitation. IEEE/ASME Transaction on Mechatronics, 12(3), 265–273.
Liégeois, A., Fournier, A., & Aldon, M. J. (1980). Model reference control of high-velocity
industrial robots. In Joint Automatic Control Conference, San Francisco.
Merlet, J. P. (2006). Introduction (chapter 1). Structural synthesis and architectures (chapter 2).
In Gladwell, G. (Ed.), Parallel robots (Solid mechanics and its applications, 2nd ed.). The
Netherlands: Springer.
Muller, A. (2009). Effects of geometric imperfections to the control of redundantly actuated par-
allel manipulators. In IEEE International Conference on Robotics and Automation (ICRA’09)
(pp. 1782–1787).
Muller, A., & Hufnagel, T. (2011) A projection method for the elimination of contradicting
control forces in redundantly actuated PKM. In IEEE International Conference on Robotics
and Automation (ICRA’11) (pp. 3218–3223).
Natal, G. S., Chemori, A., & Pierrot, F. (2012). Dual-space adaptive control of redundantly actuated
parallel manipulators for extremely fast operations with load changes. In IEEE International
Conference on Robotics and Automation (ICRA), Saint Paul, MN, USA.
Natal, G. S., Chemori, A., & Pierrot, F. (2015). Dual-space control of extremely fast parallel
manipulators: Payload changes and the 100G experiment. IEEE Transaction on Control
Systems Technology, 23(4), 1520–1535.
Reyes, F., & Kelly, R. (2001). Experimental evaluation of model-based controllers on a direct-drive
robot arm. Mechatronics, 11, 267–282.
Santibanez, V., & Kelly, R. (2000). PD control with feedforward compensation for robot
manipulators: Analysis and experimentation. Robotica, 19, 11–19. Cambridge University Press.
Shayya, S. (2015). Towards rapid and precise parallel kinematic machines. Ph.D. thesis, Université
Montpellier (Ex UM2).
Shayya, S., Krut, S., & Company, O. (2014). Dimensional synthesis of 4 dofs (3t-1r) actuatedly
redundant parallel manipulator based on dual criteria: Dynamics and precision. In IEEE/RSJ
International Conference on Intelligent Robots and Systems (IROS’14) (pp. 1716–1723).
Shoham, M., Burman, M., & Zehavi, E. (2003). Bone-mounted miniature robot for surgical
procedures: Concept and clinical applications. IEEE Transaction on Robotics and Automation,
19(5), 893–901.
Stewart, D. (1965). A platform with six degrees of freedom. Proceedings of the Institution of
Mechanical Engineers, 180, 371–386. ARCHIVE.
3 Redundant Parallel Robotic Machining Tool 79

Su, Y., Duan, B., & Zheng, C. (2004). Nonlinear pid control of a six-dof parallel manipulator. IEEE
Proceedings-Control Theory and Applications, 151, 95–102.
Yang, H. (2012). Agile mobile manufacturing for large workpieces. Ph.D. thesis, Université
Montpellier.
Youssef, H. A., & El-Hofy, H. (2008). Machining technology: Machine tools and operations. Boca
Raton: Taylor & Francis.
Zhang, Y. X., Cong, S., & Shang, W. W. (2007). Modeling, identification and control of a redundant
planar 2-dof parallel manipulator. International Journal of Control, Automation and Systems,
5, 559–569.
Ziegler, J., & Nichols, N. (1942). Optimum settings for automatic controllers. Transaction of the
American Society of Mechanical Engineer, 64, 759–768.
Chapter 4
Cable-Driven Parallel Robot Modelling
for Rehabilitation Use

Hachmia Faqihi, Maarouf Saad, Khalid Benjelloun, Mohammed Benbrahim,

and M. Nabil Kabbaj

Abstract The aim of this chapter is to presents a Kinematic analysis of a Cable-

Driven Robot for rehabilitation use of human lower limb, by taking into account the
constraints required by the entrainment system and the mobile platform (human
leg). The proposed approach is focused on optimizing the manipulability and
the human performance of the human leg, as being a physiologically constrained
three-link arm. The obtained forward kinematic model leads to define the feasible
workspace of the human leg in the considered configuration. Using an effective
optimization-based human performance measure that incorporates a new objective
function of musculoskeletal discomfort, and the mapping relation between articular
joints actuator, length cables and articular joint mobile platform, the optimal inverse
kinematic (IK) model is obtained.

Keywords Cable-Driven Robot · Rehabilitation · Lower limb · Optimization ·

Inverse Kinematics · Trajectory generation · Minimum Jerk

H. Faqihi () · K. Benjelloun

Automatic and Industrial Informatics Laboratory (LAII), Ecole Mohammadia d’Ingenieurs,
Mohammed V University, Rabat, Morocco
e-mail: [email protected]; [email protected]
M. Saad
Electrical Engineering Department, Ecole de Technologie Superieure, Montreal, QC, Canada
e-mail: [email protected]
M. Benbrahim · M. N. Kabbaj
Faculty of Sciences, Integration of Systems and Advanced Technologies Laboratory (LISTA),
University of Fez, Fez, Morocco
e-mail: [email protected]; [email protected]

© Springer Nature Singapore Pte Ltd. 2019 81

N. Derbel et al. (eds.), New Developments and Advances
in Robot Control, Studies in Systems, Decision and Control 175,
https://doi.org/10.1007/978-981-13-2212-9_4
82 H. Faqihi et al.

4.1 Introduction

Recently, the number of patients with neurological disorders is in perpetual growth.

These disorders are carried out from the neuromuscular diseases, stroke, and
spinal cord injury cerebellar disorders, or the impaired functions of the member
musculature after a serious injury, illness or surgery, which lead among others to
the problems in the lower limb joint and generate atypical gait movement, thereby
reducing patient’s quality of life.
Rehabilitation process based on physical therapy are considered as the best
method to restore patient’s strength, and the mobility fitness. Traditionally physical
therapy sessions were carried out manually, and require a highly qualified staff.
However, due to the increase in the number of patients who need such care,
and the poor performances reached in traditionally rehabilitation process in term
of quickiness, and efficience many researchers have been encouraged to require
the robotic, where a good repeatability, and a precisely controllable assistance,
providing quantitative measures of the subject’s performance and reducing the
required labor of physical therapists, and the financial burden on patients (Pennycott
et al. 2012).
Several robotic solutions have been proposed for rehabilitation use. Their designs
change depending on the desired level of sophistication. However, the major
robotic solutions developed for neurologically impaired patients, use rigid links
and mechanical joints attached to human limbs and joints, which adds extra weight
and inertia to the human limbs, and thereby changes the natural motion dynamics
of the patient. Furthermore, these devices require accurate alignment between
its mechanical joints and the patient limbs, which is difficult or impossible to
accomplish due to the complex geometry of the human body.
To overcome that, a special kind of parallel robots defined as Cable-Driven
Parallel Robots (CDPRs) are increasingly used, especially in physical rehabilitation.
Generally, their entrainment system is the flexible cables which are attached to
fixed base, to move the mobile platform. By changing the lengths of the cables
actuated by the fixed motors and winches, the displacement of the mobile platform
is reached.
In review of the literature, some researchers are been motivated to develop and
design CDPRs to physical rehabilitation use, especially for lower human limb. Each
of them adopts some specific configuration and geometry according to the type of
defficience to restore. These types of devices can be classified as end-effector-based
devices. In this case, one talk about multi-body cable-driven mechanisms, which is
an extension of the basic cable robots where the moving platform is replaced by the
human limbs, as being multi-body system (MBS) (Bryson et al. 2013). Thereby, the
use of cables ensure to this type of robots a low inertia, ability to reach high speeds,
relatively a large workspace, low fabrication costs, a possibility to reconfiguration,
tolerance to anatomical differences among patients, tolerance to misalignment of
apparatus with patient bodies, etc, which are widely recommended in rehabilitation
exercises.
4 Cable-Driven Parallel Robot Modelling for Rehabilitation Use 83

For all of the proposed rehabilitation device in this order, the use of CDPRs
generally, still have some challenges to be taken in consideration. Due to the
compliance of cables, the stiffness analysis of CDPRs becomes a vital concern
(Gouttefarde and Gosselin 2012). Indeed, defficient static stiffness can decrease the
positioning accuracy of CDPRs, and bad dynamic stiffness characteristics can lead
to vibration and long settling time.
Among challenges also for such mechanisms are the risk of interference between
the cables and the links, or between the cables themselves (Marler et al. 2009).
Though, the major challenge is ensuring their ability to have the best force
distribution in all cables (positive tension), known as tensionability condition
(wrench closure or force closure) (Rezazadeh and Behzadipour 2011).
The proposed CDPR for rehabilitation of lower limb in this study, has a specific
design, compared to those existed. It allows to applying rehabilitative motion for
lower patient limbs, following two possible embodiments. The first embodiment,
with standing position of the patient, requires the control of three degrees of
freedom in the sagittal plane for its right and left lower limbs. However, the second
embodiment, where the patient is in an elongated position, allows the movement of
one leg at a time in the seven degrees of freedom.
The study of this kind of mechanism require not only to take into account the
challenges faced by the usual CDPRs, but also the constraints required in medical
applications. Generally, the study of robotic real application require to resorting to
the Inverse Kinematic (IK) task, to find a configuration at which the end-effector
of the robot reaches given point in the task space. Specially for human body
application, the IK resolution complexity is enhanced with the increased Degrees
Of Freedom (DOF). The application of the classical IK methods, remains just viable
mathematically, do not take into account the physiological feasibility and biofidelity
of human posture, and suffer from numerical problems (Abdel-Malek 2004).
Optimization based approaches can be suitable ways to overcome the above
mentioned problems. It refers to predict the realistic posture of human limb in
its feasible workspace. As any optimization problem, for the posture prediction
problem, the joint angles of the human leg are considered as the design variables,
the constraints are considered according to physiological feasibility and motion
precision, and for the objective function, the human performance measures are used.
There are many forms used in the literature to define the human performance
measure, such as physical fatigue defined as reduction of physical capacity. It
is mainly the result of three reasons: magnitude of the external load, duration
and frequency of the external load, and vibration (Chen 2000). However, for the
movements required low speed such as rehabilitation exercises, the physical fatigue
is not so significant. Indeed, the required movements can lead to some human
discomfort (Abdel-Malek 2004), where its evaluation may vary from person to
person, such as potential energy (Mi et al. 2009), torque joints, muscle fatigue, or
perturbation from a neutral position (Spong et al. 2006).
In the current work, the passive modeling approach is adopted, with the thorough
description with a customization of anthropometric parameters, and taking into
account, the different constraints which can be appears in this system. Arising, a
84 H. Faqihi et al.

unique and sophisticated biomechanical model, compared to those found in the

literature. The obtained biomechanical model characterize a class of nonlinear
systems with unknown dynamics.
To better illustrate these aspects, the remainder of the paper is organized as
follows: In the first the robot description will be presented according to the
associated configurations. In the following the human lower limb consideration
will be discussed. Therefore the forward kinematics has been developed for the tow
configurations where motions of the human lower limb occur with three and seven
degrees of freedom. According to that, in following section, the feasible workspace
have been established. Thereby, in the followed section, the new optimal posture
prediction has been described and thereby applied on the human lower limb for the
provided motion configuration. To check the effectiveness of the proposed approach,
a simulation model has been developed using Matlab package. To sum up, the results
of the study are outlined in finally section.

4.2 Modeling Robot

The use of Cable Driven Robotic Rehabilitation (CDRR) of lower limb, is based on
eight cables, attached to the fixed winches, then, to the lower limb (Faqihi et al.
2016). The length of the cables is adjusted by motors to reproduce the desired
rehabilitation movement. For a given position, the robot must calculate and adjust
the length of the necessary cables so that the leg reaches the desired place in
space. These calculations must be carried out dynamically and continuously during
a movement.
The robot is designed to apply rehabilitation exercises according to two possible
configurations:

4.2.1 First Configuration

In the first configuration, the patient is in a standing position. This configuration is

adopted to imitate the walking movement, in each legs can move according to three
degrees of freedom, in the sagittal plane, as illustrated in Fig. 4.1.
With the eight used cables, the robot in this configuration has one Degree
Of Redundancy (DOR) for each legs at a time, which will thereafter serve as a
necessary condition to maintain the tensionability of the mechanism.
The provided motion of the human leg can be modeled by three revolute degrees
of-freedom: A hip extension-flexion degree-of-freedom, a knee extension-flexion
degree-of-freedom, and an ankle dorsiflexion-plantar degree-of-freedom.
4 Cable-Driven Parallel Robot Modelling for Rehabilitation Use 85

Structure porteuse Moteur Siège Poulie

2 jambes à la fois Support jambe Support pied Câbles

2 × 3 degrés de liberté
8 càbles et moteurs

Fig. 4.1 First configuration of CDRR Legend in Frensh !!!

Structure porteuse Moteur Siège Poulie

1 jambes à la fois Support jambe Câbles

Support pied
7 degrés de liberté
8 càbles et moteurs

Fig. 4.2 Second configuration of CDRR Legend in Frensh !!!

4.2.2 Second Configuration

For this configuration, the patient is in an elongated position, which ensure a single
leg movement at the time in the three dimensional space, as illustrated in Fig. 4.2.
With this configuration also, the robot has one DOR.
The motion of the human leg is provided for sevens degree-of-freedom,
defined as: 3 DOF hip (extension-flexion degree-of-freedom, abduction-
adduction degree-of-freedom, and inversion-eversion degree-of-freedom), 1 DOF
knee (extension-flexion degree-of-freedom), 3 DOF ankle (extension-flexion
86 H. Faqihi et al.

Fig. 4.3 Coordinate systems

on 7 DOF of the human leg

degree-of-freedom, abduction-adduction degree-of-freedom, and inversion-eversion

degree-of-freedom), as depicted in Fig. 4.3.
In general, the flexibility between the two configurations, is provided with an
appropriate cable routing with respect to the positions of the pulleys and the
attachment points of the connected cables.
According to the previous studies which have been conducted on Cable-Driven
Parallel Robot for rehabilitation of lower limbs use, the proposed robot has the
particularity and the uniqueness of control the three joints of the lower limb,
especially the hip, the knee and the ankle. Furthermore, and according to the second
configuration of the robot, the leg movement can be done in three dimensional space,
which extend the spectrum of use CDPR for different rehabilitation exercises of
lower limbs.

4.3 Human Lower Limb Considerations

As described above, the studied CDRR lead to move human lower limb by
entrainment system of cables according to tow possible configurations. In order to
give a credible model of this robot, it is necessary to give a precision of its mobile
platform.
Generally, the human body is a complex system, its biomechanical modeling
represents a simplification of its real operating. The introduction of assumptions is
necessary in this order, which are selected according to the desired performances.

4.3.1 Modeling Hypothesis

The model adopted for the lower limb based on three segments to model its
anatomical structure (bone structure): thigh, shank and foot considered as the length
between ankle and metatarsal.
4 Cable-Driven Parallel Robot Modelling for Rehabilitation Use 87

Since, exercised movement on the lower limb, during the rehabilitation sessions,
is of small variation of acceleration. In this order, some hypothesis were fixed, and
which will be valid for all structures:
H1: The body segments are considered to be non-deformable rigid bodies. This
assumption enables the application of the mechanical rigid body approach. As
the body segments are rigid, their center of mass and their moments of inertia
does not change during movement.
H2: The joints assumed to be perfect mechanical connections.
H3: The actions that generate movement are localized in the joint centers.
H4: The body mass is assumed concentrated on the center of mass.
H5: The length of each segment, remains constant during the period of executing
movement.
H6: To ignore artifacts soft tissue, the movement marrow masses (skin and fat) is
negligible and does not affect the inertial properties during the execution of
the movement.
H7: The joints are defined with viscoelastic behavior in anatomical axes of the
joint. In addition, these axes are fixed, thus the translations of a bone on the
other (knee) are also negligible.
H8: The muscular action is due to a single muscle group. Therefore, there are no
antagonist muscles that oppose the movement created by the agonists.
The connection of all three segments is ensured naturally by ligaments and muscles,
and should be kinematically redundant to ensure biofidelity of the human leg motion
(Winter 2009).
For a Kinematic analysis, the human leg is modeled by a kinematic chain
of rigid bodies, interconnected by kinematic joints, which can be either simple
or complex according to required physiological behavior, and thus the degree of
freedom associated with the possible joints (9).
According to the second configuration, the leg motion is conduct with seven
degree of freedom, distributed as depicted in Fig. 4.3.

4.3.2 Anthropometric Parameters

The modeling of body segments must take into account some anthropometric param-
eters. In order to customize the model, accurate measurements of the anthropometric
parameters are required and can be obtained from statistical tables proposed in
Winter (2009). It refers to adopt a proportional anthropometric model, based on
statistical regression equations to estimate these segmental inertial parameters (PIS),
from a comprehensive knowledge of the body such as: size and body weight. The
Table 4.1, provides the necessary information to calculating inertial parameters. For
this study, physical length segments will be used. They can be computed using total
body height (H), where W is the total body weight, and L is the total body length;
88 H. Faqihi et al.

The segment moment of inertia around its center of mass Icm , can be expressed
by:

Icm = Ws × (Ls × R)2 (4.1)

4.3.3 Range of Motion

To ensure users security during rehabilitation session, physiological safety of

humans lower limb must be considered in the design of robot. Indeed, the suspension
of the human leg according to used cable, allows to maintain the patients balance
during training, when he is unable to do it. In the other hand, defining ROM of the
human lower limb model, is not limited to the designed mechanical structure, but
also to the human physiological factors, such as the age, body build, gender, health
condition (Chaffin et al. 1992), so that the Maximum Range Of Motion (MROM)
and Maximum Speed (MS) must be taked into account, to increase the physiological
safety.
These specifications are used as constraints of the desired trajectory, ensuring that
MROM and MS will not be exceeded during motion exercise. The designed robot,
can assist patients to perform joint training, gait training, or other desired trajectory
tracking in hip, knee and ankle joints, according to the controller’s performance.
Due to the biomechanical constraints of human body motion, the bounds of the
joint variables are fixed, which define the Range of Motion (ROM).
Generally, the ROM of human legs, based on a previous study by Hernandez
et al. (2011) is given in Chaffin et al. (1992).

4.3.4 Comfort Zone

To ensure the comfort motion of the human body, each joint variable can be defined
by its comfort zone, which must belong to the range of motion (ROM) of the
associated joint variable.Referring to the literature, the comfort zone represents 35%
of the range of motion (ROM). The center of the comfort zone, qic , is calculated by
the following expression (Glowinski and Krzyzynski 2016):

qiC = 0.5(qicz
u
+ qicz
l
) + qih (4.2)

u , and q l are respectively, the upper and lower angles of the comfort zone,
where qicz icz
for the i th joint variable associated. qih is the home position angle of the i th joint
variable. Generally, the home position angle can differ from tested tasks (standing,
recumbent, seating,...).
4 Cable-Driven Parallel Robot Modelling for Rehabilitation Use 89

4.4 Kinematic Analysis of CDRR

The studied robot is a mechanical system that transforms the movements of its
actuators on movement of the platform (lower limb). In order to describe the
kinematics of the cable mechanism, the following notation is adopted:
• The joint coordinates of the actuators (angles of rotation of the motors) are
denoted by q = [q1 , . . . qm ], where m is the number of actuators.
• The articulare coordinates of the platform are noted θ = [θ1 , . . . θd ], where d is
the Degree Of Freedom of the platform.
• The operational coordinates describing the platform position, are defined in R n
and denoted by x = [x1 . . . xn ], where n the number of DDLs of the platform.
In other words, the robotic system transforms joint forces: the cable tensions,
noted t = [t1 , . . . tm ] ∈ R m , in operational forces f = [f1 , . . . fn ] ∈ R n . These
efforts represent the cable effects on the platform.
Since the platform represents a multi-body, we will therefore have j = 1 . . . m
referencial marks RPj associated to each segment of the platform.
In general, CDPR with m cables, whose geometry is illustrated in Fig. 4.4, has m
output points of the cables, denoted Ai on the base and m points of attachment of
the cables on the platform, Bi .
The coordinates of these points are grouped in the vectors ai and bi , respectively.
Where ai is expressed in the fixed frame of the robot, while bi is expressed in the
mobile frame attached to the platform. The vector p gives the Cartesian position
of the platform of geometric center P . The orientation of the fixed reference
with respect to the fixed coordinate system is given by the orthogonal matrix of
rotation Q. The length of Cable i is denoted by li . By writing ei a unit vector, we
can write:

Fig. 4.4 General

Architecture of CDPR
90 H. Faqihi et al.

Ai Bi = li ei (4.3)

The problem of inverse geometry model ifor CDPRs consists to determine the
lengths of the cables l, required for a precise displacement of the actuators described
by their articular coordinates q, from the vector describing the placement (position
and orientation) x of the mobile platform such as:

q = f (x) (4.4)

In the kinematic chain that connects the base to the platform, only the m cables
connect the platform to the fixed base of the robot. From Eq. 4.3, we can write
directly:

li = Ai Bi  = x + Qbi − ai  (4.5)

It is assumed that there exists a linear relationship between the cable length li ,
i = 1 . . . m and the rotation angle of the motor qi expressed in meters and radians,
respectively (El-Ghazaly et al. 2014) :

li = ±ri qi (4.6)

where ri is the ratio expressed by:

"
ρi2
ri = κi ei2 + (4.7)
2π

with κi is the transmission ratio of the ith winder with the ith motor, ei and ρi the
radius and the pitch of the ith winder respectively. Note that κi and ρi are constants.
The relation (4.7) is true in the case of an inextensible cable whose winding
system guarantees that ei remains constant over time (the cable does not curl up on
itself). Sign of the said relation depends on the direction of rotation of the ith motor
and on the winding direction of the cable. The sign must be therefore established
according to the particular mechanics of the robot.
The finally relation is given by:

1
qi = x + Qbi − ai  (4.8)
ri

The Jacobian matrix J of CDPR is the matrix formed by the gradient composants li
of l, which connects operational speeds dotx of the platform to joint speeds q̇.
Assuming that the cables are taut, for a robot with m DDL, the J is expressed
by: Noted that l˙ = [l˙1 , l˙2 , . . . l˙m ]T the linear velocity vector dotq and cable lengths
dotl are connected by equation:

l˙ = R q̇ (4.9)
4 Cable-Driven Parallel Robot Modelling for Rehabilitation Use 91

where R is diagonal matrix defined by diag{ri } = {r1 , . . . rm }. According to the

operational platform speeds:

l˙ = J (x)ẋ (4.10)

Hence the relationship between joint speeds (speeds of rotation of the motors) dotq,
and the operational speeds dotx of the platform as follows:

q̇ = R −1 J (x)ẋ (4.11)

In the studied CDPR the mobile plate-form is MBS, characterized by their

articulation coordinates θi . For safety and comfort purpose of patient it would be
necessary to take into account the constraints of θi to determine f (x) defined in
Eq. 4.4.
Therefore, the main goal is to find the displacement of the actuators qi for a
given position x of the end-effector, by defining the associated articular joints of the
plateforme θi according to the discussed constraints, and thereby associated cable
lengths li .
In this order, the remain of this section focuses on introducing a general
optimization-based formulation to predict the θi from given end-effector position
x, and deduce therefrom the articulation coordinates θi . The new proposed idea,
is to use an objective function incorporating three factors that contribute to
musculoskeletal discomfort as human performance measure.

4.4.1 Forward Kinematic Analysis

The forward kinematic (FK) model in n plane, defined by g function, can determine
the pose of the end-effector (xj , j = 1 . . . n), from given articular joints variables of
the paltforme (θi , i = 1 . . . DOF ). It is a necessary step in the kinematic analysis
process.

xj = g(θi ) (4.12)

For the rigid bodies robotic systems, several methods can be used to resolve this
problem.
Despite of the human body complexity, for all practical purposes, it has
been shown that approximated modeling of gross human motion, in order to
ensure human motion simulation, ergonomic analysis, or rehabilitation process,
can be achieved using homogenous transformation matrices method, and the
Denavit Hartenberg (DH) representation, based on appropriate kinematic coordi-
nates. Indeed, the DH method provides an adequate, and systematic method for
embedding the local coordinate systems for each link.
92 H. Faqihi et al.

Table 4.1 DH human leg Link(i) αi ai di qi

parameters for 3 DOF
1 0 a1 0 q1
2 0 a2 0 q2
3 0 a3 0 q3

Table 4.2 DH human leg Joint(i) αi ai di qi

parameters for 7 DOF
1 − π2 0 0 q1
π
2 2 0 0 q2
3 0 a1 0 q3
4 0 a2 0 q4
5 0 a3 0 q5
6 − π2 0 0 q6
π
7 2 0 0 q7

The forward kinematic human leg model is developed from the DH parameters
(Dombre and Khalil 2010) where each degree-of-freedom can be modeled as
a revolute joints. The DH parameters are depicted in Table 4.1 for the first
configuration and Table 4.2 for the second configuration, from the defined kinematic
coordinates.
From the provided D-H parameters, the forward kinematic model can be
computed using the transfomation matrix, given in Eq. 4.13. Generally, the transfor-
mation matrix is the relationship expression between two consecutive frames i − 1
and i, which depends on the described parameters (θi ,αi ,ai ,di ) given in the above
tables.
⎛ ⎞
cθi −sθi cαi sθi cαi αi cθi
⎜ sθi cθi sαi −cθi sαi αi sθi ⎟
i−1
Ti = ⎜
⎝ 0
⎟ (4.13)
sαi cαi di ⎠
0 0 0 1

The global transformation matrix for the first configuration can be expressed by:
0
T3 =0 T1 .1 T2 .2 T3 (4.14)

and for the second one:

0
T7 =0 T1 .1 T2 .2 T3 .3 T4 .4 T5 .5 T6 .6 T7 (4.15)

From Eqs. 4.14 and 4.15, the forward kinematic model is given by the following
equations respectively for the first and second configurations:
   
x c123 a3 + a2 c12 + a1 c1
= (4.16)
y s123 a3 + a2 s123 + a1 s1
4 Cable-Driven Parallel Robot Modelling for Rehabilitation Use 93

where: c123 = cos(q1 + q2 + q3 ), s123 = sin(q1 + q2 + q3 ), c12 = cos(q1 + q2 ),

s12 = sin(q1 + q2 ), c1 = cos(q1 ), and s1 = sin(q1 ).
⎛ ⎞ ⎛ ⎞
x a2 c1234 + a1 c123 + a3 c1234 c4 − a3 s1234 s5
⎝ y ⎠ = ⎝ a2 s1234 + a1 s123 + a3 s1234 c4 + a3 c1234 s5 ⎠ (4.17)
z −a1 s23 − a2 s234 − a3 s234 c4 − a3 c234 s5

where: c1234 = cos(θ1 + θ2 + θ3 + θ4 ), s1234 = sin(θ1 + θ2 + θ3 + θ4 ), c123 =

cos(θ1 + θ2 + θ3 ), s123 = sin(θ1 + θ2 + θ3 ), c12 = cos(θ1 + θ2 ), s12 = sin(θ1 + θ2 ),
c4 = cos(θ4 ), s4 = sin(θ4 ), and s5 = sin(θ5 ).
In terms of velocity and acceleration, the forward kinematic is given by:

ẋj = J θ̇i , ẍj = J θ̈i + J˙θ̇i (4.18)

where ẋj , ẍj , represent respectively the end-effector velocity, the end-effector
acceleration in task space, J and J˙, represent the Jaccobian matrix of the system and
its derivative. Finally, and θ̇i and θ̈i represent respectively the end-effector velocity
and acceleration in joint space.

4.4.2 Feasible Workspace

In order to analyze the feasible workspace associated to the human lower limb,
where we plot the different possible positions of foot which can be achieved, the
direct kinematic model f (θi ) is used. Thus, the workspace can be defined as the set
of all the possible positions in the task space according to the ROM, as following:
# $
Ep = θi ∈ ROM/ xj = f (θi ) (4.19)
j

Using the appropriate forward kinematic model given in Eq. 4.17, and the ROM
described, the feasible workspace can be plotted for the motion provided according
to the used configuration.

4.4.3 Proposed Optimization Approach

4.4.3.1 Problem Formulation

The optimal posture prediction is considered to be a constrained optimization

problem (CO), using a constraint to find a realistic configuration.
Generally, the CO problem can have equality and/or inequality constraints
according to the described problem, and the objective function, which requires some
94 H. Faqihi et al.

assumptions according to the continuity and differentiability. In that fact, in the

following the optimization problem model are described.

4.4.3.2 Design Variables

The design variables represent in this case, the joint variables θi , i = 1 . . . DOF ,
following the used configuration. However, the initial guess consists of determining
initial values, for the design variables. Indeed it should be suitable to use a feasible
point. In this study, an initial guess is used to satisfy the ROM.

4.4.3.3 Constraints

The first constraints consider the difference between the current end-effector
position, velocity, and acceleration, and the given target position, velocity and
acceleration respectively in cartesian space, as following:

computed
||xj (θi ) − xjdesired (θi )|| ≤ ε1
computed
||ẋj (θi ) − ẋjdesired (θi )|| ≤ ε2 (4.20)
computed
||ẍj (θi ) − ẍjdesired (θi )|| ≤ ε3 (4.21)

computed
where ||.|| define the Euclidean norm. The end-effector position xj hits a
predetermined target point xj desired in cartesian space, within a specified tolerance
ε1 a small positive number that approximates zero, similarly to end-effector velocity
and acceleration.
It should be noted that, determining the end-effector position, velocity and
acceleration are ensured using the forward kinematic model.
On the other hand, each joint variable is constrained to lower and upper limits,
represented by θil and θiu , respectively. These limits ensure that the human posture
does not assume an unrealistic position to achieve the target point.
To more rigorous biofidelity end, we can choose that each joint variable is
constrained to lie between upper and lower angles of the comfort zone, designed
u and θ l respectively.
by θicz icz

l
θicz ≤ θi ≤ θicz
u
(4.22)

Finally, the constraints in term of velocity and acceleration limits are used, as
following.

θ̇il ≤ θ̇i ≤ θ̇iu (4.23)

θ̈il ≤ θ̈i ≤ θ̈iu (4.24)
4 Cable-Driven Parallel Robot Modelling for Rehabilitation Use 95

where θ̇il and θ̈il are the lower limit of velocity and acceleration, θ̇iu , and θ̈iu are
the upper limits of velocity and acceleration respectively. These limits are fixed
according to biomedical studies, where the physiological and health state of patient
are taken into account.

4.4.3.4 Cost Function

As described previously, the posture prediction requires a human performance

measure, where its rigorous choice ensures the optimal realistic posture.
To this end, the modeling musculoskeletal discomfort is used as human perfor-
mance measure, which can be somewhat ambiguous, as it is a subjective quantity,
thus its evaluation may vary from one person to another (Abdel-Malek 2004).
According to the last described researches (Abdel-Malek 2004; Mi et al. 2009),
in this order, different forms of human performance measures have been adopted,
but it often results in postures with joints extended to their limits, and thus to some
uncomfortable positions.
As remedy, we can add factors associated with moving while joint variables are
near their respective limits in terms of position, velocity and acceleration. In this
respect, it is possible to incorporate different factors that contribute to discomfort.
The first factor, is referred to the tendency to move different segments of the
body sequentially. The second factor, is referred to the tendency to gravitate to a
reasonably comfortable position. Finally, the discomfort associated to the motion
while joints are near their ROM in term of position, velocity and acceleration,
expresses the third factor.
According to the previous studies, the proposed objective function is similar to
that adopted by Yang et al. (2004) applied for the upper limb. However, there is
currently no research focused on prediction of human leg posture, by applying this
form of discomfort, with the restriction particularity of 15%, taking into account the
end-effector position, velocity and acceleration.
In order to incorporate the first factor, we can find several strategies which induce
motion in a certain order, or with higher weighted joints than others.
Consider qic the comfortable position of ith joint variable, measured from the
home configuration defined by qih = 0. Then, conceptually, the displacement from
the comfortable position for a particular joint position is given by: |qi − qic |.
However, to avoid numerical difficulties and non-differentiability, we can use:
(qi − qic )2 .
Generally, the terms should be combined using a weighted sum wi , to emphasize
the importance of particular joints depending on the characteristics of each patient.
Thereby, the joint displacement function is given as follows:

fdisplacment = wi (qi − qic )2 (4.25)

i

The weights are used to approximate the lexicographic approach (Marler et al.
2009).
96 H. Faqihi et al.

In order to incorporate the tendency to gravitate to a reasonably comfortable

neutral position, each term in Eq. (4.25) is normalized, as described in the following:

qi − qic
Δqi = (4.26)
qiu − qil

Each term of (Δqi )2 of this normalization scheme is considered as a fitness function

with each individual joint and has normalized values, which lie between zero and
one.
The principal limitation of this approach often results in postures with joints
limits extended, thereby an uncomfortable joint. In this order, the third factor is
introduced, which defines the discomfort of moving while joints are near their
respective limits. This factor requires to add some designed penalty terms to increase
significantly the discomfort where joint values are close to their limits.
– Barrier Penalty Function
Generally, the new designed penalty term P (d) is a barrier penalty function
(Eschenauer and Thierauf 1989), of d argument, expressed by:

P (d) = [sin(ad + b)]p (4.27)

The P (d) function is adapted to penalize any number d, considered as normalized

parameters, which is approaching zero at some number value.
The proposed idea is that the penalty term remains zero until the d value reaches
d ≤ 0.15, which defines the desired curve data.
Thereby, the parameters a, b, and p of the basic structure of barrier penalty
function are fitted to reach the desired curve data.
The penality function P (d) is computed, according to the desired curve data
as explained previously, by using Curve Fitting Package in Matlab, as depicted in
Fig. 4.5.
The obtained parameters of the penalty function are given by : a = 2.5, b =
7.855, p = 100. Thereby the final expression of the penality term is given by:

P (d) = [sin(2.5d + 7.855)]100 (4.28)

According to the three described factors, the consequent discomfort function is

obtained as follows:
DOF
fdicomf ort = [wi (Δqi )2 + P (Rpui ) + P (Rpli ) (4.29)
i=1
+P (Rvui ) + P (Rcli ) + P (Raui ) + P (Rali )]

where P (Rpli ) and P (Rpui ), P (Rvli ) and P (Rvui ), P (Rali ) and P (Raui ) are the
penality terms with joint values that approach their lower limits, and their upper
limits, respectively for end-effector position, velocity and acceleration.
4 Cable-Driven Parallel Robot Modelling for Rehabilitation Use 97

Fig. 4.5 Penality term P(d)

qiu − qi qi − qil
Rpui = ; Rpli = (4.30)
qiu − qi l qiu − qil
q̇iu − q̇i q̇i − q̇il
Rvui = ; Rvli =
q̇iu − q̇i
l q̇iu − q̇il
q̈iu − q̈i q̈i − q̈il
Raui = ; Rali =
q̈iu − q̈i
l q̈iu − q̈il

The penalty term that depends on parameter d, remains zero as long as the upper or
lower joint value does not reach 15% of its range, as depicted in Fig. 4.5.

4.4.3.5 Constrained Optimization Model

From the described design variables, constraints and cost function, the final
optimization problem can be formulated, as the following:
98 H. Faqihi et al.

⎧
⎪
⎪ min : fDiscomf ort (qi )
⎪
⎪
⎪
⎪ subject to : ||xj
computed
(qi ) − xjdesired (qi )|| < ε
⎪
⎪
⎪
⎪ computed
||ẋj (qi ) − ẋjdesired (qi )|| < ε
⎨
computed
⎪ ||ẍj (qi ) − ẍjdesired (qi )|| < ε
⎪
⎪
⎪
⎪
l ≤ q ≤ qu
qicz
⎪
⎪
i icz
⎪
⎪ q̇il ≤ q̇i ≤ q̇iu
⎪
⎩
q̈il ≤ q̈i ≤ q̈iu

4.4.4 Proposed Optimal Solution

The inverse kinematic optimization problem formulated previously is a Nonlinear

Optimization Problem (NLP). Several numerical solutions of constrained nonlin-
ear optimization problems have been presented in the literature. For resolution
feasibility Sequential Quadratic Programming (SQP) is considered to be suitable
method (Gill and Saunders 2002) to resolve the proposed optimization problem.
The algorithm resolution is divided into three main steps, as presents in the flowchart
Fig. 4.6:
– Step1: The first step begins by the initialization, where it is necessary to
determine the total body height of a subject and to calculate the length of thigh,
shank, and foot segments. Then, we fix the initial joint position, velocity and
acceleration variables, and calculate the initial guess as being initial effector
position, velocity and acceleration by using forward kinematics, according to
the used configuration. Next, taking into account the joints constraints, the
workspace and the comfort zone of each joint are computed.
– Step2: Giving the end position, velocity and acceleration coordinates, in the
second step, we check if the position is in the workspace, else it is necessary to
find new coordinates.
– Step3: SQP optimization technique is applied in this step to find the optimal
postures. This is performed using Matlab fmincon constrained function. The
obtained solution is a matrix with position, velocity and acceleration angles.
Taking into account the comfort zone of each joint, the obtained angles are the
most comfortable.

4.4.5 Trajectory Generation

In rehabilitation robots, the reference trajectory must be predefined as a human

limb motion practiced during activities of daily live. Indeed, motion therapy
can be carried out in different modes including passive, active, active-resistive,
4 Cable-Driven Parallel Robot Modelling for Rehabilitation Use 99

1 step

Initialization

Initial parameters

Workspace

The optimum user angles of each joint

2 step

The final effector coordinates

Is the end position located in the workspace? no

yes

3 step

Optimization fmincon

The final joint angles satisfy? no

yes
New parameters

Fig. 4.6 Proposed flowchart solution

active-assistive, and bilateral exercises, which differ depending on the degree of

patient involvements. Selecting the proper mode strategy requires an appropriate
rehabilitation robot choice, with concerned patients.
To determine the appropriate trajectory for the movement of the rehabilitation
robot, there are several methods such as a prerecorded trajectory obtained by gait
analysis, and a prerecorded trajectory during therapist assistance, which require data
use, and modelling the trajectory based on normative movements which can be
based on kinematics and/or dynamics constraints during the path motion in terms
of fitting more realistic motion (Rastegarpanah and Mozafar 2016). Generally, it
is desirable to use reference trajectories ensuring the feasibility and biofidelity of
rehabilitation session.
According to passive rehabilitation exercises, the desired trajectory used to check
the effectiveness of the proposed approach is defined from Clinical Gait Database
100 H. Faqihi et al.

Desired trajectory
2

X(cm)
1

0
0 10 20 30 40 50 60 70 80 90 100
position index
0.28

0.26
Y(cm)

0.24

0.22
0 10 20 30 40 50 60 70 80 90 100
position index
0.3
Z(cm)

0.2

0.1
0 10 20 30 40 50 60 70 80 90 100
position index

Fig. 4.7 Desired trajectory

(CGD) generated during rehabilitative motion. The used data is reported in relation
to the right-handed global coordinate system, where +z points up and subject is
walking in +x direction (Fig. 4.7).

4.5 Simulation Results and Discussion

In order to show the effectiveness of the proposed approach, this section presents a
simulation results of the developed Kinematic analysis of the Cables-Driven Robot
for Rehabilitation use, for the tow discussed configurations.
The validation test is held on a patient of 1.80 m height. Using the anthropometric
associated data, the developed forward kinematic model and the described ROM, the
associated workspaces are plotted, as shown in Figs. 4.8 and 4.9.
As described above, the main goal is to find the associated joint coordinates of
the actuators qi defining the length cables li , which lead to move the end-effector to
the given position xi .
To define the mapping between the end-effector position xi and the associated
articular joints θi the developed optimal posture algorithm is used. Indeed, desired
end-effector position for the two configurations are fixed to lie their associated
workspace.
4 Cable-Driven Parallel Robot Modelling for Rehabilitation Use 101

Fig. 4.8 First configuration workspace

Fig. 4.9 Second configuration workspace

102 H. Faqihi et al.

The initial value is fixed in joint space according to the feasible biomechanical
posture for the used configuration as q1initguess = [0, 0, 0], and for the second
configuration q2initguess = [0, 0, 0, 0, 10, 0, 0] (in degree), where the result of the
optimization is usually sensitive to the initial guess.
The remain parameters of the optimal posture algorithm, are given by ε =
0.0001, the weight wi for the joints variables of the lower limb, are defined in Marler
et al. (2009).
One of the most important factors in the development of any optimization
problem is the selection of the appropriate fitness function, and well defined
constraints according to the problem complexity as described in this paper. Using
SQP algorithm resolution for the proposed optimization problem formulation the
inverse kinematic model is obtained.
The optimum kinematic parameters are obtained using the desired motion in
terms of end effector position for the two configurations, and optimization designed
routine, where the objective function, and constraints are defined, as described
previously. From the provided checking function, the end-effector coordinates of
desired motion are validated, and then the optimal position joints for the tow
configurations are predicted in Figs. 4.10 and 4.11.
In order to check the effectiveness of the developed optimization algorithm for
the first configuration in sagittal plane, another objective function, already used in
the same context (Glowinski and Krzyzynski 2016), is considered. The results of
running optimization routine using tow objective functions are shown in Fig. 4.12.

Fig. 4.10 Obtained joint-space position (First configuration)

4 Cable-Driven Parallel Robot Modelling for Rehabilitation Use 103

Fig. 4.11 Obtained joint-space position (Second configuration)

Fig. 4.12 Obtained results (First configuration)

104 H. Faqihi et al.

Table 4.3 Optimal posture Optimization routine ERMS

comparaison
F1 2.0811 × 10−4
F2 9.5271 × 10−4

Fig. 4.13 Obtained results (Second configuration)

The error in each axis of motion is depicted for the used objective functions,
where F1 is the described objective function, and F2 the objective function used
in Glowinski and Krzyzynski (2016).
Using the root-mean-squared RMS error expression given as:
%
&
&1 N
ERMS =' ||error(x)||2 (4.31)
N
i

where N denotes the number of position points using in the simulation. The
quantitative analysis is presented in Table 4.3.
From the above comparison, we have found that the developed optimization
algorithm presents the lowest error.
For the second configuration the cartesian trajectory from the obtained inverse
kinematic model, have been predicted as shown in Fig. 4.13, which presents a
good result according to the used desired trajectory. This simulation test represent
to our knowledge the first study carried out in the all three plane of human leg
configuration,
4 Cable-Driven Parallel Robot Modelling for Rehabilitation Use 105

4.6 Conclusion

A new general optimization-based formulation for optimal kinematic analysis for

the human lower limb has been developed. The proposed method is developed
specially in three dimensional space, in terms of the kinematic parameters, using
an objective function incorporating three factors that contribute to musculoskeletal
discomfort as human performance measure.
The performance measure is referred to the tendency to move sequentially
all segments of the human leg, the tendency to move leg while joints are near
their ROM, and the discomfort associated with gravitating around a reasonably
comfortable position.
This new form of objective function is developed principally for rehabilitation
use, where physiological patient constraints need to be taken into consideration.
In this order, according to the physiological constraints and the developed forward
kinematic model, the feasible workspace is presented. To validate the feasibility and
the effectiveness of the proposed kinematic method to predict the inverse kinematic
model of 3D space human leg configuration, a reference trajectory is generated to
be suitable in rehabilitation case, and thereby, applied in the proposed algorithm
solution.
The simulation results, were present an optimal joint space parameters defining
the inverse kinematics.
However, this study still valid for the static purpose, and can be improved
according to a large definition of discomfort according to dynamical parameters
in term of muscle, fatigue... which can be developed in the future research.

References

Abdel-Malek, K., Yang, J., Yu, W., & Duncan, J. (2004). Human performance measures:
Mathematics. Iowa: University of Iowa.
Bryson, J. T., & Agrawal, S. K. (2013). Methodology to identify and analyze optimal cable config-
urations in the design of cable-driven serial manipulators. In International Design Engineering
Technical Conference and Computers and Information in Engineering Conference, 37th ASME
Conference Mechanisms and Robotics IDETC/CIE.
Chaffin, D. B., Andersson, G. B. J., & Martin, B. J. (1992). Occupational biomechanics. New York:
Wiley.
Chen, Y. (2000). Changes in lifting dynamics after localized arm fatigue. International Journal of
Industrial Ergonomics, 25(6), 611–619.
Dombre, E., & Khalil, W. (2010). Modeling, performance analysis and control of robot manipula-
tors. London: Wiley.
El-Ghazaly, G., Gouttefarde, M., & Creuze, V. (2014). Adaptive terminal sliding mode control of a
redundantly-actuated cable-driven parallel manipulator: Cogiro. In Proceedings of the Second
International Conference on Cable-Driven Parallel Robots. Springer.
Eschenauer, H. A., & Thierauf, G. (1989). Discretization methods and structural optimization,
procedures and applications. In Proceedings of a GAMM-Seminar October 5–7, 1988, Siegen,
FRG (Lecture notes in engineering). Springer.
106 H. Faqihi et al.

Faqihi, H., Saad, M., Ben Jelloun, K., Benbrahim, M., & Kabbaj, M. N. (2016). Tracking trajectory
of a cable-driven robot for lower limb rehabilitation. International Journal of Electrical,
Computer, Energetic, Electronic and Communication Engineering, 10(8), 1036–1041.
Gill, P. E., Murray, W., & Saunders, M. A. (2002). Snopt: An SQP algorithm for large-scale
constrained optimization. SIAM Review, 47(1), 99–131.
Glowinski, S., & Krzyzynski, T. (2016). An inverse kinematic algorithm for human leg. Journal of
Theoretical and Applied Mechanics, 54(1), 53–61.
Gouttefarde, M., & Gosselin, C. M. (2012). Analysis of the wrench-closureworkspace of planar
parallel cable-drivenmechanisms. IEEE Transaction on Robotics, 22(3), 434–445.
Hernandez-Santos, C., Soto, R., & Rodriguez, E. (2011). Design and dynamic modeling of
humanoid biped robot e-robot. IEEE Conference on Electronics, Robotics and Automotive
Mechanics CERMA, Cuernavaca, Mexico.
Marler, R. T., Arora, J. S., Yang, J., Kim, H. J., & Abdel-Malek, K. (2009). Use of multi-objective
optimization for digital human posture prediction. Engineering Optimization, 41(10), 925–943.
Mi, Z., Yang, J., & Abdel-Malek, K. (2009). Optimization-based posture prediction for human
upper body. Robotica, 27(4), 607–622.
Pennycott, A., Wyss, D., Vallery, H., Klamroth-Marganska, V., & Riener, R. (2012). Towards more
effective robotic gait training for stroke rehabilitation: A review. Journal of NeuroEngineering
and Rehabilitation, 9, 65.
Rastegarpanah, A., & Mozafar, S. (2016). Lower limb rehabilitation using patient data. Applied
Bionics and Biomechanics, 2016, 1–10.
Rezazadeh, S., & Behzadipour, S. (2011). Workspace analysis of multibody cable-driven mecha-
nisms. Journal of Mechanisms and Robotics. ASME, 3, 021005.
Spong, W. M., Hutchinson, S., & Vidyasagar, M. (2006). Robot modeling and control. New Delhi:
Wiley.
Winter, D. A. (2009). Biomechanics and motor control of human movement. New York: Wiley.
Yang, J., Marler, R. T., Kim, H. J., Arora, J. S., & Abdel-Malek, K. (2004). Multi-objective
optimization for upper body posture prediction. In Proceedings of the 10th AIAA/ISSMO
Multidisciplinary Analysis and Optimization Conference, Albany, NY (pp. 1–18). American
Institute of Aeronautics and Astronautics.
Chapter 5
Control of Robot Manipulators Using
Modified Backstepping Sliding Mode

Yassine Kali, Maarouf Saad, and Khalid Benjelloun

Abstract In this chapter, a robust backstepping sliding mode controller with

time delay estimation method is presented for high accuracy joint space tracking
trajectory of robot manipulators with unknown dynamics and external disturbances.
The proposed method is based on backstepping scheme, on a switching function
and on a time delay estimation in the final step. This structure is used to estimate
unknown dynamics and disturbances, to adapt the switching gains in term of the
error and its first time derivative and to reduce the control effort and the chattering
phenomenon. Lyapunov theory is used to establish sufficient condition that ensures
the stability of the closed-loop system. The proposed controller is simulated on a
two-link robot and implemented in real time on the 7-DOF ANAT robot to show the
effectiveness regarding unknown dynamics and external disturbances and chattering
reduction.

Keywords Backstepping · Sliding mode · Time delay estimation ·

Uncertainties · Robot manipulators · Tracking position

5.1 Introduction

During the last two decades, nonlinear control of uncertain robotic manipulators has
been the topic of many researches. In general, all developed controllers are based
on the mathematical model which is always an approximation of reality due to com-
putational complexity and external disturbances which make the implementation

Y. Kali () · K. Benjelloun

Department of Electrical Engineering, Ecole Mohammadia d’Ingénieurs, University of
Mohammed V, Rabat, Morocco
e-mail: [email protected]; [email protected]
M. Saad
Department of Electrical Engineering, École de Technologie Supérieure, Montreal, QC, Canada
e-mail: [email protected]

© Springer Nature Singapore Pte Ltd. 2019 107

N. Derbel et al. (eds.), New Developments and Advances
in Robot Control, Studies in Systems, Decision and Control 175,
https://doi.org/10.1007/978-981-13-2212-9_5
108 Y. Kali et al.

difficult (Lewis et al. 1993). In literature, many control algorithms have been
developed for uncertain robotic manipulators, including adaptive control (Cheah
et al. 2006; Seraji 1987), intelligent controllers such as fuzzy logic control (Guo
and Chung 2003; Yi and Woo 1997) and neural network (Hsia and Jung 1995),
Backstepping (Slotine and Li 1991), Sliding Mode Control (SMC) (Utkin 1992;
Utkin et al. 1999).
Sliding mode control where a switching control is designed to drive the systems
trajectories onto a specific hyperplane in the state space, also called sliding surface,
has been proposed for uncertain systems (Slotine and Sastry 1983; Utkin et al.
1999). Nonetheless, this control has many drawbacks, the major one is the well-
known chattering phenomenon which is due to the choice of switching gain larger
than the over-estimated bound of unknown dynamics and disturbances (Boiko and
Fridman 2005; Fridman 2001).
As a solution to this aforementioned phenomenon, some works have suggested
the use of continuous functions as saturation or hyperbolic tangent instead of the
sign one, but the steady state error does not converge in a finite-time (Boiko
2013; Utkin et al. 1999). Another approach is to use the observer-based sliding
mode (de Wit and Slotine 1991; Liu and Wang 2012; Slotine et al. 1986) where the
goal is to provide exact and robust estimation in order to allow chattering reduction
by a small choice of the switching gain matrix. However, the control performance
can be reduced if the estimation is not accurate. Moreover, sliding mode using
adaptive switching gains (Jamoussi et al. 2013) has been proposed, which provide an
adaptation of the switching gain to be as small as possible and sufficient to eliminate
the effect of unknown dynamics and external disturbances. However, the bounds of
the uncertainties must be known.
Otherwise, a Higher Order Sliding Mode (HOSM) controller (Fridman and
Levant 2002; Geng et al. 2013; Levant 1993; Ling et al. 2012; Utkin 2016) has been
proposed for chattering reduction, but the increasing of required informations and
the knowledge of the superior bound of the unknown dynamics and disturbances
is always required. This in turn makes the implementation difficult. Otherwise,
a combination of sliding mode with intelligent controller as neural-network and
fuzzy logic has been considered (Dotoli 2003; Jezernik et al. 1997; Lin 2003).
Even if these controllers provide a good approximation of unknown dynamics and
disturbances; however, their implementation is still difficult due to the large number
of parameters or fuzzy rules.
In addition, one can notice that Time Delay Estimation method (TDE) (Toumi
and Ito 1990; Toumi and Shortlidge 1991), can estimate nonlinear unknown
dynamics and unexpected disturbances effectively and simply. In this method, the
unknown parts of dynamics are assumed continuously differentiable with respect
to the time variable, and are estimated by using time delayed signals of the
system state derivatives and control. The TDE is proven to be efficient because the
estimation does not need an exact knowledge of the robot model. In Kali et al.
(2015), a combination of sliding mode control using saturation function instead the
signum one and time delay estimation method has been proposed. This controller is
chattering free but leads to a large steady state error.
5 Control of Robot Manipulators Using Modified Backstepping Sliding Mode 109

In this chapter, an accurate trajectory tracking control of robot manipulators

with unknown dynamics and unexpected disturbances is proposed for joint space
trajectory tracking by using a new method based on a combination of Backstepping
Sliding Mode and TDE. The choice of using backstepping sliding mode based on
exponential reaching law instead of conventional SMC is to deal with the chattering
problem since the exponential function allows adaptation to the variations of the
sliding surface (Fallaha et al. 2011; Rahman et al. 2013) and to eliminate the effect
of TDE error.
In the design procedure, the controller will be designed recursively by designing
intermediate control laws for some of the state variables considered as “virtual
controls” (Chang 2011; Slotine and Li 1991), and at the end of the procedure, the
switching manifold will be introduced. Finally, the term consisting on unknown
dynamics and external disturbances is approximated using time delay estimation
method.
The chapter is organized as follows. Section 5.1 introduces the chapter. In
Sect. 5.2, the dynamics of a general n-DOF robot manipulator and problem for-
mulation are presented. The new method is designed for high accuracy joint space
tracking and a Lyapunov based stability analysis for the closed loop dynamics with
sufficient condition are provided in Sect. 5.3. Simulation results are provided in
Sect. 5.4, to illustrate the advantages of the proposed nonlinear control technique.
In Sect. 5.5, experimental results of the proposed method applied on 7-DOF ANAT
robot arm installed at GREPCI Laboratory in real-time are given to prove the
effectiveness of the proposed controller. Finally, Sect. 5.6 concludes the chapter.

5.2 Mathematical Model of Robot Manipulators

and Preliminaries

5.2.1 Robot Dynamics

Consider the dynamics of n-DOF robot manipulator in the following matrix

equation (Spong et al. 2005):

M(q(t))q̈(t) + C(q(t), q̇(t))q̇(t) + G(q(t)) = τ (t) + τd (t) (5.1)

where:
• q(t), q̇(t), q̈(t) ∈ R n represent the joint space position, velocity and acceleration
vectors, respectively.
• M(q(t)) ∈ R n×n denotes the inertia matrix which is symmetric definite positive
and its inverse always exists.
• C(q(t), q̇(t)) ∈ R n×n is the centrifugal and Coriolis matrix.
• G(q(t)) ∈ R n represents the gravitational vector.
• τ (t) ∈ R n is the torque input vector.
• τd (t) ∈ R n denotes the disturbance vector.
110 Y. Kali et al.

Introducing a diagonal matrix where the elements are strictly positive constants:
⎡ ⎤
m11 0 · · · 0
⎢ ⎥
⎢ 0 m22 . . . : ⎥
⎢
M=⎢ ⎥ (5.2)
. . ⎥
⎣ : . . . 0 ⎦
.
0 · · · 0 mnn

The dynamic equation of the robot manipulator given in partitioned form in Eq. (5.1)
can be rewritten as follows:

M q̈(t) + N(q(t), q̇(t)) + H (q(t), q̇(t), q̈(t)) = τ (t) (5.3)

where:

N(q(t), q̇(t)) = C0 (q(t), q̇(t))q̇(t) + G0 (q(t))

H (q(t), q̇(t), q̈(t)) = (M(q(t)) − M)q̈(t) + Cu (q(t), q̇(t))q̇(t) + Gu (q(t)) − τd (t)

with C0 (q(t), q̇(t)) and G0 (q(t)) are the nominal parts while Cu (q(t), q̇(t)) and
Gu (q(t)) denote the uncertain parts caused by load variations, unknown parame-
ters, unmodel dynamics and external disturbances of C(q(t), q̇(t)) and G(q(t)),
respectively. For simplicity, let us denote M(t) = M(q(t)), N (t) = N (q(t), q̇(t))
and H (t) = H (q(t), q̇(t), q̈(t)).

5.2.2 Problem Formulation

The objective is to design a robust joint space controller that ensures high accuracy
tracking trajectory even in presence of unknown dynamics and unexpected distur-
bances. To that end, the controller will be designed and its stability analysis will be
carried out based on the following properties and assumptions:
Property 1 The inertia matrix M(q(t)) is symmetric definite positive and bounded
such as:

0 < m1 ≤ M(q(t)) ≤ m2

with m1 and m2 are two known positive constants (Spong et al. 2005).
Property 2 The diagonal positive constant matrix M is chosen such as the follow-
ing condition in Spong et al. (2005) is verified:

In×n − M −1 (q(t))M < 1

5 Control of Robot Manipulators Using Modified Backstepping Sliding Mode 111

Assumption 1 The joint position and velocity states are measurable.

Assumption 2 The functions Hi (t) for i = 1, · · · , n of the uncertain vector
H (t) = [H1 (t), · · · , Hn (t)]T are continuously differentiable with respect to the
time variable and don’t vary largely during a small L period of time (Toumi and Ito
1990).

5.3 Controller Design

In this section, a combination of backstepping sliding mode with exponential

reaching law and time delay estimation method will be designed to ensure the
convergence of the joint space trajectories to the known desired ones with high
accuracy even in presence of uncertainties. The architecture of the closed-loop
system is represented in Fig. 5.1.
The design of the proposed controller is described step-by-step as follows:
Step 1 This first step consists on defining the regulated variable to be the joint space
tracking error as follows:

e(t) = q(t) − qd (t) (5.4)

where qd (t) ∈ R n is the desired joint space trajectory vector. Therefore, its first time
derivative is computed as:

ė(t) = q̇(t) − q̇d (t) (5.5)

Fig. 5.1 Block diagram of the proposed controller

112 Y. Kali et al.

Our objective in this step is to force the convergence of the regulated variable to zero
e(t) → 0 by designing a virtual control. To reach this goal, the following Lyapunov
function is selected:
1 T
V1 (t) = e (t)e(t) (5.6)
2
Its first time derivative is obtained as follows:

V̇1 (t) = eT (t)ė(t)

= eT (t)(q̇(t) − q̇d (t)) (5.7)

Choosing q̇(t) as virtual control variable. Then, an appropriate stabilizing function

is selected to ensure stability as follows:

q̇(t) = β1 (t)
= q̇d (t) − K1 e(t) (5.8)

where K1 = diag (k11 , k12 , · · · , k1n ) is a diagonal positive matrix.

Substituting the stabilizing function β1 (t) in the first time derivative of the
Lyapunov function (5.7) leads to:

V̇1 (t) = −eT (t)K1 e(t) (5.9)

From the above equation, it is clear that V̇1 (t) is negative definite, which proves that
the convergence of the tracking error e(t) to zero is ensured.
Step 2 Defining now the switching function to be the difference between the virtual
control and the stabilizing function as:

σ (t) = q̇(t) − β1 (t)

= q̇(t) − q̇d (t) + K1 e(t) (5.10)

Differentiating the above sliding surface (5.10) with respect to time and using the
robot model in (5.3) gives:

σ̇ (t) = q̈(t) − β̇1 (t)

= q̈(t) − q̈d (t) + K1 ė(t)
−1 !
=M τ (t) − N(t) − H (t) − q̈d (t) + K1 ė(t) (5.11)

Therefore, based on Eq. (5.10), the first time derivative of the error ė(t) in (5.5) can
be rewritten as follows:

ė(t) = σ (t) − K1 e(t) (5.12)

5 Control of Robot Manipulators Using Modified Backstepping Sliding Mode 113

Thus, the second Lyapunov function is selected as:

1
V2 (t) = V1 (t) + σ T (t)σ (t) (5.13)
2
Its first time derivative is calculated as:

V̇2 (t) = V̇1 (t) + σ T (t)σ̇ (t)

= eT (t)ė(t) + σ T (t)σ̇ (t)
 
= eT (t) σ (t) − K1 e(t)
 −1
+σ T (t) M [τ (t) − N(t) − H (t)] − q̈d (t) + K1 ė(t)
 −1
= σ T (t) M [τ (t) − N(t) − H (t)] − q̈d (t) + K1 ė(t) + e(t)

−eT (t)K1 e(t) (5.14)

By choosing:

−1
M [τ (t) − N(t) − H (t)] − q̈d (t) + K1 ė(t) + e(t) =
−K2 σ (t) − K(σ (t)) sign (σ (t)) (5.15)

where K2 = ( (k21 , k22 , · · · , k2n ) is a diagonal )positive matrix and

diag
k1 k2 kn
K(σ (t)) = diag , ,··· , with ki > 0 and
D1 (σ1 (t)) D2 (σ2 (t)) Dn (σn (t))
the term Di (σi (t)) for i = 1, · · · , n is defined in Fallaha et al. (2011) as follows:

Di (σi ) = δi + (1 − δi )e−γi |σi (t)|

pi
(5.16)

where 0 < δi < 1, γi > 0, pi > 0. Finally, the signum function is defined by
!T
sign (σ (t)) = sign (σ1 (t)), · · · , sign (σn (t)) with:
⎧
⎨ 1, if σi (t) > 0
sign (σi (t)) = 0, if σi (t) = 0 (5.17)
⎩
−1, if σi (t) < 0

Remark 1 Notice that the exponential reaching law allows chattering reduction due
ki
to the gradually variations of the term in Eq. (5.16) that increases and
Di (σi (t))
ki
decreases between ki and depending on the sliding surface variations. In other
δi
ki
words, increases during the reaching phase which make the attraction of
Di (σi (t))
114 Y. Kali et al.

σi (t) faster (i.e, when the value of |σi (t)| becomes high, then, Di (σi ) converges to
ki ki
δi and to ) while decreases during the sliding phase when the error and its
Di (σi ) δi
first time derivative converge to zero (i.e, when σi (t) converges to 0, then, the term
ki
Di (σi ) converges to 1 and to ki ).
Di (σi )
Then, the Backstepping Sliding Mode Control (BSMC) is obtained by resolving
Eq. (5.15) as:

τ (t) = Mu(t) + N(t) + H (t) (5.18)

where:

u(t) = q̈d (t) − K1 ė(t) − e(t) − K2 σ (t) − K(σ (t)) sign (σ (t)) (5.19)

Since H (t) is uncertain, the control performance will be affected. Then, based on
Assumption 2 given in Sect. 5.2, H (t) can be estimated using time delay estimation
method (Toumi and Ito 1990) as:

Ĥ (t) ∼
= H (t − L)
= τ (t − L) − N(t − L) − M q̈(t − L) (5.20)

where L is the estimation time delay. Clearly the accuracy of Ĥ (t) improves as L
decreases. In practice, L can’t be chosen smaller than the sampling period.
Finally, the proposed robust backstepping sliding mode with time delay estima-
tion is obtained as:

τ (t) = Mu(t) + N(t) + Ĥ (t)

!
= τ (t − L) + M u(t) − q̈(t − L) + N (t) − N (t − L) (5.21)

where u(t) is given in Eq. (5.19) and the time delayed acceleration q̈(t − L) can be
obtained by one of the following approximations:

1 !
q̈(t − L) = q̇(t − L) − q̇(t − 2L) (5.22)
L
1 !
q̈(t − L) = 2
q(t − L) − 2q(t − 2L) + q(t − 3L) (5.23)
L
Step 3 For the stability analysis, the proposed controller in Eq. (5.21) is replaced in
the first time derivative of the selected second Lyapunov function in Eq. (5.14) as:
 −1
V̇2 (t) = σ T (t) M [Ĥ (t) − H (t)] − K2 σ (t) − K(σ ) sign (σ (t))

−eT (t)K1 e(t)

5 Control of Robot Manipulators Using Modified Backstepping Sliding Mode 115

 
= −eT (t)K1 e(t) − σ T (t)K2 σ (t) + σ T ε(t) − K(σ ) sign (σ (t))
n  
  ki
= − k1i ei2 (t) + k2i σi2 (t) + σi (t) εi (t) − sign (σi )
Di (σi )
i=1
n  
  ki
= − k1i ei (t) + k2i σi (t) + σi (t)εi (t) −
2 2
|σi (t)|
Di (σi )
i=1
n  
  ki
≤ − k1i ei2 (t) + k2i σi2 (t) − |σi (t)| − |εi (t)| (5.24)
Di (σi )
i=1

−1
where ε(t) = M [Ĥ (t) − H (t)] denotes the time delay estimation error. To ensure
that V̇2 (t) is a negative definite function, the following condition must be verified:

ki > |εi (t)|, f or i = 1, · · · , n. (5.25)

Therefore, the error e(t) and its first time derivative ė(t) are asymptotically bounded
(i.e. the closed loop system is stable) if the TDE error ε(t) is asymptotically
bounded. Therefore, the asymptotic boundedness of ε will be proved as in Jin et al.
(2008).
Substituting the controller given in Eq. (5.21) in the mathematical model of the
robot (5.3) leads to:

ε(t) = q̈(t) − u(t) (5.26)

Multipliying both sides of the above equation by M(t) and using Eq. (5.1) gives:

M(t)ε(t) = M(t)(q̈(t) − u(t))

= τ (t) + τd (t) − C(q(t), q̇(t))q̇(t) − G(q(t)) − M(t)u(t) (5.27)

Then, substituting (5.21) in the above equation leads to:

M(t)ε(t) = Mu(t) + N(t) + Ĥ (t) + τd (t) − C(q(t), q̇(t))q̇(t)

−G(q(t)) − M(t)u(t)
= [M − M(t)]u(t) + Ĥ (t) + τd (t) − Cu (q(t), q̇(t))q̇(t)
−Gu (q(t)) (5.28)
!
As H (t) = M(t) − M q̈(t) + Cu (q(t), q̇(t))q̇(t) + Gu (q(t)) − τd (t). Then, Ĥ (t)
can be expressed as:
!
Ĥ (t) = M(t − L) − M q̈(t − L) + Gu (q(t − L)) − τd (t − L)
+Cu (q(t − L), q̇(t − L))q̇(t − L) (5.29)
116 Y. Kali et al.

Therefore, Eq. (5.28) becomes:

!
M(t)ε(t) = [M − M(t)]u(t) + M(t − L) − M q̈(t − L) + Δ (5.30)

where:

Δ = Cu (q(t − L), q̇(t − L))q̇(t − L) − Cu (q(t), q̇(t))q̇(t)

+Gu (q(t − L)) − Gu (q(t)) + τd (t) − τd (t − L) (5.31)

Obviously, for a sufficiently small L, Δ is bounded. Then, using Eq. (5.26), the
following expression can be stated:

q̈(t − L) = ε(t − L) + u(t − L) (5.32)

Then:
! !
M(t)ε(t) = M − M(t) u(t) + M(t − L) − M(t) q̈(t − L) + Δ
!
+ M(t) − M q̈(t − L)
! !
= M − M(t) u(t) + M(t − L) − M(t) q̈(t − L) + Δ
! !
+ M(t) − M ε(t − L) + u(t − L)
! ! !
= M(t) − M ε(t − L) + M(t) − M u(t − L) − u(t)
!
+ M(t − L) − M(t) q̈(t − L) + Δ (5.33)

Therefore, the TDE error is expressed as:

ε(t) = Dε(t − L) + Dζ1 + ζ2 (5.34)

where

D = I − M −1 (t)M

ζ1 = u(t − L) − u(t)

ζ2 = M −1 (t)[M(t − L) − M(t)]q̈(t − L) + M −1 (t)Δ

Using Assumptions 1 and 2, it is clear that ζ1 and ζ2 are bounded for a small L. In
addition, by choosing an appropriate M, the following inequation is verified (Spong
et al. 2005):

D = I − M −1 M < 1 (5.35)

5 Control of Robot Manipulators Using Modified Backstepping Sliding Mode 117

In the discrete-time domain, Eq. (5.34) is represented as:

ε(k) = D(k)ε(k − 1) + Dζ1 (k) + ζ2 (k) (5.36)

Since D < 1, the roots of D(k) reside inside a unit circle, thus, the TDE error in
Eq. (5.36) is asymptotically bounded with bounded functions ζ1 and ζ2 .
Hence, the stability of the closed-loop system is proved.
Remark 2 In real-time, the measured signals are contaminated by noise. The
noise effect might be amplified when q̈(t − L) is obtained using one of the
approximations (5.22) and (5.23). As a solution to solve this problem, a Low Pass
Filter (LPF) may be introduced before calculating the time delayed acceleration.
However, the attenuation of noise without using a LPF is possible by choosing small
values for M.
If a digital LPF with the cutoff frequency β is adopted, the control law can be
modified as follows:

τ f (t) = βL(1 + βL)−1 τ (t) + (1 + βL)−1 τ f (t − L) (5.37)

where τ (t) denotes the calculated input before the filter while τ f (t) denotes the
filtered control input. Substituting τ (t) by its expression in (5.21), gives:
 
τ f (t) = τ f (t − L) + βL(1 + βL)−1 M u(t) − q̈(t − L)
 
+βL(1 + βL)−1 N(t) − N(t − L) (5.38)

Comparing (5.38) with the controller in (5.21), then:

M = βL(1 + βL)−1 M (5.39)

Since βL(1 + βL)−1 < 1, then, for very small value of M, a similar result as using
a digital LPF will be obtained.

5.4 Simulation Studies

In this section, three numerical simulations are presented for trajectory tracking
of a two-link rigid robotic manipulator shown in Fig. 5.2 using Matlab/Simulink
software with ODE 4 Solver and a sampled time equal to 0.001 s.
For this two-link manipulator, the dynamic equation (5.1) has the following
matrices,
( )
l22 m2 + 2l1 l2 m2 c2 + l12 (m1 + m2 ) + J1 l2 m2 (l1 + l2 )
M(q(t)) = ,
l2 m2 (l1 + l2 ) l22 m2 + J2
118 Y. Kali et al.

Fig. 5.2 Considered two-link

robot manipulator

( )
−l1 l2 m2 s2 q̇22 (t) − 2l1 l2 m2 s2 q̇1 (t)q̇2 (t)
C(q(t), q̇(t))q̇(t) = ,
l1 l2 m2 s2 q̇22 (t)

( )
l2 m2 gc12 + (m1 + m2 )l1 gc1
G(q(t)) = ,
l2 m2 gc12

⎧( )
⎪
⎪ 2 sin t + 0.5 sin(200π t) , if t < 6 s
⎪
⎪
⎪
⎨ cos 2t + 0.5 sin(200π t)
τd (t) =
⎪
⎪ ( )
⎪
⎪ 20 sin t + 5 sin(200π t)
⎪
⎩ 10 cos 2t + 5 sin(200π t) , else

where si , ci and cij are defined by si = sin(qi (t)), ci = cos(qi (t)) and cij =
cos(qi (t) + qj (t)), respectively. The parameters of the above elements are reported
in Table 5.1.

Table 5.1 Parameters of the two-link robot manipulator

Symbol Definition Value S.I. Unit
l1 Length of the first link 1 (m)
l2 Length of the second link 0.85 (m)
J1 Moment of inertia of the first motor 5 (kg m2 )
J2 Moment of inertia of the second motor 5 (kg m2 )
m1 Mass of link 1 0.5 (kg)
m2 Mass of link 2 1.5 (kg)
m̂1 Nominal mass of link 1 0.4 (kg)
m̂2 Nominal mass of link 2 1.2 (kg)
g Gravitational constant 9.81 (m/s2 )
5 Control of Robot Manipulators Using Modified Backstepping Sliding Mode 119

The control objective here is to follow the reference signals defined by:

q1d (t) = 1.25 − (7/5)e−t + (7/20)e−4t (5.40)

−t −4t
q2d (t) = 1.25 + e − (1/4)e (5.41)

The initial states (joint positions velocities) are selected to be q1 (0) =0.4 rad,
q2 (0) = 1.8 rad and q̇1 (0) = q̇2 (0) = 0 rad/s.

5.4.1 Classical SMC Simulated on Robotic Arm

A complete study of classical sliding mode theory can be found in Utkin (1992).
The control law for the considered robot is given as:
!
τ (t) = M0 (q(t)) q̈d (t) − λė(t) − sign (σ (t)) + C0 (q(t), q̇(t))q̇(t) + G0 (q(t))

where M0 (q(t)), C0 (q(t), q̇(t)) and G0 (q(t)) are the nominal matrices calculated
using the nominal masses m̂1 and m̂2 given in Table 5.1. The controller gains used
in the simulation are:

λ = diag (2, 2), K = diag (5, 5)

The simulation results are given in Fig. 5.3.

5.4.2 SMC with TDE Simulated on Robotic Arm

The theoretical development of the sliding mode control with time delay estimation
can be found in Kali et al. (2015, 2017a). The control law for robot is given as:
!
τ (t) = τ (t −L) − M0 (q(t −L))q̈(t −L) + M0 (q(t)) q̈d (t)−λė(t)−Ksat (σ (t))
+C0 (q(t), q̇(t))q̇(t)+G0 (q(t))−C0 (q(t −L), q̇(t −L))q̇(t −L)−G0 (q(t −L))

In this simulation, the used parameters for sliding mode control using saturation
function with time delay estimation are:

λ = diag (2, 2), K = diag (5, 5), L = Ts = 0.001 s

Moreover, the time delayed acceleration q̈(t − L) is obtained using the second
approximation (5.23). The obtained simulation results are shown in Fig. 5.4.
120 Y. Kali et al.

Fig. 5.3 Obtained simulation results via classical sliding mode control. (a) Joint 1 tracking. (b)
Joint 2 tracking. (c) Joint 1 error. (d) Joint 2 error. (e) Torque of joint 1. (f) Torque of joint 2

5.4.3 Proposed Controller Simulated on Robotic Arm

The proposed modified backstepping sliding mode simulated on the two-link robot
is given in Eq. (5.21) with:

K1 = diag (2, 2), K2 = diag (5, 5), k1 = k2 = 3

δ1 = δ2 = 0.6, γ1 = γ2 = 20, p1 = p2 = 1

and L = Ts = 0,001 s while the time delayed acceleration q̈(t − L) is obtained

using the second approximation (5.23). The obtained simulation results are depicted
in Fig. 5.5.

5.4.4 Discussion
5.4.4.1 Classical Sliding Mode Control

The joint positions q1 (t) and q2 (t) shown respectively in Fig. 5.3a, b did not follow
their desired references properly (the tracking error doesn’t converge totally to zero).
5 Control of Robot Manipulators Using Modified Backstepping Sliding Mode 121

Fig. 5.4 Obtained simulation results via sliding mode control with time delay estimation. (a) Joint
1 tracking. (b) Joint 2 tracking. (c) Joint 1 error. (d) Joint 2 error. (e) Torque of joint 1. (f) Torque
of joint 2

In addition, Fig. 5.3e, f show that the chattering phenomenon and the control effort
are very important due to the large choice of the switching gains.

5.4.4.2 Sliding Mode Controller with Time Delay Estimation Method

Figure 5.4e, f show that the control effort and the chattering phenomenon in control
torque inputs is reduced except at the beginning. However, the robot system deals
with a large steady state error as depicted in Fig. 5.4c, d.

5.4.4.3 Proposed Backstepping Sliding Mode Controller with Time Delay

Estimation Method

The obtained simulation results show the effectiveness of the modified backstepping
sliding mode method, since the joints q1 and q2 shown respectively in Fig. 5.5a, b
convergence to their desired trajectories in finite time. This is confirmed in the
tracking error figures (Fig. 5.5c, d). In addition, the proposed controller allows
chattering reduction as shown in Fig. 5.5e, f.
122 Y. Kali et al.

Fig. 5.5 Obtained simulation results via the proposed controller. (a) Joint 1 tracking. (b) Joint 2
tracking. (c) Joint 1 error. (d) Joint 2 error. (e) Torque of joint 1. (f) Torque of joint 2

From the above obtained results, the proposed controller shows the superiority
such that the steady state error converges to zero faster due to the exponential
reaching law. The chattering is still present SMC with TDE method, which can
damage the controlled robot, while is reduced significantly in the proposed method.

5.5 Case Study

5.5.1 System Description

In this section, the real time implementation of the proposed optimal super-
twisting with time delay estimation on the 7-DOF ANAT robot shown in Fig. 5.6
using Simulink with Real-Time Workshop is presented. ANAT (Articulated Nimble
Adaptable Trunk) Robot is built by Robotics Design inc. (Canada). The first joint of
the used ANAT is prismatic, followed by six rotary joints (Kali et al. 2017b).
As shown in Fig. 5.7, the frames of references are selected using the modified
Denavit-Hartenberg (Craig 2005) (D-H) convention. The D-H parameters that
determine the homogeneous transformation matrices are given in Table 5.2.
5 Control of Robot Manipulators Using Modified Backstepping Sliding Mode 123

Fig. 5.6 7-DOF ANAT robot arm

Fig. 5.7 D-H frames of the

ANAT robot arm

Table 5.2 D-H parameters Joints αi−1 (rad) ai−1 (m) di (m) qi (rad)
of the 7-DOF ANAT robot
1 0 0 q1 0
2 0 L1 0 q2
3 0 L0 0 q3
4 0 L0 0 q4
5 0 L0 L2 q5
6 π/2 L3 0 q6
7 −π/2 0 −L4 q7
124 Y. Kali et al.

The initial task space position of the end-effector that is placed on the last joint
(7th joint) is: x(0) = [0.6764 0 − 0.19]T while the initial joint positions are 0 rad
and joint velocities are 0 rad/s. The equation of motion are obtained such in Eq. (5.3)
with:

M = diag (0.23, 0.11, 0.14, 0.12, 0.12, 0.12, 0.12).

In addition, the ANAT model verifies the properties and the assumptions given in
Sect. 5.2.

5.5.2 Real-Time Setup

All robot joints are actuated by DC motors. The actuators encoders provide position
measurements. The ATMEGA 16 microcontrollers are used to send the angles
positions to simulink, and using the forward kinematics, the joint space trajectory
is transformed to the workspace trajectory. The microcontrollers are also used to
transform the control input signals into a Pulse Width Modulation (PWM) signal.
Then, this PWM signal is applied to the H-bridge drive of the actuators of the 7-
DOF ANAT robot. A current sensor located in the H-bridge drive provides the
current measurement of each actuator. In the experimentation, the sixth, and seventh
joints are chosen to stay at their initial position. A sampling time Ts = 0.03 s has
been selected for the control loop (Kali et al. 2017b). The above real-time setup is
illustrated in Fig. 5.8.

Fig. 5.8 Experimental setup

5 Control of Robot Manipulators Using Modified Backstepping Sliding Mode 125

5.5.3 Controller Setting

The proposed robust backstepping sliding mode control with time delay estimation
has four diagonal matrices with constant elements. This makes the controller setting
simplify, which is briefly described as follows:
• Determine the matrix K1 of the sliding surface in (5.11).
• Determine the matrix K2 such as a fast convergence will be obtained.
• Select L as small as possible (equal to the sampling time interval Ts ).
• Tuning the matrix M based on the Property 2 given in Sect. 5.2. This step is very
important since M plays a key role in the stability and the noise attenuation. To
that end, the diagonal elements of M are increased gradually from small positive
values, while checking the control performance by trial and error.
• Finally the constants ki for i = 1, · · · , n are chosen such as the condition of
stability in (5.25) is verified, they should be further tuned to achieve the optimal
performance.

5.5.4 Experimental Results

The control objective here is to follow a known spiral desired trajectory defined in
the task space. For this application the desired spiral is chosen such as all kinematic
singularities are avoided, the relation between the desired accelerations in the task
space and the joint space is given by:

q̈d (t) = J + ẍd (t) − J + J˙J + ẋd (t) (5.42)

where J + = J T (J J T )−1 denotes the generalized inverse of the Jacobian matrix

while J˙ is its derivative, q̈d (t) is the desired joint acceleration vector, ẍd (t) and ẋd (t)
are the task space desired acceleration and velocity vectors, respectively. The desired
velocity in the joint space q̇d (t) is obtained from q̈d (t) using an integrator while
the desired position in the joint space qd (t) is obtained from q̇d (t) using another
integrator.
In Sect. 5.3, the proposed controller that ensures the stability of the closed loop
system is given in (5.21). The time delay estimation is chosen to be the sampled
time: L = Ts = 0.03 s and the time delayed acceleration q̈i (t − L) for i = 1, · · · , 7
is obtained using the approximation based on the joint position as:

1 !
q̈i (t − L) = 2
qi (t − L) − 2qi (t − 2L) + qi (t − 3L)
Ts
126 Y. Kali et al.

Fig. 5.9 3D Task space tracking trajectory

The controller gains are chosen such as the condition of stability in Eq. (5.25) is
verified:

K1 = diag (2, 2, 2, 2, 2, 2, 2)

K2 = diag (1.2, 1.2, 1.2, 1.2, 1.2, 1.2, 1.2)

ki = 3.8; δi = 0.7; γi = 20; pi = 1 f or i = 1 : 7

The experimental results are shown in Figs. 5.9, 5.10, 5.11, 5.12, 5.13 and 5.14:
3D task space tracking, workspace tracking position and error, joint space tracking
position and error and control torque inputs. In these results, the performances
produced by the proposed controller are illustrated for the ANAT robot by neglecting
the Coriolis matrix and adding a load of 5.5 kg.
The proposed controller ensures the convergence of the end-effector position
to the task space desired position with high accuracy due to a good estimation of
uncertainties as shown in Figs. 5.9, 5.10 and 5.11.
Figure 5.12 illustrates the high accuracy tracking position in joint space even in
presence of uncertainties and external disturbances confirmed by the small tracking
errors as depicted in Fig. 5.13.
Furthermore, it can be seen from Fig. 5.14 that the control torque inputs are
smooth and chattering free with small values that are acceptable for our robot, this
is thanks to the exponential reaching law that make the controller adapted to the
variations of the sliding surface and forcing ki /Ni (Si ) to vary between ki and ki /δi
for i = 1, · · · , n (here n = 7).
5 Control of Robot Manipulators Using Modified Backstepping Sliding Mode 127

Fig. 5.10 Task space tracking position. (a) X-tracking. (b) Y-tracking. (c) Z-tracking

To evaluate the performance of the proposed modified backstepping sliding

mode, it is compared with the conventional SMC and the SMTDC (Kali et al. 2015)
for the spiral desired trajectory in presence of uncertainties using two performance
indices.The first one, the Root-Mean-Squared (RMS) error ei RMS is a numerical
measure of average tracking performance while and the second one is the RMS
torque τi RMS which is the average control torque input allowing the evaluation of
the control effort. The expressions of ei RMS and τi RMS are given as follows:
128 Y. Kali et al.

Fig. 5.11 Task space tracking error. (a) X-error. (b) Y-error. (c) Z-error
5 Control of Robot Manipulators Using Modified Backstepping Sliding Mode 129

Fig. 5.12 Joint space tracking position. (a) Tracking position of joint 1. (b) Tracking position of
joint 2. (c) Tracking position of joint 3. (d) Tracking position of joint 4. (e) Tracking position of
joint 5
130 Y. Kali et al.

Fig. 5.12 (continued)

%
&
&1 N
eRMS =' e(k)2 (5.43)
N
k=1

%
&
&1 N
τ RMS =' τ (k)2 (5.44)
N
k=1

where the first one is the Root-Mean-Squared (RMS) error eRMS which is a
numerical measure of the average tracking performance and the second one is
the RMS torque τ RMS which is the average control torque input allowing the
5 Control of Robot Manipulators Using Modified Backstepping Sliding Mode 131

Fig. 5.13 Joint space tracking error. (a) Tracking error of joint 1. (b) Tracking error of joint 2. (c)
Tracking error of joint 3. (d) Tracking error of joint 4. (e) Tracking error of joint 5
132 Y. Kali et al.

Fig. 5.13 (continued)

evaluation of the control effort and N is equal to the number of sampling steps of
the experimentation. The quantitative analysis is reported in Tables 5.3 and 5.4.
From the above comparison, one can notice that SMC can’t stabilize the system
in presence of uncertainties and external disturbances due to the values of the
switching gain matrix that are large, which make an excessive control effort not
accepted by our robot. Therefore, SMTDC leads with a large steady state error while
the proposed method can guarantee better results regarding the energy reduction and
high accuracy tracking even in presence of uncertainties.
5 Control of Robot Manipulators Using Modified Backstepping Sliding Mode 133

Fig. 5.14 Control torque inputs. (a) Torque input of joint 1. (b) Torque input of joint 2. (c) Torque
input of joint 3. (d) Torque input of joint 4. (e) Torque input of joint 5
134 Y. Kali et al.

Fig. 5.14 (continued)

Table 5.3 Comparison of RMS errors (rad)

Controller Joint 1 Joint 2 Joint 3 Joint 4 Joint 5
SMC 0.11 0.07 0.105 0.087 0.14
SMTDC 0.05 0.85 0.071 0.091 0.096
Proposed controller 0.008 0.019 0.014 0.016 0.021

Table 5.4 Comparison of RMS torques input (N m)

Controller Joint 1 Joint 2 Joint 3 Joint 4 Joint 5
SMC 1.067 1.293 1.53 1.47 1.628
SMTDC 0.728 0.637 0.94 0.824 0.641
Proposed controller 0.233 0.392 0.381 0.271 0.48
5 Control of Robot Manipulators Using Modified Backstepping Sliding Mode 135

5.6 Conclusion

In this work, a high accuracy tracking control based on backstepping sliding mode
with time delay estimation for robot manipulators systems with unknown dynamics
and external disturbances has been presented. Sufficient condition is established
to guarantee the stability by defining an appropriate Lyapunov function. The
experimental results show the merit and the effectiveness of this new control. The
results were satisfactory. The tracking trajectory was done with high accuracy and
convenable control torque input even in the presence of uncertainties and external
disturbance.

Acknowledgements The authors are grateful to Nabil Derbel (University of Sfax, Tunisia),
Jawhar Ghommam (University of Tunis, Tunisia) and Quanmin Zhu (University of the West of
England) for the opportunity to contribute to the New developments and advances in the field of
Robotics.

References

Boiko, I. (2013). Chattering in sliding mode control systems with boundary layer approximation
of discontinuous control. International Journal of Systems Science, 44, 1126–1133.
Boiko, I., & Fridman, L. (2005). Analysis of chattering in continuous sliding-mode controllers.
IEEE Transactions on Automatic Control, 50, 1442–1446.
Chang, Y. (2011). Block backstepping control of MIMO systems. IEEE Transactions on Automa-
tion Control, 56(5), 1191–1197.
Cheah, C., Liu, C., & Slotine, J. (2006). Adaptive Jacobian tracking control of robots with
uncertainties in kinematic, dynamic and actuator models. IEEE Transactions on Automatic
Control, 51, 1024–1029.
Craig, J. (2005). Introduction to robotics: Mechanics and control (3rd ed.). Upper Saddle River,
NJ: Pearson/Prentice-Hall.
de Wit, C. C., & Slotine, J.-J. (1991). Sliding observers for robot manipulators. Automatica, 27(5),
859–864.
Dotoli, M. (2003). Fuzzy sliding mode control with piecewise linear switching manifold. Asian
Journal of Control, 5(4), 528–542.
Fallaha, C., Saad, M., Kanaan, H., & Al-Haddad, K. (2011). Sliding mode robot control with
exponential reaching law. IEEE Transactions on Industrial Electronics, 58, 600–610.
Fridman, L. (2001). An averaging approach to chattering. IEEE Transactions on Automatic
Control, 46, 1260–1265.
Fridman, L., & Levant, A. (2002). Higher order sliding mode. In J. P. Barbot & W. Perruguetti
(Eds.), Sliding mode control in engineering (Systems and control book series). Marcel Dekker.
Geng, J., Sheng, Y., Liu, X., & Liu, B. (2013). Second order time varying sliding mode control
for uncertain systems. In 25th Chinese Control and Decision Conference (CCDC), Guiyang
(pp. 3880–3885).
Guo, Y., & Chung, M.-J. (2003). An adaptive fuzzy sliding mode controller for robotic manipula-
tors. IEEE Transactions on Systems, Man, and Cybernetics, 33(2), 149–159.
Hsia, T., & Jung, S. (1995). A simple alternative to neural network control scheme for robot
manipulators. IEEE Transactions on Industrial Electronics, 42(4), 414–416.
136 Y. Kali et al.

Jamoussi, K., Ouali, M., Chrifi-Alaoui, L., & Benderradji, H. (2013). Robust sliding mode control
using adaptive switching gain for induction motors. International Journal of Automation and
Computing, 10(4), 303–311.
Jezernik, K., Rodic, M., Safaric, R., & Curk, B. (1997). Neural network sliding mode robot control.
Robotica, 15, 23–30.
Jin, M., Kang, S. H., & Chang, P. H. (2008). Robust compliant motion control of robot with
nonlinear friction using time-delay estimation. IEEE Transactions on Industrial Electronics,
55(1), 258–269.
Kali, Y., Saad, M., Benjelloun, K., & Benbrahim, M. (2015). Sliding mode with time delay control
for mimo nonlinear systems with unknown dynamics. International Workshop on Recent
Advances in Sliding Modes, Istanbul, Turkey, April 9–11.
Kali, Y., Saad, M., Benjelloun, K., & Benbrahim, M. (2017a). Sliding mode with time delay control
for robot manipulators (pp. 135–156). Singapore: Springer.
Kali, Y., Saad, M., Benjelloun, K., & Fatemi, A. (2017b). Discrete-time second order sliding mode
with time delay control for uncertain robot manipulators. Robotics and Autonomous Systems,
94, 53–60.
Levant, A. (1993). Sliding order and sliding accuracy in sliding mode control. International
Journal of Control, 58, 1247–1263.
Lewis, F., Abdallah, C., & Dawson, D. (1993). Control of robot manipulators. New York:
Macmillan Publishing Company.
Lin, C. (2003). Nonsingular terminal sliding mode control of robot manipulators using fuzzy
wavelet networks. IEEE Transactions on Fuzzy Systems, 14(6), 849–859.
Ling, R., Wu, M., Dong, Y., & Chai, Y. (2012). High order sliding mode control for uncertain
nonlinear systems with relative degree three. Nonlinear Science and Numerical Simulation, 17,
3406–3416.
Liu, J., & Wang, X. (2012). Advanced sliding mode control for mechanical systems.
Berlin/Heidelberg: Springer.
Rahman, M., Saad, M., Kenné, J., & Archambault, P. (2013). Control of an exoskeleton robot arm
with sliding mode exponential reaching law. International Journal of Control, Automation and
Systems, 11(1), 92–104.
Seraji, H. (1987). Adaptive control of robotic manipulators. In 26th IEEE Conference on Decision
and Control, Los Angeles (p. 26).
Slotine, J., & Li, W. (1991). Applied nonlinear control. Taipei: Printice-Hall International.
Slotine, J. J., & Sastry, S. S. (1983). Tracking control of non-linear systems using sliding surfaces,
with application to robot manipulators. International Journal of Control, 38(2), 465–492.
Slotine, J. J., Hedrick, J. K., & Misawa, E. A. (1986). On sliding observers for nonlinear systems.
In American Control Conference, Seattle (pp. 1794–1800).
Spong, M., Hutchinson, S., & Vidyasagar, M. (2005). Robot dynamics and control. New York:
Wiley.
Toumi, K., & Ito, O. (1990). A time delay controller for systems with unknown dynamics. ASME
Journal of Dynamic System, Measurement and Control, 112, 133–141.
Toumi, K., & Shortlidge, C. (1991). Control of robot manipulators using time delay. In IEEE
International Conference of Robotics and Automation.
Utkin, V. (1992). Sliding mode in control and optimization. Berlin: Springer.
Utkin, V., Guldner, J., & Shi, J. (1999). Sliding mode control in electomechanical systems (2nd
ed.). Boca Raton/London: CRC Press.
Utkin, V. (2016). Discussion aspects of high-order sliding mode control. IEEE Transactions on
Automatic Control, 61(3), 829–833.
Yi, S., & Woo, P.-Y. (1997). A robust fuzzy logic controller for robot manipulators with
uncertainties. IEEE Transactions on Systems, Man, and Cybernetics, 24(4), 599–602.
Chapter 6
Robot Manipulator Control Using
Backstepping with Lagrange’s
Extrapolation and PI Compensator

Yassine Kali, Maarouf Saad, Jean-Pierre Kenné, and Khalid Benjelloun

Abstract A robust nonlinear backstepping technique with Lagrange’s extrapolation

and PI compensator is proposed in this chapter for high accuracy trajectory tracking
of robot manipulators with uncertain dynamics and unexpeted disturbances. The
proposed controller is synthesized by using Lagrangian extrapolation method with
PI compensator to estimate the uncertainties and disturbances and to deal with
the effect of hard nonlinearities caused by the estimation error while nonlinear
backstepping technique is used to ensure good tracking. The stability analysis is
accomplished recursively using appropriate Lyapunov functions candidate. As a
result, the proposed control technique shows better performances via experimental
results on a 7-DOF robot arm in comparison with the classical backstepping and
sliding mode control.

Keywords Backstepping · Lagrange’s extrapolation · PI compensator ·

Uncertain robot manipulators · Lyapunov · Trajectory tracking

6.1 Introduction

One of the challenging problems in control engineering practice is to design robust

controllers for robot manipulators in the presence of uncertainties and unexpected
disturbances. Several controllers are proposed in literature; among them, Predictive
control (Wilson et al. 2016), Fuzzy control (Al-Hadithi et al. 2007), Neural
Network control (Fierro and Lewis 1998), Backstepping (Khalil 1992; Slotine and
Li 1991), Sliding Mode control (Utkin 1992; Utkin et al. 1999). The design of

Y. Kali () · K. Benjelloun

Ecole Mohammadia d’Ingénieurs, University of Mohammed V, Rabat, Morocco
e-mail: [email protected]; [email protected]
M. Saad · J.-P. Kenné
École de Technologie Supérieure, University of Quebec, Montreal, QC, Canada
e-mail: [email protected]; [email protected]

© Springer Nature Singapore Pte Ltd. 2019 137

N. Derbel et al. (eds.), New Developments and Advances
in Robot Control, Studies in Systems, Decision and Control 175,
https://doi.org/10.1007/978-981-13-2212-9_6
138 Y. Kali et al.

those controllers is based on the mathematical model of the robots. Unfortunately,

in reality, these models are uncertain due to uncertain parameters, load variations,
unmodelled dynamics. Therefore, its difficult to fulfill the control objective.
One of the most important nonlinear control techniques is the Backstepping
approach. This control scheme was introduced in the beginning of 1990s by Koko-
tovic et al. (1995). Due to its design and implementation simplicity, Backstepping
method has been used to control many systems in different area. A recursive
Lyapunov procedure is the basic idea in the design. This idea consists of designing
a controller by considering virtual controls that represent in reality some state
variables and then, ensuring them stability by selecting appropriate stabilization
functions (Lewis et al. 1993; Slotine and Li 1991). This technique is developed
to ensure the stability of the controlled system at any time, in addition to the control
objective (tracking and/or stabilization). However, as all nonlinear controllers, this
method has many drawbacks. The major one is its sensitivity to uncertainties
and external disturbances. Otherwise, the control may easily cause unacceptable
practical complications and degrade the control objective.
To cope with this problem, an adaptive backstepping has been proposed that pro-
vide an adaptation of the control to ensure robustness towards uncertain nonlinear
dynamics and disturbances by an overparametrization that might cause inequality of
the number of parameter estimates and the number of unknown parameters (Hu et al.
2012; Shieh and Hsu 2008; Zhou and Wen 2008). In Chen and Fu (2015), Choi and
Farrell (2000), and Zhou and Er (2007), by considering the problem of controlling
and stabilizing a particular class of systems with uncertainties, an observer-based
backstepping control has been considered. Moreover, other researchers proposed
a combination of backstepping and intelligent control techniques (neural-network
and/or fuzzy logic) (Jagannathan and Lewis 1998; Su et al. 2015; Weisheng et al.
2015; Yoo and Ham 2000). It is known that these techniques have the merit to
estimate uncertainties and disturbances. However, they are difficult to implement
due to the large number of fuzzy rules introduced (case of fuzzy logic) and the large
number of parameters (case of neural-network) that increase when the number of
robot’s DOF increases.
In this chapter, a robust backstepping controller with Lagrange’s extrapola-
tion and PI compensator is designed for high accuracy trajectory tracking for
robot manipulators in the presence of uncertainties and external disturbances.
The aim of Lagrange’s extrapolation method is to provide a good estimation of
the unknown dynamics and disturbances using time delayed signals of control
torque inputs, joints position and velocity without a prior exact knowledge of
robot model (Yaramasu and Wu 2014). However, hard nonlinearities will appear
and make the controlled system leads with a large steady state error. To deal
with these hard nonlinearities, the PI compensator is used. Moreover, nonlinear
backstepping (Skjetne and Fossen 2004; Tan et al. 2000) is used to ensure robustness
and high accuracy tracking. The stability of the closed-loop system is proved
recursively using Lyapunov theory.
The rest of the chapter is organized as follows: in the next section, the dynamics
equation of n-link rigid robot manipulators is presented and the problem formulation
6 Robot Manipulator Control Using Backstepping with Lagrange’s. . . 139

is stated. The proposed control scheme and its stability analysis for the resulting
closed-loop are provided in Sect. 6.3. Simulation results are presented in Sect. 6.4,
to illustrate the superiority of the proposed technique over the classical sliding mode
and backstepping technique. In Sect. 6.5, the proposed controller is implemented in
real time on the 7-DOF ANAT robot arm and experimental results are presented.
The paper is closed Sect. 6.6.

6.2 Preliminaries

Consider the dynamics of n-DOF robot manipulator in the following matrix

equation (Spong et al. 2005):

M(q(t))q̈(t) + C(q(t), q̇(t))q̇(t) + G(q(t)) + F (q̇(t)) = τ (t) + τd (t) (6.1)

where:
• q(t), q̇(t), q̈(t) ∈ R n represent the joint space position, velocity and acceleration
vectors, respectively.
• M(q(t)) ∈ R n×n denotes the inertia matrix which is symmetric definite positive
and its inverse always exists.
• C(q(t), q̇(t)) ∈ R n×n is the centrifugal and Coriolis matrix.
• G(q(t)) ∈ R n represents the gravitational vector.
• F (q̇(t)) = Fv q̇(t) + Fs sign(q̇(t)) ∈ R n is the vector of viscous/static friction
torque at the joints with Fv , Fs ∈ R n×n are positive diagonal matrices.
• τ (t) ∈ R n is the torque input vector.
• τd (t) ∈ R n denotes the disturbance vector.
Due to load variations, unknown parameters, unmodel dynamics, the dynamic
matrices can be divided into two parts as:
⎧
⎪
⎪ M(q(t)) = M0 (q(t)) + ΔM(q(t))
⎨
C(q(t), q̇(t)) = C0 (q(t), q̇(t)) + ΔC(q(t), q̇(t))
(6.2)
⎪
⎪ G(q(t)) = G0 (q(t)) + ΔG(q(t))
⎩
F (q̇(t)) = F0 (q̇(t)) + ΔF (q̇(t))

where M0 (q(t)), C0 (q(t), q̇(t)), G0 (q(t)) and F0 (q̇(t)) are the nominal parts while
ΔM(q(t)), ΔC(q(t), q̇(t)), ΔG(q(t)) and ΔF (q̇(t)) denote the uncertain parts of
the inertia matrix M(q(t)), the centrifugal and Coriolis matrix C(q(t), q̇(t)), the
gravitational vector G(q(t)) and the vector of viscous/static friction torque at the
joints F (q̇(t)), respectively.
Then, the dynamic equation of the robot manipulator given in partitioned form
in Eq. (6.1) can be rewritten as follows:

M0 (q(t))q̈(t) + N(q(t), q̇(t)) + H (q(t), q̇(t), q̈(t)) = τ (t) (6.3)

140 Y. Kali et al.

where:

N (q(t), q̇(t)) = C0 (q(t), q̇(t))q̇(t) + G0 (q(t)) + F0 (q̇(t)) (6.4)

H (q(t), q̇(t), q̈(t)) = ΔM(q(t))q̈(t) + ΔC(q(t), q̇(t))q̇(t)
+ΔG(q(t)) + ΔF (q̇(t)) − τd (t) (6.5)

For simplicity, let us denote M0 (q(t)) = M0 (t), N(t) = N (q(t), q̇(t)) and H (t) =
H (q(t), q̇(t), q̈(t)).
In general, the control objective is to make the joint positions track a desired
trajectory with high accuracy. However, since H (t) is unknown, the control
performance will be affected. Then, the objective is to design a robust joint space
controller which ensures the convergence of the error even in the presence of
unknown dynamics and unexpected disturbances. To this end, we will design the
controller and carry out its stability analysis based on the following property and
assumptions:
Proper The nominal inertia matrix M0 (t) is symmetric definite positive and
bounded such as:

0 < m1 ≤ M0 (t) ≤ m2

with m1 and m2 are two known positive constants (Spong et al. 2005).
Assumption 1 The joint position and velocity states are available for measure-
ments.
Assumption 2 The functions Hi (t) for i = 1, · · · , n of the uncertain vector
H (t) = [H1 (t), · · · , Hn (t)]T are continuously differentiable with respect to the
time variable and don’t vary largely during a small L period of time (Toumi and Ito
1990).

6.3 Controller Design

To achieve the control objective, a combination of nonlinear backstepping and

Lagrange’s extrapolation and PI compensator will be designed in this section. The
architecture of the closed-loop system is represented in Fig. 6.1.
The control design consists on three steps described as follows:
Step 1 This step consists on defining the regulated variable to be the joint position
tracking error variable as follows:

E1 (t) = e(t)
= q(t) − qd (t) (6.6)
6 Robot Manipulator Control Using Backstepping with Lagrange’s. . . 141

Fig. 6.1 Block diagram of the proposed controller

where qd (t) ∈ R n denotes the desired joint space position trajectory. Therefore, the
first time derivative of the dynamics of the new coordinate (i.e. regulated variable)
is computed as:

Ė1 (t) = ė(t)

= q̇(t) − q̇d (t) (6.7)

Considering that q̇(t) is a control variable and defining a virtual control law β1 (t) for
Eq. (6.7), and let E2 (t) be a new error variable representing the difference between
the actual control variable and the virtual control law of (6.7) as follows:

E2 (t) = q̇(t) − β1 (t) (6.8)

Hence, the derivative of the regulated variable Ė1 (t) can be rewritten as:

Ė1 (t) = E2 (t) + β1 (t) − q̇d (t) (6.9)

In this first step, the control objective is to ensure the convergence of the regulated
variable to zero (E1 (t) → 0) by designing an appropriate virtual control law. To
that end, let us select the following Lyapunov function:
1 T
V1 (t) = E (t)E1 (t) (6.10)
2 1
Its time derivative is calculated as:

V̇1 (t) = E1T (t)Ė1 (t)

= E1T (t) (E2 (t) + β1 (t) − q̇d (t))
= E1T (t) (β1 (t) − q̇d (t)) + E1T (t)E2 (t) (6.11)
142 Y. Kali et al.

Now, the appropriate virtual control β1 (t) that will make the first order system
stabilizable is selected:

β1 (t) = q̇d (t) − k1 E1 (t) (6.12)

where k1 = diag (k11 , k12 · · · , k1n ) is a diagonal positive matrix. Then, the time
derivative of the Lyapunov function becomes:

V̇1 (t) = −E1T (t)k1 E1 (t) + E1T (t)E2 (t) (6.13)

Notice that if E2 (t) = 0, then V̇1 (t) = −E1T (t)k1 E1 (t) is negative definite and the
convergence to zero of the regulated variable E1 (t) is guaranteed. Therefore, the
aim of the next step of the design procedure is to ensure that the error between the
actual control variable and the virtual control law E2 (t) converges to zero.
Step 2 Now, let us recall the expression of E2 (t):

E2 (t) = q̇(t) − β1 (t)

= q̇(t) − q̇d (t) + k1 E1 (t) (6.14)

Differentiating the above equation with respect to time leads to:

Ė2 (t) = q̈(t) − β̇1 (t)

= q̈(t) − q̈d (t) + k1 Ė1 (t)
= M0−1 (t) [τ (t) − N(t) − H (t)] − q̈d (t) + k1 Ė1 (t) (6.15)

Therefore, using Eqs. (6.6) and (6.14), Ė1 (t) can be rewritten as follows:

Ė1 (t) = E2 (t) − k1 E1 (t) (6.16)

Thus, the second Lyapunov function is selected as:

1 1
V (t) = V1 (t) + E2T (t)E2 (t) + εT (t)Γ ε(t) (6.17)
2 2
* +
where ε(t) = M0−1 (t) Ĥ (t) − H (t) denotes the estimation error with Ĥ (t) is the
estimate of the unknown vector H (t) (its expression will be determinate later) and
Γ = diag (Γ1 , Γ2 · · · , Γn ) is an adaptive gain positive matrix.
Taking the time derivative of the second Lyapunov function (6.17) gives:

V̇ (t) = V̇1 (t) + E2T (t)Ė2 (t) + εT (t)Γ ε̇(t)

= E1T (t)Ė1 (t) + E2T (t)Ė2 (t) + εT (t)Γ ε̇(t)
6 Robot Manipulator Control Using Backstepping with Lagrange’s. . . 143


= E2T (t) M0−1 (t) [τ (t) − N(t) − H (t)] − q̈d (t) + k1 Ė1 (t)

+E1T (t)(E2 (t) − k1 E1 (t)) + εT (t)Γ ε̇(t)


= E2T (t) M0−1 (t) [τ (t) − N(t) − H (t)] − q̈d (t) + k1 Ė1 (t)

+E2T (t)E1 (t) − E1T (t)k1 E1 (t) + εT (t)Γ ε̇(t) (6.18)

Choosing:

M0−1 (t) [τ (t) − N(t) − H (t)] − q̈d (t) + k1 Ė1 (t)

+E1 (t) = −k2 E2 (t) + ε(t) (6.19)

where k2 = diag (k21 , k22 , · · · , k2n ) is a diagonal positive gain matrix, the
expression of the torque control input is obtained as:

τ (t) = M0 (t)u(t) + N(t) + Ĥ (t) (6.20)

where

u(t) = q̈d (t) − k1 Ė1 (t) − E1 (t) − k2 E2 (t) (6.21)

Step 3 In this step, the expression of Ĥ (t) will be determinate. For that, let us
substitute the obtained control law (6.20) in the derivative of the second selected
Lyapunov function (6.18) as:

V̇ (t) = −E1T (t)k1 E1 (t) + E2T (t) (−k2 E2 (t) + ε(t)) + εT (t)Γ ε̇(t)
= −E1T (t)k1 E1 (t) − E2T (t)k2 E2 (t) + εT (t) (Γ ε̇(t) + E2 (t)) (6.22)

To guarantee V̇ (t) is a negative definite function (i.e., the error variable E2 (t)
converges to zero), the following condition is set:

Γ ε̇(t) = −E2 (t) (6.23)

Integrating the above equation between 0 and t gives:

, t
ε(t) = −Γ −1 E2 (t)dt
0
,
−1
t  
= −Γ Ė1 (t) + k1 E1 (t) dt
0
 , t 
−1
= −Γ E1 (t) − E1 (0) + k1 E1 (t)dt (6.24)
0
144 Y. Kali et al.

 -t
where −Γ −1 E1 (t) − E1 (0) + k1 0 E1 (t)dt represents the PI compensator with
Γ −1 and Γ −1 k1 are the proportional gain matrix and the integral gain matrix,
respectively, and E1 (0) is the initial value of the tracking error. Remark that the
PI compensator is updated to the direction of decreasing ε(t). Consequently:
* +  , t 
M0−1 (t) Ĥ (t) − H (t) = −Γ −1 E1 (t) − E1 (0) + k1 E1 (t)dt (6.25)
0

Then, the estimate Ĥ (t) is obtained as:

 , t 
−1
Ĥ (t) = H (t) − M0 (t)Γ E1 (t) − E1 (0) + k1 E1 (t)dt (6.26)
0

where H (t) can be obtained using Lagrange’s extrapolation using the last three
points:

H (t) = 3H (t − L) − 3H (t − 2L) + H (t − 3L) (6.27)

with L is the estimation time delay (in practice, L is chosen to be the sampled time
period) and H (t − aL) for a = 1, 2, 3 is obtained using Eq. (6.3) as:

H (t − aL) = τ (t − aL) − M0 (t − aL)q̈(t − aL) − N (t − aL) (6.28)

The past acceleration q̈(t − aL) at time t − aL can be obtained using one of the
following approximations:

1 
q̈(t − aL) = q̇(t − aL) − q̇(t − (a + 1)L) (6.29)
L
1  
q̈(t − aL) = 2
q(t − aL) − 2q(t − (a + 1)L) + q(t − (a + 2)L) (6.30)
L

Finally, using the obtained expression of Ĥ (t) in Eq. (6.26), the time derivative of
the Lyapunov function V (t) becomes:

V̇ (t) = −E1T (t)k1 E1 (t) − E2 (t)T k2 E2 (t)

= −E T (t)KE(t)
≤ −λmin (K)E(t)2 (6.31)

where E(t) = [E1 (t), E2 (t)]T and λmin (K) denotes the minimum eigenvalue of
the matrix K defined by:
6 Robot Manipulator Control Using Backstepping with Lagrange’s. . . 145

( )
k1 0n×n
K= (6.32)
0n×n k2

Hence, the stability of the closed loop system is proven.

6.4 Numerical Simulations

In this section, three numerical simulations are presented for trajectory tracking
of a two-link rigid robotic manipulator shown in Fig. 6.2 using Matlab/Simulink
software with ODE 4 Solver and a sampled time equal to 0.001 s.
For this two-link manipulator, the dynamic equation (6.1) has the following
matrices,
( )
l22 m2 + 2l1 l2 m2 c2 + l12 (m1 + m2 ) + J1 l2 m2 (l1 + l2 )
M(q(t)) = ,
l2 m2 (l1 + l2 ) l22 m2 + J2
( )
−l1 l2 m2 s2 q̇22 (t) − 2l1 l2 m2 s2 q̇1 (t)q̇2 (t)
C(q(t), q̇(t))q̇(t) = ,
l1 l2 m2 s2 q̇22 (t)
( )
l2 m2 gc12 + (m1 + m2 )l1 gc1
G(q(t)) = ,
l2 m2 gc12
⎧( )
⎪
⎪ 2 sin t + 0.5 sin(200π t)
⎪
⎪ , if t < 6 s
⎪
⎨ cos 2t + 0.5 sin(200π t)
τd (t) =
⎪
⎪ ( )
⎪
⎪ 20 sin t + 5 sin(200π t)
⎪
⎩ 10 cos 2t + 5 sin(200π t) , else

Fig. 6.2 Considered two-link

robot manipulator
146 Y. Kali et al.

Table 6.1 Parameters of the two-link robot manipulator

Symbol Definition Value S.I. Unit
l1 Length of the first link 1 (m)
l2 Length of the second link 0.85 (m)
J1 Moment of inertia of the first motor 5 (kg m2 )
J2 Moment of inertia of the second motor 5 (kg m2 )
m1 Mass of link 1 0.5 (kg)
m2 Mass of link 2 1.5 (kg)
m̂1 Nominal mass of link 1 0.4 (kg)
m̂2 Nominal mass of link 2 1.2 (kg)
g Gravitational constant 9.81 (m/s2 )

where si , ci and cij are defined by si = sin(qi (t)), ci = cos(qi (t)) and cij =
cos(qi (t) + qj (t)), respectively. The parameters of the above elements are reported
in Table 6.1.
The control objective here is to follow the reference signals defined by:

q1d (t) = 1.25 − (7/5)e−t + (7/20)e−4t (6.33)

q̇1d (t) = (7/5)e−t − (7/5)e−4t (6.34)
−t −4t
q̈1d (t) = −(7/5)e + (28/5)e (6.35)
q2d (t) = 1.25 + e−t − (1/4)e−4t (6.36)
−t −4t
q̇2d (t) = −e +e (6.37)
q̈2d (t) = e−t − 4e−4t (6.38)

The initial states (joint positions and joint velocities) are selected to be q1 (0) =
0.4 rad, q2 (0) = 1.8 rad and q̇1 (0) = q̇2 (0) = 0 rad/s.

6.4.1 Results Using Sliding Mode Control

A complete study of sliding mode theory can be found in Utkin (1992). The control
law for the considered robot is given as:
!
τ (t) = M0 (q(t)) q̈d (t) − λė(t) − Ksign(σ (t)) + C0 (q(t), q̇(t))q̇(t) + G0 (q(t))

where M0 (q(t)), C0 (q(t), q̇(t)) and G0 (q(t)) are the nominal matrices calculated
using the nominal masses m̂1 and m̂2 given in Table 6.1. The used controller gains
in the simulation are:

λ = diag (2, 2), K = diag (5, 5)

The simulation results are given in Fig. 6.3.

6 Robot Manipulator Control Using Backstepping with Lagrange’s. . . 147

Fig. 6.3 Obtained simulation results via classical sliding mode control. (a) Joint 1 tracking. (b)
Joint 2 tracking. (c) Joint 1 error. (d) Joint 2 error. (e) Torque of joint 1. (f) Torque of joint 2

6.4.2 Results Using Backstepping Method

The theoretical development of the nonlinear backstepping scheme can be found

in Slotine and Li (1991) and Liu and Wang (2012). The control law for robot is
given as:
!
τ (t) = M0 (q(t)) q̈d (t) − (λ1 + λ2 )ė(t) − (I + λ1 λ2 )e(t)
+C0 (q(t), q̇(t))q̇(t) + G0 (q(t))

In this simulation, the used parameters for sliding mode control using saturation
function with time delay estimation are:

λ1 = diag (2, 2), λ2 = diag (3, 3)

The obtained simulation results are shown in Fig. 6.4.

148 Y. Kali et al.

Fig. 6.4 Obtained simulation results via backstepping method. (a) Joint 1 tracking. (b) Joint 2
tracking. (c) Joint 1 error. (d) Joint 2 error. (e) Torque of joint 1. (f) Torque of joint 2

6.4.3 Results Using the Proposed Method

The proposed nonlinear backstepping with Lagrange’s extrapolation and PI com-

pensator simulated on the two-link robot is given in Eq. (6.20) where the controller
gains are chosen as:

K1 = diag (2, 2), K2 = diag (5, 5)

Γ −1 = diag (0.04, 0.04), L = Ts = 0, 001 s

Moreover, the time delayed accelerations q̈(t − aL) for a = 1, 2, 3 are calculated
using the second approximation given in Eq. (6.30). The simulation results are
depicted in Fig. 6.5.

6.4.4 Discussion
6.4.4.1 Classical Sliding Mode Controller

The joint positions q1 and q2 shown respectively in Fig. 6.3a, b follow their desired
references properly. However, Fig. 6.3e, f show that the chattering phenomenon and
6 Robot Manipulator Control Using Backstepping with Lagrange’s. . . 149

Fig. 6.5 Obtained simulation results via the proposed controller. (a) Joint 1 tracking. (b) Joint 2
tracking. (c) Joint 1 error. (d) Joint 2 error. (e) Torque of joint 1. (f) Torque of joint 2

the control effort are very important due to the large choice of the switching gains
which give an unacceptable torque inputs.

6.4.4.2 Nonlinear Backstepping

Figure 6.4e, f show that the control effort is very reduced when it is compared to the
obtained results with the robust classical sliding mode. However, the robot system
do not converge to zero due to its sensitivity to uncertainties and disturbances as
depicted in Fig. 6.4c, d.

6.4.4.3 Proposed Backstepping with Lagrange’s Extrapolation and PI

Compensator

The obtained simulation results show the effectiveness of the proposed nonlinear
backstepping sliding mode method with Lagrange’s extrapolation. The joints q1 and
q2 shown respectively in Fig. 6.5a, b converge with high accuracy to their desired
trajectories. This is confirmed in the tracking error figures (Fig. 6.5c, d). In addition,
the control torque inputs show acceptable effort and values as depicted in Fig. 6.5e, f.
150 Y. Kali et al.

From the above obtained results, the proposed controller shows the superiority
such that the steady state error converges to zero faster due to the good estimation
of uncertainties and disturbances and due to the compensator of hard nonlinearities.
The sliding mode control can achieve high accuracy tracking. However, the
chattering is present and the control effort is high, which can damage the controlled
robot, while the classical backstepping can’t stabilize the robot system because its
sensitivity to uncertainties and external disturbances.

6.5 Case Study

6.5.1 System Description

In this section, the real time implementation of the proposed optimal super-
twisting with time delay estimation on the 7-DOF ANAT robot shown in Fig. 6.6
using Simulink with Real-Time Workshop is presented. ANAT (Articulated Nimble
Adaptable Trunk) Robot is built by Robotics Design inc. (Canada). The first joint of
the used ANAT is prismatic, followed by six rotary joints (Kali et al. 2017).
As shown in Fig. 6.7, the frames of references are selected using the modified
Denavit-Hartenberg (Spong et al. 2005) (D-H) convention. The D-H parameters that
determine the homogeneous transformation matrices are given in Table 6.2.
The initial task space position of the end-effector that is placed on the last joint
(7th joint) is: x(0) = [0.6764 0 − 0.19]T while the initial joint positions are 0 rad
and joint velocities are 0 rad/s. The equation of motion of the ANAT robot are as in
Eq. (6.3). In addition, the ANAT model verifies the assumptions given in Sect. 6.2.

Fig. 6.6 7-DOF ANAT robot

arm
6 Robot Manipulator Control Using Backstepping with Lagrange’s. . . 151

Fig. 6.7 D-H frames of the

ANAT robot arm

Table 6.2 D-H parameters Joints αi−1 (rad) ai−1 (m) di (m) qi (rad)
of the 7-DOF ANAT robot
1 0 0 q1 0
2 0 L1 0 q2
3 0 L0 0 q3
4 0 L0 0 q4
5 0 L0 L2 q5
6 π/2 L3 0 q6
7 −π/2 0 −L4 q7

6.5.2 Real-Time Setup

All robot joints are actuated by DC motors. The actuators encoders provide position
measurements. The ATMEGA 16 microcontrollers are used to send the angles
positions to simulink, and using the forward kinematics, the joint space trajectory
is transformed to the workspace trajectory. The microcontrollers are also used to
transform the control input signals into a Pulse Width Modulation (PWM) signal.
Then, this PWM signal is applied to the H-bridge drive of the actuators of the 7-
DOF ANAT robot. A current sensor located in the H-bridge drive provides the
current measurement of each actuator. In the experimentation, the sixth, and seventh
joints are chosen to stay at their initial position. A sampling time Ts = 0.03 s has
been selected for the control loop (Kali et al. 2017). The above real-time setup is
illustrated in Fig. 6.8.
152 Y. Kali et al.

Fig. 6.8 Experimental setup

6.5.3 Controller Setting

The proposed nonlinear backstepping with Lagrange’s extrapolation has three

diagonal matrices with constant elements. This makes the controller setting simplify,
which is briefly described as follows:
• Determine the matrix K1 of the virtual control β1 in Eq. (6.12).
• Tuning the matrix K2 such as a fast convergence will be obtained.
• Select L as small as possible (equal to the sampling time interval Ts ).
• Finally the constants Γi for i = 1, · · · , n are chosen such as the condition of
stability in Eq. (6.24) is verified, they should be further tuned to reduce the effect
of hard nonlinearities caused by Lagrange’s extrapolation.

6.5.4 Experimental Results

The control objective here is to track a triangular desired trajectory. For this
application the desired trajectory is chosen such as all kinematic singularities are
avoided, the relation between the desired accelerations in the task space and the
joint space is given by:

q̈d (t) = J + ẍd (t) − J + J˙J + ẋd (t) (6.39)

6 Robot Manipulator Control Using Backstepping with Lagrange’s. . . 153

where J + = J T (J J T )−1 denotes the generalized inverse of the Jacobian matrix

while J˙ is its derivative, q̈d (t) is the desired joint acceleration vector, ẍd (t) and ẋd (t)
are the task space desired acceleration and velocity vectors, respectively. The desired
velocity in the joint space q̇d (t) is obtained from q̈d (t) using an integrator while
the desired position in the joint space qd (t) is obtained from q̇d (t) using another
integrator.
In Sect. 6.2, the proposed controller that ensures the convergence of the states
trajectories to is given in Eq. (6.20). The time delayed acceleration q̈i (t − aL) for
a = 1, 2, 3 and for i = 1, · · · , 7 are obtained using the approximation based on the
joint position as:

1 !
q̈i (t − aL) = qi (t − aL) − 2qi (t − (2 + a)L) + qi (t − (3 + a)L)
Ts2

while the controller gains are reported in Table 6.3.

The experimental results are shown in Figs. 6.9, 6.10, 6.11, 6.12, and 6.13: 2D
task space tracking (The arrows in Fig. 6.9 shows the direction of displacements), X
and Y tracking, joint space tracking position, joint space tracking error and control

Table 6.3 Controller Parameter Value

parameters
λ diag (0.75; 0.75; 0.75; 0.75; 0.75; 0.75; 0.75)
k1 diag (1.5; 1.5; 1.5; 1.5; 1.5; 1.5; 1.5)
k2 diag (2; 2; 2; 2; 2; 2; 2)
Γ −1 diag (0.02; 0.02; 0.02; 0.02; 0.02; 0.02; 0.02)

Fig. 6.9 2D Task space tracking

154 Y. Kali et al.

Fig. 6.10 Task space tracking position (m). (a) X-position tracking. (b) Y-position tracking.

torque inputs. In these results, the performances produced by the proposed controller
are illustrated for the ANAT robot by neglecting some parts of dynamics and by
adding a load of 5.5 kg on the fifth joint.
The proposed controller ensure the convergence of the end-effector position to
the task space desired position with high accuracy due to a good estimation of
uncertainties as shown in Fig. 6.10. The joint space tracking is good as depicted in
Fig. 6.12 and confirmed by the small joint space tracking error in Fig. 6.13. Notice
that the PI compensator eliminate the hard nonlinearities that can affect the tracking
accuracy.
6 Robot Manipulator Control Using Backstepping with Lagrange’s. . . 155

Fig. 6.11 Task space tracking error (m). (a) X-position tracking error. (b) Y-position tracking error

Furthermore, it can be seen from Fig. 6.14 that the control torque inputs are
smooth with small values that are acceptable for our robot, this is thanks to the
good estimation.

6.6 Conclusion

For a class n-DOF uncertain robot manipulators in presence of uncertainties and

unexpected perturbations, a robust backstepping control with Lagrange’s extrapola-
tion and PI compensator is presented in order to achieve ourcontrol objective with
156 Y. Kali et al.

Fig. 6.12 Joint space tracking position (rad). (a) Joint 2 tracking. (b) Joint 3 tracking. (c) Joint 4
tracking. (d) Joint 5 tracking
6 Robot Manipulator Control Using Backstepping with Lagrange’s. . . 157

Fig. 6.12 (continued)

Fig. 6.13 Joint space tracking error (rad). (a) Joint 2 tracking error. (b) Joint 3 tracking error. (c)
Joint 4 tracking error. (d) Joint 5 tracking error
158 Y. Kali et al.

Fig. 6.13 (continued)

high accuracy. The control scheme is designed step-by step and the stability analysis
is established using a Lyapunov functions candidate. The proposed controller allows
good estimation of unknown dynamics and disturbances. Experimental results
on the ANAT robot arm showed the effectiveness of the proposed controller,
remarkably improved tracking performance even in presence of uncertain dynamics
and unexpected disturbances. Further research should be pursued to deal with the
problem of uncertain kinematics and to implement the proposed controller on other
nonlinear systems.
6 Robot Manipulator Control Using Backstepping with Lagrange’s. . . 159

Fig. 6.14 Joint control torques (Nm). (a) Torque of joint 2. (b) Torque of joint 3. (c) Torque of
joint 4. (d) Torque of joint 5
160 Y. Kali et al.

Fig. 6.14 (continued)

Acknowledgements The authors are grateful to Nabil Derbel (University of Sfax, Tunisia),
Jawhar Ghommam (University of Tunis, Tunisia) and Quanmin Zhu (University of the West of
England) for the opportunity to contribute to the New developments and advances in the field of
Robotics.

References

Al-Hadithi, B. M., Matía, F., & Jiménez, A. (2007). Fuzzy controller for robot manipulators
(pp. 688–697). Berlin/Heidelberg: Springer.
Chen, S.-H., & Fu, L.-C. (2015). Observer-based backstepping control of a 6-DOF parallel
hydraulic manipulator. Control Engineering Practice, 36, 100–112.
Choi, J. Y., & Farrell, J. (2000). Observer-based backstepping control using online approximation.
American Control Conference, 5, 3646–3650.
Fierro, R., & Lewis, F. L. (1998). Control of a nonholonomic mobile robot using neural networks.
IEEE Transactions on Neural Networks, 9(4), 589–600.
Hu, Q., Xu, L., & Zhang, A. (2012). Adaptive backstepping trajectory tracking control of robot
manipulator. Journal of the Franklin Institute, 349(3), 1087–1105.
Jagannathan, S., & Lewis, F. (1998). Robust backstepping control of robotic systems using neural
networks. Journal of Intelligent and Robotic Systems, 23, 105–128.
Kali, Y., Saad, M., Benjelloun, K., & Fatemi, A. (2017). Discrete-time second order sliding mode
with time delay control for uncertain robot manipulators. Robotics and Autonomous Systems,
94, 53–60.
Khalil, H. (1992). Nonlinear systems. New York: Macmillan Publishing Company.
Kokotovic, P., Krstic, M., & Kanellakopoulos, I. (1995). Nonlinear and adaptive control design.
New York: Wiley.
Lewis, F., Abdallah, C., & Dawson, D. (1993). Control of robot manipulators. New York:
Macmillan Publishing Company.
Liu, J., & Wang, X. (2012). Advanced sliding mode control for mechanical systems.
Berlin/Heidelberg: Springer.
6 Robot Manipulator Control Using Backstepping with Lagrange’s. . . 161

Shieh, H.-J., & Hsu, C.-H. (2008). An adaptive approximator-based backstepping control approach
for piezoactuator-driven stages. IEEE Transactions On Industrial Electronics, 55, 1729–1738.
Skjetne, R., & Fossen, T. (2004). On integral control in backstepping: Analysis of different
techniques. In American Control Conference, Boston, Massachusetts.
Slotine, J., & Li, W. (1991). Applied nonlinear control. Taipei: Prentice-Hall International.
Spong, M., Hutchinson, S., & Vidyasagar, M. (2005). Robot modeling and control. Wiley.
Su, C.-Y., Li, G., Li, Z., & Su, H. (2015). Fuzzy approximation-based adaptive backstepping
control of an exoskeleton for human upper limbs. IEEE Transactions on Fuzzy Systems, 23,
555–566.
Tan, Y., Chang, J., Tan, H., & Jun, H. (2000). Integral backstepping control and experimental
implementation for motion system. In IEEE International Conference on Control Applications,
Anchorage, AK.
Toumi, K., & Ito, O. (1990). A time delay controller for systems with unknown dynamics. ASME
Journal of Dynamic System, Measurement and Control, 112, 133–141.
Utkin, V. (1992). Sliding mode in control and optimization. Berlin: Springer.
Utkin, V., Guldner, J., & Shi, J. (1999). Sliding mode control in electromechanical systems (2nd
ed.). Boca Raton, London.
Weisheng, C., Ge, S., Jian, W., & Maoguo, G. (2015). Globally stable adaptive backstepping neural
network control for uncertain strict-feedback systems with tracking accuracy known a priori.
IEEE Transactions on Neural Networks and Learning Systems, 26, 1842–1854.
Wilson, J., Charest, M., & Dubay, R. (2016). Non-linear model predictive control schemes
with application on a 2 link vertical robot manipulator. Robotics and Computer-Integrated
Manufacturing, 41, 23–30.
Yaramasu, V., & Wu, B. (2014). Model predictive decoupled active and reactive power control
for high-power grid-connected four-level diode-clamped inverters. IEEE Transactions on
Industrial Electronics, 61(7), 3407–3416.
Yoo, B. K., & Ham, W. C. (2000). Adaptive control of robot manipulator using fuzzy compensator.
IEEE Transactions on Fuzzy Systems, 8, 186–199.
Zhou, J., & Er, M. J. (2007). Adaptive output control of a class of uncertain chaotic systems.
Systems and Control Letters, 56(6), 452–460.
Zhou, J., & Wen, C. (2008). Adaptive backstepping control of uncertain systems. Berlin: Springer.
Chapter 7
Nonlinear Observer-Based Fault
Detection and Isolation for a Manipulator
Robot

Khaoula Oulidi Omali, M. Nabil Kabbaj, and Mohammed Benbrahim

Abstract Fault Detection and Isolation (FDI) techniques in robot manipulator is

becoming one of the most phenomena in robotics in order to ensure higher levels
of safety and productivity. Research, has been produced a considerable effort in
seeking systematic approaches to fault detection for both linear and nonlinear
dynamical systems. In the last decade considerable research efforts have been spent
to seek for systematic approaches to Fault Detection (FD) in dynamical systems.
Special attention has been addressing for robotic systems, especially for those
operating in remote or hazardous environments, where a high degree of safety as
well as self-detection capabilities are required. On the other hand, the development
of effective strategies of fault detection for robot manipulators operating in an
industrial context is a critical research task. Several FD techniques for robot
manipulators have been proposed in the literature, although the problem of their
application to industrial robots has not been extensively investigated.
In this chapter, we present a high-gain observer based fault detection and isola-
tion scheme for a class of affine nonlinear systems. In order to test the effectiveness
and the robustess of the proposed approach, a case study is developed for a special
robot manipulator named Articulated Nimble Adaptable Trunk “ANAT” with a five-
degree-of-freedom in order to detected and isolated sensor fault.

Keywords Fault detection and isolation · Robot manipulator · High-gain

observer · Residuals · Nonlinear systems

K. Oulidi Omali () · M. N. Kabbaj · M. Benbrahim

Sidi Mohamed Ben Abdellah University Fez, Fez, Morocco
e-mail: [email protected]; [email protected];
[email protected]

© Springer Nature Singapore Pte Ltd. 2019 163

N. Derbel et al. (eds.), New Developments and Advances
in Robot Control, Studies in Systems, Decision and Control 175,
https://doi.org/10.1007/978-981-13-2212-9_7
164 K. O. Omali et al.

7.1 Introduction

Robotics is a relatively young field of modern technology that crosses tradi-

tional engineering boundaries. Understanding the complexity of robots and their
applications requires knowledge of electrical engineering, mechanical engineering,
systems and industrial engineering, computer science, economics, and mathematics.
New disciplines of engineering, such as manufacturing engineering, applications
engineering, and knowledge engineering have emerged to deal with the complexity
of the field of robotics and factory automation.
Recently, robot manipulators are the most important components of manufactur-
ing and control processes. They have the impact: Improving productivity, increasing
the quality of manufactured products, and reducing the cost of labor.
Associated with the increasing industrial demands on safety, reliability, depend-
ability and assuring the normal operation of this robot manipulator, the issue of FDI
is important.
The most important steps are fault detection and fault isolation. The methods
of fault detection can be separated into methods which don’t use plant models and
model-based approaches. For the first category are classical limit and trend value
checks as well as signal analysis methods, for example, autocorrelation techniques,
spectral analysis, etc. Normally, those methods need dedicated sensor systems.
Observer-based, parameter estimation and parity relations techniques, on the other
hand, the second class of fault detection schemes are the most relevant approaches.
Therefore, different methods have been developed for detecting and isolating
actuator and sensor faults on robot manipulators, such as parameter estimation,
parity relations and observer based approaches (Filaretov et al. 2003; Jollie 2016;
Ma and Yang 2016; Schneider and Frank 1996).
In the literature, the parameter estimation method requires knowledge of the
model structure of the investigated process and actual process measurements. The
actual values of model parameters are then determined from this information. The
fault detection is accomplished by a comparison of identified with predetermined
parameter values by using thresholds. Note that parameter estimation methods
require appropriate excitation of the system. An overview of theory and applications
of parameter estimation methods can be found in Isermann (1984) (Fig. 7.1).
The parity space approach for nonlinear systems is studied only in a few papers
(Guernez et al. 1997; Krishnaswami and Rissoni 1994; Mironovsky 1989; Shumsky
1998). The papers Mironovsky (1989) and Shumsky (1998) consider the problem of
fault detection only; Krishnaswami and Rissoni (1994) gives only the most general
recommendations to obtain the parity equations; Guernez et al. (1997) extends
the parity space approach to nonlinear polynomial dynamic systems. It is known,
however, that the description of manipulation robots includes non polynomial, non
linearities such as backlash and Coulomb friction (Filaretov et al. 1999), so the
approach (Guernez et al. 1997) cannot be used in this case. There are many papers
where the parity space approach was considered for linear systems (see surveys
7 NLO Based FDI for a Manipulator Robot 165

Fig. 7.1 Parameter

estimation scheme

Fig. 7.2 Parity space scheme

Frank 1990; Gertler and Monajemy 1995; Patton 1994). It is well known (Frank
1990; Gertler 1993; Gertler and Monajemy 1995) that the parity space approach
is equivalent to the observer-based one in the linear case. However, they aren’t
equivalent in the nonlinear case (Fig. 7.2).
Observer-based method requires a model of the investigated process also. The
model is operated parallel to the real process, the inputs to the model are the same
as those to the real process. In contrast to classical state-space observers, e.g.,
Kalman filters, the diagnostic observers are output observers which are operated
in an open-loop configuration. Assuming an exact model of the plant, the difference
of measured and calculated process outputs (which is called a residual) will produce
a nonzero value when a fault has occurred. A thorough overview of robust observer-
based fault detection methods can be found in Frank (1990, 1993). The main
advantage of observer-based methods over parameter estimation methods is given
by the fact that no special excitation is required. The observer-based method will
work in steady-state operation conditions as well.
In any of the systems that were discussed above, in order to have the efficient
operation of the process and to increase the reliability and safety, prompt detection
of anomalous situations (fault detection) and the fast identification (isolation) of
the most probable causes (faults) need to be addressed. FDI can be carried out
using analytical or functional information about the system being monitored, i.e.,
based on a mathematical model of the system. This approach is known as analytical
redundancy, which is also known as model based or quantitative FDI. Model based
FDI is currently the subject of extensive research and is being used in highly reliable
control systems due to the fact that analytical redundancy based techniques are
more economical and at times more powerful. These methods have the potential
of detecting soft incipient faults even during the system’s transient operation. One
166 K. O. Omali et al.

Fig. 7.3 Basic configuration of observer-based fault diagnosis

of the most popular of the model based approaches to FDI is the observers based
techniques. The basic idea of observers-based methods consists reconstruct some
or all system outputs from accessible process variables. The fault indicators are
generated by forming the differences between the estimated outputs and the actual
outputs. However, special attention has to be paid when applying observer theory
for fault detection and isolation. The basic configuration of observer-based fault
diagnosis is shown in Fig. 7.3.
One of the most challenging topics in observer based in the FDI research and
application areas is high-gain observers.This type of observers are relatively fast and
simple to design as you don’t need to solve complex differential equations nor use
complicated formulaes, also for large class of nonlinear systems, they can provide
global or semi-global stability results for class of systems. This means that their
can provide stability guarantees for any arbitrarily initial conditions. For nonlinear
systems represented in the normal form, it was shown that when high-gain observers
are used in output feedback control they can recover the performance of the state
feedback controller. In other words, assuming that all the states are measured, you
can design a controller that time. Then, when the states are provided by the high-gain
observers, the output feedback controller can provide a very similar performance to
the state feedback case, therefore, they can be robust to modeling uncertainty and
7 NLO Based FDI for a Manipulator Robot 167

external disturbances. They were first introduced by Gauthier et al. (1992). They
are non-linear observers that take into account the non-stationarity, non-linearity of
industrial processes and at the same time ensure a good estimate of the real state
with easy adjustment of Gain vector (Farza et al. 2005; Koubaa et al. 2004; Nadri
2001).
High-gain observers have been developed around 30 years ago. In the works
(Esfandiari and Khalil 1987; Khalil and Saberi 2007; Saberi and Sannuti 1990;
Tornambe 1988), also the recent survey (Khalil and Praly 2014). Choosing the
observer gain K large enough, the observer error can be made arbitrarily small,
see e.g. Tornambe (1992). The advantages of high-gain observers is that they
can be used to estimate the system states without knowing the exact parameters;
only some structural assumptions, such as a known relative degree, are necessary.
Furthermore, they are robust with respect to input noise. The drawback is that it is
not known a priori how large K must be chosen and appropriate values must be
identified by offline simulations. If K is chosen unnecessarily large, the sensitivity
to measurement noise increases dramatically.
In Gauthier et al. (2001), the authors gave a necessary and sufficient condition
that characterizes the class of nonlinear affine systems mono-input, mono-output
observable for any input. The authors shows that this class of systems is diffeo-
morphic to the canonical form which consists of a fixed linear dynamic and a
triangular nonlinear dynamic. Using this structure, the authors proposed a high-gain
observer synthesized under a Lipschitz hypothesis on nonlinear dynamics. The gain
of the proposed observer is derived from the resolution of an algebraic Lyapunov
equation which can be explicitly computed. Several generalizations of this result
to multi-input systems are proposed in Busawon et al. (1998), Farza et al. (2004),
Gauthier and Kupka (2001), Hammouri and Farza (2003), Hou et al. (2000) and
Shim et al. (2001). The main characteristic of the high-gain observer proposed
in Farza et al. (2005), is its simplicity of implementation since the observer is a
copy of the dynamics of the system with a gain whose expression can be explicitly
calculated. In addition, the observer setting can be completed by selecting a single
synthesis parameter. The effectiveness of this type of observer is proved by several
industrial applications (Farza et al. 1999; Hammouri et al. 2001). Indeed, some tests
are necessary for a judicious choice of a gain ensuring the compromise between
the speed of convergence and the robustness with respect to measurement noises. In
this chapter, high-gain observer will be applied to a robot manipulator in objective
to detect and isolate the faults.
The chapter is organized as follows. The next section exposes system description
and modeling. High-gain observers used for the design of residuals generators are
developed in Sect. 7.3. Section 7.4 provides the application for a robot manipulator.
The conclusion is given in Sect. 7.5.
168 K. O. Omali et al.

7.2 System Description and Modeling

Consider the MIMO nonlinear system with m inputs and p outputs defined by the
following state representation (Slotine and Weiping 1991):
⎧
⎨ẋ = f (x, d, t) + . gi (x, t)ui (t)
⎪ m

i=1 (7.1)
⎪
⎩y = Cx

where x ∈ Rn is the state variable vector, u(t) ∈ Rm is the input control vector
and y ∈ Rp is the output vector, f (x, d, t) is the n-dimensional unknown nonlinear
dynamics and d is the disturbance, g(x, t) is the (n×m) nonlinear control dynamics
matrix, C is the (p × n) output distribution matrix.
Such systems must satisfy the following assumptions:
A1 : the integers m, p and n are known such as m ≤ n and p = m;
A2 : the system is controllable and observable, in limit detectable and stabilizable;
A3 : the unknown nonlinear dynamics and unexpected disturbances are continu-
ously differentiable with respect to the time variable.

7.2.1 The ANAT Robot

A.N.A.T. (Articulated Nimble Adaptable Trunk) technology is a new innovation

in robotic architecture that allows the creation of highly robust and intelligent
robots with reconfigurable modular architectures. This technology was invented
by the founder of Robotics Design Inc., Mr. Charles Kharallah. In use in over
seven ground-breaking products worldwide, ANAT technology represents robotic
architecture of the future.
ANAT was exclusively designed and patented by Robotics Design Inc. (US
patent 6323 615) ANAT has many properties that give it flexibility and dexterity in
larger movements than most existing manipulators. This asset comes mainly from
the robot’s redundant modular trunk, which enables it to perform auxiliary obstacle
avoidance tasks, and therefore allows it to work in more complex and difficult-
to-access workspaces. In addition, the robot’s redundancy ensures the main task
continues even in the case where one or more motors of the trunk are defective.
Thus, it is possible to implement on ANAT avoidance tasks of obstacle in the
cartesian space, provided that they have the necessary resources for the realization.
The invention consists of a series of motorized U and H shaped modules that
function much like cells of the human body; they work together to achieve a
common goal. These modules connect to each other along four points, allowing
them to evenly distribute pressure amongst themselves like the simple yet timeless
design of the Roman arches. This allows ANAT robots to carry exceptionally heavy
7 NLO Based FDI for a Manipulator Robot 169

Fig. 7.4 Robot ANAT

payloads, and withstand pressure applied on any point of their modules. ANAT
modules can be configured into a diverse spectrum of robot configurations such as
fixed manipulators, mobile robots, flexible ergonomic arms (up to 32 D.O.F.), and
much more. Motors are contained in the central axis of each module, so regardless
of the quantity or impact point of outside pressure placed on the module, no strain
is placed on the motor. These modules have one degree of freedom each, bend and
fold relatively to each other along their axis, are highly flexible and able to avoid
obstacles with ease. Modules can be connected in a myriad of configurations like
Lego blocks, and can be reconfigured from their initial form to form another robot.
This allows the same modular technology used in the design of an arc-welder or pick
and place manipulator to be re-used in forming a mobile robot such as an unmanned
mining vessel or vehicle. A robot using this simple and innovative architecture
which performs a single specialized application can be formed by modules varying
only in size, or a single robot can be formed to specialize in several applications,
with the only variable being attached accessories. This also drastically simplifies
and reduces costs for maintenance, as a faulty module can be easily replaced with
an identical one. In short, nimble, adaptable and durable modular robots are now
possible through our technology, and new products are born into the ANAT robot
family daily (Fig. 7.4).
Figure 7.5 below shows the schematic representation of the robot. It has seven
degrees of freedom composed as follows: a first prismatic joint, a part redundant
system consisting of three parallel rotary joints. Finally, an effective effector to
orient the tool in a desired position and consisting of three rotary joints whose axes
of rotation are perpendicular to one another.
We say redundancy when the number of degrees of freedom of the manipulator
is greater than the number of degrees of freedom of the task to be performed. When
there is redundancy, there is an infinity of possible positions for a fixed placement
of the tool; thus some members of the manipulator may be in motion. Let us call it
the internal movement of the robot.
Redundancy allows to optimize the performance of a robot and to avoid
constraints imposed by the articular limits or the obstacles.
170 K. O. Omali et al.

Fig. 7.5 ANAT robot

Fig. 7.6 One of the robot

modules

An interesting property of ANAT lies in its modularity which facilitates its

maintenance and performs “Plug and Play” tasks. Figure 7.6 shows one of the ANAT
modules.
Moreover, the mechanical structure of the robot also has interesting advantages
over conventional manipulators. First the aluminum structure of ANAT makes it
7 NLO Based FDI for a Manipulator Robot 171

Table 7.1 Robot workspace Articulation Type Space

1 Prismatique of L0 =0.57 m at 1.27 m
2 Rotative of −90◦ +90◦
3 Rotative of −90◦ +90◦
4 Rotative of −90◦ +90◦
5 Rotative of −90◦ +90◦
6 Rotative of −90◦ +90◦
7 Rotative of −nπ at nπ

relatively light. On the other hand, the load distribution on the robot motors is
optimized, in contrast to the usual manipulators on which the stresses are unequally
distributed. Therefore, for common manipulators, motors may be oversized to
accomplish the same task. The industrial applications of such a robot are multiple.
Indeed, ANAT can be particularly used in the paint and finishing industries, the
automotive industries, as well as the aeronautical and naval industries.
The robot is subjected to some mechanical displacement stresses qm shown in
Table 7.1.
The system (7.1) comes from the dynamic equation of robot (7.2). So the
dynamics of n-DOF robot manipulator in the following matrix equation (Craig
1989):

M(q)q̈ + C(q, q̇)q̇ + G(q) + F q̇ = τ + τd (7.2)

where q ∈ Rn is the vector of the generalized coordinates in the joint space, q̇, q̈
∈ Rn the joint velocity and acceleration vectors, respectively, M(q) ∈ Rn×n is the
inertia matrix, C(q, q̇) ∈ Rn×n is the centrifugal and coriolis matrix, G(q) ∈ Rn
is the gravitational vector, F q̇ ∈ Rn is the vector of viscous friction torque at the
joints, τ , τd ∈ Rn denotes the disturbance and torque input vectors, respectively.
The dynamic equation of the robot manipulator given in partitioned form in
Eq. (7.2) can be rewritten as follows: with the disturbance and the viscous friction
are negligibles. (τd = 0), (F q̇ = 0).

q̈ = −M(q)−1 F (q, q̇) + M(q)−1 τ (7.3)

where M is the inertia matrix, which is symmetric and positive definite. Thus,
M(q)− 1 always exits. F is the centrifugal, coriolis, and gravity vector; q is the joint
position vector; τ is the torque input vector of the manipulator. Let x = [X1T , X2T ]T
the state vector with X1 = [q1 , q2 , q3 , q4 , q5 ]T and X2 = [q˙1 , q˙2 , q˙3 , q˙4 , q˙5 ]T , and
y = X1 is the output vector. The description of the system can be given in state
representation form as follows:
172 K. O. Omali et al.

⎧
⎪
⎪ ẋ1 = x6
⎪
⎪
⎪
⎪
⎪
⎪ ẋ2 = x7
⎪
⎪
⎪
⎪ẋ3 = x8
⎪
⎪
⎪
⎪
⎪
⎪ ẋ4 = x9
⎪
⎪
⎪
⎪ = x10
⎪
⎪ẋ5
⎪
⎪
⎪
⎪ .
5
⎪
⎨ẋ6 = f1 (x, d, t) + g1i (x, t)ui (t)
i=1
(7.4)
⎪
⎪ .5
⎪
⎪ ẋ = f2 (x, d, t) + g2i (x, t)ui (t)
⎪ 7
⎪
⎪
⎪
i=1
⎪
⎪ .5
⎪
⎪ ẋ = f (x, d, t) + g3i (x, t)ui (t)
⎪
⎪ 8 3
⎪
⎪ i=1
⎪
⎪ .5
⎪
⎪
⎪
⎪ ẋ9 = f4 (x, d, t) + g4i (x, t)ui (t)
⎪
⎪
⎪
⎪
i=1
⎪
⎪ .5
⎪
⎩ẋ10 = f5 (x, d, t) + g5i (x, t)ui (t)
i=1

where

g(x, t) = M(q)−1

ui = τi , for i=1:5

f (x, d, t) = −M(q)−1 F (q, q̇)

7.3 High Gain Observer

In general, an observer is a dynamic system that provides estimations of the current

state of the system, by using the previous knowledge of the inputs and outputs of
the system.
Consider the following class of affine nonlinear system:
⎧
⎨ẋ = f (x, d, t) + . gi (x, t)ui (t)
⎪ m

i=1 (7.5)
⎪
⎩y = Cx

The system (7.5) has the input u(t) ∈ Rm which has a set of admissible values of the
input. It is also assumed that there exists a physical domain Ω ∈ Rn (open, bounded)
of evolution of the input and that is the domain of interest of the system.
7 NLO Based FDI for a Manipulator Robot 173

Suppose that system (7.5) is observable in the sense of rank and that u = 0 is an
universal input, then the jacobian {h1 , Lf h1 , . . . , Ln−1 n−1 n−1
f h1 , h2 , Lf h2 , Lf hp }. In
the surroundings of a regular point one can select a subset of full rank:

φ = {z1 , . . . , zn }{h1 , Lf h1 , . . . , Ln−1 n−1 n−1

f h 1 , h 2 , Lf h 2 , Lf h p } (7.6)
.p
with k=1 ηk = n and L is the lie derivative.
The input hk (x) belongs in the order ηk . This determines a local coordinate
system in which the system (7.5) is written as:

ż = Az + ϕ̃(z) + ϕ̄(z)u
(7.7)
y = Cz

with z ∈ Rn , u ∈ Rm , y ∈ Rp
⎡ ⎤ ⎡ ⎤
A1 C1
⎢ .. ⎥ ⎢ .. ⎥
A=⎣ . ⎦; B=⎣ . ⎦; (7.8)
Ap Cp
⎡ ⎤ ⎡ ⎤
ϕ˜1 (z) ϕ¯1 (z)
⎢ ⎥ ⎢ ⎥
ϕ̃(z) = ⎣ ... ⎦ ; ϕ̄(z) = ⎣ ... ⎦ ; (7.9)
ϕ˜p (z) ϕ¯p (z)

with
⎡ ⎤ ⎡ ⎤
0 1 ... 0 0
⎢ .. .. . . .. ⎥ ⎢ .. ⎥
⎢ . . .⎥ ⎢ ⎥
Ak = ⎢ . ⎥; ϕ˜k (z) = ⎢ . ⎥ ; (7.10)
⎣0 0 . . . 1⎦ ⎣ 0 ⎦
0 0 ... 0 ϕ˜k (z)
⎡ ⎤
ϕ1¯ k(z)
⎢ ⎥ !
ϕ¯k (z) = ⎣ ... ⎦ ; Ck = 1 0 . . . 0 (7.11)
ϕη¯k k (z)

ηk
with ϕ˜k (z) = Lf hk , ϕ̄i k(z) = Lg Li−1
f hk ; dim (Ak ) = (ηk × ηk ),
dim (Ck ) = (1 × ηk ) and dim (ϕ˜k (z)) = dim (ϕ¯k (z)) = (ηk × 1) for k = 1 . . . p,
i = 1 . . . ηk .
174 K. O. Omali et al.

For the following theorem, is not used the linearity in u, the following system is
then considered:

ż = Az + ϕ(z, u)
(7.12)
y = Cz

Let K be a matrix (n × p, p) such that

⎡ ⎤
K1
⎢ .. ⎥
K=⎣ . ⎦ (7.13)
Kp

(with kk of dimension n × 1), such that for each block k, the matrix Ak − Kk C have
negative real parts. Then the system is uniformly locally observable, and it exists
T0 > 0, such that, for every T , such 0 < T < T0 , the following system Constitutes
an observer for the system (7.12):

z + ϕ(/
ż = A/ /
z, u) + Λ−1 (T , δ)K(y − C Z) (7.14)

with /
zμ k = yk
/
zj = /
zj = μk ,
⎡ δ ⎤
T 1 Δ1 (T δ1 )
⎢ .. ⎥
Λ(T , δ) = ⎣ . ⎦ (7.15)
T δp Δ p (T δp )

with
⎡ δk ⎤
T
⎢ .. ⎥
⎢ . ⎥
⎢ ⎥
Δk (T ) = ⎢
⎢ T 2δk ⎥
⎥ (7.16)
⎢ .. ⎥
⎣ . ⎦
T ηk δk

Moreover, the standard of the observation error is bounded by an exponential whose

decay rate can be chosen arbitrarily large.
Remark 1 The system

/̇
z = A/
z + ϕ(/ /
z, u) + Λ−1 (T , δ)K(y − C Z) (7.17)
7 NLO Based FDI for a Manipulator Robot 175

is also an observer for system (7.12). If a change of variable z = φ(x) has been
x = φ −1 (/
necessary, it is necessary to return to the old database by/ z). By applying
this change of coordinates to the previous system, we obtain the observer in the old
coordinates.
m ( )−1
∂φ(x)
/̇
x(t) = f (/
x (t)) + x (t))ui (t) +
gi (/ Λ−1 K[y(t) − C/
x (t)]
∂x /
x
i=1
(7.18)

Remark 2 The observer is written in the new coordinates as a system copy plus a
non-linear correction. Implementation requires writing the observer into the original
frame.
Therefore the observer becomes:
⎧ ( )
⎪
⎨ .
m ∂φ(x) −1 −1
/̇
x(t) = f (/x (t))+ gi [/x (t)]ui (t)+ Λ K[y(t) − C/ x (t)]
∂x /
⎪
⎩ y = Cx
i=1 x

(7.19)
( )−1
∂φ(x)
The correction term Λ−1 K[y(t) − C/
x (t)] is then explicited as follows:
∂x /
x
⎡ ⎤
1 0 0 0 0
( )−1 ⎢0 0⎥
⎢ 1 0 0 ⎥
∂φ(x) ⎢ ⎥
= ⎢0 0 1 0 0⎥ (7.20)
∂x / ⎢ ⎥
x ⎣0 0 0 1 0⎦
0 0 0 0 1
⎡ ⎤
⎢T −δ1 0 0 ⎥
⎢ 0 0 ⎥
⎢ ⎥
⎢ 0 T −2δ1 0 0 0 ⎥
Λ−1 =⎢ ⎥ (7.21)
⎢ 0 0 T −δ2 0 0 ⎥
⎢ ⎥
⎣ 0 0 0 T −2δ2 0 ⎦
0 0 0 0 T −δ3

The gain T −1 must be selected as 0 < T ≤ T0 < 1. T0 is defined according to

different parameters (η2 , the constant Lipschitz of the function ϕ(z, u) defined by
variable change,. . . ). And the gain K is given by:
176 K. O. Omali et al.

⎡ ⎤
K1
⎢ .. ⎥
⎢ . ⎥
⎢ ⎥
⎢ ⎥
⎢ K2 ⎥
⎢ ⎥
⎢ .. ⎥
⎢ . ⎥
⎢ ⎥
K=⎢ K3 ⎥ (7.22)
⎢ ⎥
⎢ .. ⎥
⎢ . ⎥
⎢ ⎥
⎢ K4 ⎥
⎢ ⎥
⎢ .. ⎥
⎣ . ⎦
K5

The pair (A, C) is observable, and it is an easy matter to assign eigenvalues to

the matrix (A − KC), that has the companion structure:
⎡ ⎤
−K1 1 ... 0
⎢ .. .. . . .. ⎥
⎢ . . .⎥
A − KC = ⎢ . ⎥ (7.23)
⎣−Kn−1 0 . . . 1⎦
−Kn 0 ... 0

If a n-pla λ = (λ1 ,. . . ,λn ) of eigenvalues has to be assigned, the vector K(λ) is the
vector that contains the coefficients of the monic polynomial that has λ as roots. If
the assigned eigenvalues are distinct, matrix (A − KC) can be diagonalized by a
vandermonde matrix:
⎡ n−1 ⎤
λ1 . . . λ1 1
⎢ ⎥
V ≡ V (λ) ⎣ ... . . . ... ... ⎦ (7.24)
λn−1
n . . . λn 1

So that

V (λ)(A − K(λ)C)V (λ)−1 = diag {λ} = Λ (7.25)

Remark 3 Given a set λ of n eigenvalues to be assigned to A-KC, the gain K(λ) is

readily computed through the formula:

K(λ) = −V −1 (λ)[λ1 n . . . λn n ]T (7.26)

which is not difficult to check. It is well known that a vandermonde matrix V (λ)
is singular if and only if two (or more) eigenvalues in the set λ coincide. It is
also well known that the smaller is the minimum difference between eigenvalues
7 NLO Based FDI for a Manipulator Robot 177

in λ, the larger is the norm of V 1 (λ). For reasons that will be made clear in the
following section it is important to choose eigenvalues for matrix (A − K(λ)C)
keeping bounded the norm of the inverse of the vandermonde matrix V (λ). In
Saberi and Sannuti (1990) it is shown that if the n eigenvalues are chosen as
λj = λj (w) = −w j , for j = 1, . . . , n, with w > 0, then

lim ||V −1 (λ(w))|| = 1 (7.27)

w→∞

So, the term of correction of the system is:

⎡ ⎤
T −δ1 K1 (q1 − q/1 )
( ) ⎢T −2δ1 K (q − q/ )⎥
∂φ(x) −1 −1 ⎢ 2 2 2 ⎥
⎢ ⎥
x (t)] = ⎢ T −δ2 K3 (q3 − q/3 ) ⎥
Λ K[y(t) − C/ (7.28)
∂x / ⎢ ⎥
x ⎣T −2δ2 K4 (q4 − q/4 )⎦
T −δ3 K5 (q5 − q/5 )

7.4 High Gain Observer Applied to the ANAT Robot

In the observer based scheme, reconstruct some or all system outputs from
accessible process variables. The fault indicators are generated by forming the
differences between the estimated outputs and the actual outputs. In the absence
of faults, the differences converge to zero as soon as a fault occurs, these differences
are no longer equal to zero (Dixon et al. 2000; Farza et al. 2004; Filaretov et al.
2003).
Using the system model (7.5), a high-gain observer is developed as explained in
Sect. 7.3:
⎧ ( )
⎪
⎨/̇ .m ∂φ(x) −1 −1
x(t) = f (/ x (t)) + i=1 gi (/
x (t))ui (t) + Λ K[y(t) − C/ x (t)]
∂x /
⎪
⎩y = Cx
x

(7.29)

The gain K and T , show the effectiveness of this observer. Thus, the choice of
the gain of high-gain observer is based on a compromise between the speed of
convergence of the observer and insensitivity to measurement noise. The gain K
are chosen according to Remark 3, and T between 0 and 1.
Therefore, the system model and the observer are shown in Fig. 7.7 in order to
clarified the structure of residuals generation.
178 K. O. Omali et al.

Fig. 7.7 Residual generation

7.4.1 Simulation Results

The state estimation error r(t) can be calculated as:

r(t) = y(t) − /
y (t) (7.30)

The residuals are supposed to differ from zero in case of faults r(t) = 0 and to be
zero when there are no faults on the sensors r(t) = 0 (Fig. 7.8).
So the residuals are evaluated as:

r1 = |q1 − /
q1 | (7.31)
r2 = |q2 − /
q2 | (7.32)
r3 = |q3 − /
q3 | (7.33)
r4 = |q4 − /
q4 | (7.34)
r5 = |q5 − /
q5 | (7.35)

Table 7.2 represents the fault signatures matrix for these residuals. Assuming that
simultaneous faults cannot occur, we find that the signatures for each of the failures
are quite different.
We simulate the system during T = 30 s. It is also noted that all faulty signals
are additive.
– a fault d1 is injected on the first joint (q1 ) at the time t = 2 s. Figure 7.9 shows
that the residual r1 is different from zero during the presence of fault, and the
residuals (r2 ; r3 ; r4 ; r5 ) are equal to zero.
7 NLO Based FDI for a Manipulator Robot 179

Fig. 7.8 Observer based

residual generation

Table 7.2 Fault signatures di /ri r1 r2 r3 r4 r5

matrix
d1 1 0 0 0 0
d2 0 1 0 0 0
d3 0 0 1 0 0
d4 0 0 0 1 0
d5 0 0 0 0 1

Fig. 7.9 Residuals evolution of the system with fault d1

– at t = 2 s, Fig. 7.10 shows that the residual r2 is sensitive to this fault, and the
residuals (r1 ; r3 ; r4 ; r5 ) are equal to zero.
– a fault d3 is injected on the joint (q3 ) at the instant t = 2 s. Figure 7.11 shows the
evolution of the different residuals. It can be seen that the residuals (r1 ; r2 ; r4 ; r5 )
are equal to zero and the residuals r3 is sensitive to the fault.
– a fault d4 injected on the fourth articulation (q4 ) at the instant t = 2 s. Figure 7.12
shows that the residu r4 differ from zero during the failure time.
– a fault d5 is injected on the articulation (q5 ) at the time t = 2 s. Figure 7.13 shows
that the residual r5 is sensitive to this fault.
180 K. O. Omali et al.

Fig. 7.10 Residuals evolution of the system with fault d2

Fig. 7.11 Residuals evolution of the system with fault d3

Therefore, we simulate the system at the time of T = 10 s. It is also noted that

all faulty signals are additive.
Those faults are injected in different times of appear and disappear.
The initial condition of the system chosen as q1 = 0.25 rad/s; q2 = 0.75 rad/s;
q3 = 0.5 rad/s; q4 = 1.25 rad/s; q5 = 0.45 rad/s, and for the observer are q1 =
0.75 rad/s; q2 = 1.25 rad/s; q3 = 0.25 rad/s; q4 = 0.25 rad/s; q5 = 0.75 rad/s.
– The first Fig. 7.14 represents the residual r1 when fault d1 is injected on the
articulation q1 between the time ti = 2 s and tf = 4 s. The residual r1 is sensitive
to this fault changed from zero during the failure time.
7 NLO Based FDI for a Manipulator Robot 181

Fig. 7.12 Residuals evolution of the system with fault d4

Fig. 7.13 Residuals evolution of the system with fault d5

– Figure 7.15 represents the residual r2 when a fault d2 is injected on the

articulation q2 between the time ti = 4 s and tf = 6 s, presents that the residual
r2 is sensitive to this fault changed from zero during the failure time.
– Figure 7.16 represents the residual r3 when a fault d3 is injected on the
articulation q3 between the time ti = 2 s and tf = 3 s. The residual r3 is sensitive
to this fault changed from zero during the failure time.
– Figure 7.17 represents the residual r4 when a fault d4 is injected on the
articulation q4 between the time ti = 5 s and tf = 6 s. The residual r4 is sensitive
to this fault changed from zero during the failure time.
182 K. O. Omali et al.

Fig. 7.14 Residuals evolution of the system with fault d1

Fig. 7.15 Residuals evolution of the system with fault d2

– Figure 7.18 represents the residual r5 when a fault d5 is injected on the

articulation q5 between the time ti = 4 s and tf = 8 s. The residual r5 is sensitive
to this fault changed from zero during the failure time.
A comparison of the results obtained by the first part of simulation and the second
shows that: the first gives a good results, the second part show that when the initials
conditions are differents to zeros gives also all residuals near of zero.
So, we can say that the proposed methods can give a good results and it can detect
and isolate the sensor faults in a robot manipulator.
7 NLO Based FDI for a Manipulator Robot 183

Fig. 7.16 Residuals evolution of the system with fault d3

Fig. 7.17 Residuals evolution of the system with fault d4

7.5 Conclusion

In this chapter, an approach for fault detection and isolation based on high-gain
observer, for a class of affine nonlinear systems has been proposed. The approach
is applied to a robot manipulator. Through the simulation results, it is shown that
all sensors faults defined in specifications are detected and isolated. First, high-
gain observer approach has been presented. Then, application for this methods in
184 K. O. Omali et al.

Fig. 7.18 Residuals evolution of the system with fault d5

robot manipulator “ANAT” has been performed. Finally, simulation results have
been shown and proven the effectiveness, performances and the robustness of the
approach.

References

Busawon, K., Farza, M., & Hammouri, H. (1998). Observer design for a special class of nonlinear
systems. International Journal of Control, 71, 405–418.
Craig, J. J. (1989). Introduction to robotics: Mechanics and control (2nd ed.). Boston: Addison-
Wesley Longman Publishing, Inc.
Dixon, E., Warren, W., Ian, D., Dawson, D. M., & Hartranft, J. P. (2000). Fault detection
for robot manipulators with parametric uncertainty: A prediction-error-based approach. IEEE
Transactions on Robotics and Automation, 16(6), 689–699.
Esfandiari, F., & Khalil, H. (1987). Observer-based design of uncertain systems: Recovering state
feedback robustness under matching conditions. In Allerton Conference, Monticello (pp. 97–
106).
Farza, M., Hammouri, H., Jallut, C., & Lieto, J. (1999). State observation of a nonlinear system:
Application to (bio) chemical processes. AICHE Journal, 45, 93–106.
Farza, M., M’Saad, M., & Rossignol, L. (2004). Observer design for a class of MIMO nonlinear
systems. Automatica, 40, 135–143.
Farza, M., M’saad, M., & Sekher, M. (2005). A set of observers for a class of nonlinear systems.
In International Federation of Automation Control-IFAC’05, Prague, 37.
Filaretov, V., Vukobratovic, M., & Zhirabok, A. (1999). Observer-based fault diagnosis in
manipulation robots. Mechatronics, 9, 929–939.
Filaretov, V., Vukobratovic, M., & Zhirabok, A. (2003). Parity relation approach to fault diagnosis
in manipulation robots. Mechatronics, 13, 141–152.
Frank, P. (1990). Fault diagnosis in dynamic systems using analytical and knowledge-based
redundancy? A survey and some new results. Automatica, 26, 459–474.
Frank, P. (1993). Advances in observer-based fault diagnosis. In TOOLDIAG’93, International
Conference on Fault Diagnosis, Toulouse.
7 NLO Based FDI for a Manipulator Robot 185

Gauthier, J., & Kupka, I. (2001). Observability and observers for nonlinear systems. SIAM Journal
Control Optimisation, 32, 181–188.
Gauthier, J. P., Hammouri, H., & Othman, S. (1992). A simple observer for nonlinear systems –
application to bioreactors. IEEE Transactions on Automatic Control, 37, 875–880.
Gauthier, J. P., Hammouri, H., & Othman, S. (2001). A simple observer for nonlinear systems –
application to bioreactors. IEEE Transactions on Automatic Control, 37, 875–880.
Gertler, J. (1993). Residual generation in model-based fault diagnosis. Control Theory Advanced
Technology, 9, 259–285.
Gertler, J., & Monajemy, R. (1995). The state of the art. Automatica, 31, 627–635.
Guernez, C., Cassar, J., & Staroswiecki, M. (1997). Process fault diagnosis based on modeling and
estimation methods-a survey automatica. In IFAC Symposium SAFEPROCESS’97, Kingston
upon Hull.
Hammouri, H., & Farza, M. (2003). Nonlinear observers for locally uniformly observable systems.
ESAIM, 9, 353–370.
Hammouri, H., Busawon, K., Yahoui, A., & Grellet, G. (2001). A nonlinear observer for induction
motors. European Physical Journal-Applied Physics, 15, 181–188.
Hou, M., Busawon, K., & Saif, M. (2000). Observer design for a class of MIMO nonlinear systems.
IEEE Transactions on Automatic Control, 45, 1350–1355.
Isermann, R. (1984). Process fault diagnosis based on modeling and estimation methods-a survey
automatica. International Federation of Automatic Control, 20, 387–404.
Jollie, I. T. (2016). Simultaneous fault diagnosis for robot manipulators with actuator and sensor
faults. Information Sciences, 366, 12–30.
Khalil, H., & Praly, L. (2014). High-gain observers in nonlinear feedback control. International
Journal of Robust and Nonlinear Control, 24, 993–1015.
Khalil, H., & Saberi, A. (2007). Adaptive stabilization of a class on nonlinear systems using high-
gain feedback. IEEE Transactions on Automatic Control, 32, 1031–1035.
Koubaa, Y., Farza, M., & M’saad, M. (2004). Obsevateur adaptatif pour une classe des systèmes
non linéaires. In 5th Conférence Internationale des Sciences et Techniques de l’Automatique,
STA’04.
Krishnaswami, V., & Rissoni, G. (1994). Nonlinear parity equation residual generation for fault
detection and isolation. In IFAC Symposium SAFEPROCESS’94, Espoo (Vol. 1, pp. 317–322).
Ma, H. J., & Yang, G. H. (2016). Simultaneous fault diagnosis for robot manipulators with actuator
and sensor faults. Information Sciences, 366, 12–30.
Mironovsky, L. (1989). Functional diagnosis of nonlinear discrete-time processes. Autom Remote
Control, 6, 150–157.
Nadri, M. (2001). Observation et commande des systèmes non linéaires et application aux
bioprocédés. Thèse de doctorat, Université Claude Bernard Lyon-1.
Patton, R. (1994). The state of the art. In IFAC Symposium SAFEPROCESS’94, Espoo (Vol. 1,
pp. 1–24).
Saberi, A., & Sannuti, P. (1990). Observer design for loop transfer recovery and for uncertain
dynamical systems. IEEE Transactions on Automatic Control, 35, 878–897.
Schneider, H., & Frank, P. M. (1996). Observed- based supervision and fault detection in robots
using nonlinear and fuzzy logic residual evaluation. IEEE Transactions on Control Systems
Technology, 4(3), 274–282.
Shim, H., Son, Y., & Seo, J. (2001). Semi-global observer for multi-output nonlinear systems.
Systems and Control Letters, 42, 233–244.
Shumsky, A. (1998). Parity relation method and its application to fault detection in nonlinear
dynamic systems. Automation and Remote Control, 9, 155–165.
Slotine, J. J., & Weiping, L. (1991). Applied nonlinear control. Englewood Cliffs: Printice-Hall
International.
Tornambe, A. (1988). Use of asymptotic observers having high gains in the state and parameter
estimation. In 27th IEEE Conference on Decision and Control, Austin (pp. 1791–1794).
Tornambe, A. (1992). High-gain observers for non-linear systems. International Journal of
Systems Science, 23, 1475–1489.
Chapter 8
Crone Controller Design for a Robot Arm

Ahmed Abid, Rim Jallouli–Khlif, Nabil Derbel, and Pierre Melchior

Abstract The CRONE strategy (French acronym of Commande Robuste d’Ordre

Non Entier) is assumed to give rise to a great number of applications in different
industrial fields. This paper introduces the design of a second generation controller
intended to control in position a two degrees of freedom robot arm. Great perfor-
mances and robust behaviour are obtained and validated especially by a comparative
study.

Keywords CRONE control · Second generation · Robustness · Robot arm

8.1 Introduction

In the two last decades in the field of automatic control, fractional systems have
been of growing interest in analysis, system identification and control domains.
Concerning this last domain, several authors have suggested to introduce fractional
controllers (Machado 1997; Manabe 1961; Petras and Dorcak 1999; Podlubny
1999). This approach leads indeed to a better flexibility in the structure of the
controller. In CRONE control, fractional differentiation is also used in the design
procedure of the controller to ensure the robustness of the feedback system.
Although robustness is a very wide notion, it always means insensitivity or very
nearly so. In the fractional CRONE approach (Oustaloup 1991a,b; Oustaloup and
Mathieu 1999) the severest robustness of the stability degree (stability margins)

A. Abid () · R. Jallouli–Khlif · N. Derbel

Control and Energy Management Lab (CEM-Lab LR11ES47), National Engineering School of
Sfax, Sfax, Tunisia
e-mail: [email protected]; [email protected]
P. Melchior
Integration from Materials to Systems Lab (IMS-UMR 5218 CNRS), Bordeaux INP, University
of Bordeaux, Bordeaux, France
e-mail: [email protected]

© Springer Nature Singapore Pte Ltd. 2019 187

N. Derbel et al. (eds.), New Developments and Advances
in Robot Control, Studies in Systems, Decision and Control 175,
https://doi.org/10.1007/978-981-13-2212-9_8
188 A. Abid et al.

is sought, in order to maintain the frequency or time dynamic performance which

measures this degree (performance robustness) (Oustaloup et al. 2013). Hence, the
CRONE controller is the first fractional order controller developed. It was launched
in 1975, and it was introduced using three generations (Abid et al. 2015; Jallouli-
Khlif et al. 2010; Moreau et al. 2018).
CRONE control methodology treats with different domains as automobile
(Nouillant et al. 2002; Oustaloup et al. 1996), aerospace, thermal applications
(Moreau et al. 2018), industrial systems (Oustaloup et al. 2013), robotics (Delavari
et al. 2013). . .
This work is dealing with the design of a CRONE controller for a robotic
application, notably the position control of a two degrees of freedom robot arm.
As a first step, the work deals with a servo MX − 28AT , used as an actuator
in a humanoid research robot Darwin OP2 (Maximo et al. 2017). The controller
is designed using the CRONE toolbox on the base of an experimentally validated
plant model (Maximo et al. 2017). Simulation results prove the efficiency of such
second generation controller. Moreover, a comparison study with a PID controller
arguments the advantages of the fractional approach especially in settling time and
robustness against parametric uncertainties. In a second step, the designed fractional
controller is introduced to control an XY plane robot arm. Simulations run upon
nominal case and two modified ones validate the approach notably by comparing
with the classical PID approach.
The remainder of the paper is divided as follows. An overview on the second
CRONE generation is given in Sect. 8.2. The robot servo model subject to control
is briefly presented in Sect. 8.3. Section 8.4 explains the control strategy based on
the CRONE controller and gives simulation results. A comparison study versus
a PID controller is achieved in Sect. 8.5. Section 8.6 deals with the use of the
designed fractional controller to command in position a robot arm with two degrees
of freedom. Efficiency and robustness of the approach are mentioned. Finally, some
concluding remarks are drawn in the last Section.

8.2 Second CRONE Generation

For control effort-level problems, it can be difficult to have a crossover frequency

ωcg in a frequency interval where the plant phase is constant. For the second
generation CRONE approach, as the transfer function of the open loop system is
synthesized a posteriori, it is defined as the transfer function of a fractional order
integrator:
ω
cg n
β(s) = , (8.1)
s
where ωcg is the frequency for which the uncertainties do not lead to any phase
variation, n ∈ R and n ∈ [1, 2]. The closed loop transfer function T (s) is defined as
follows:
8 Crone Controller Design for a Robot Arm 189

1
T (s) =  n, (8.2)
1+ s
ωcg

Around the crossover frequency ωcg , the Black N ichols plot of the open-loop
transfer function β(s) is a vertical asymptote with a constant phase. This asymptote
allows having (Moreau et al. 2018):
• a robust phase margin Mφ equal to (2 − n) × π/2;
• a robust resonance factor;
• a robust gain module;
• a robust damping ratio of the closed loop system;
In order to manage accuracy, robustness and control effort problems, Eq. (8.1) of
the controller is truncated in frequency, and a low pass filter as well as an integrator
are added. Therefore, the new description of the open-loop transfer function will be
written as follows:
 n  n  
1 + ωsl l 1 + ωsh s −nh
β(s) = β0 1+ (8.3)
s
ωl 1 + ωsl ωh

where ωl and ωh are respectively the low and the high transitional frequencies, n is
the fractional order around the frequency ωcg , nl and nh define the system behavior
at low and high frequencies, and β0 is a constant that guarantees a crossover
frequency ωcg . It is expressed in Eq. (8.4) (Oustaloup 1991a):

  nl 0   1 n−nl 0   1 nh −n
ωcg ωcg 2 2 ωcg 2 2
β0 = 1+ 1+ (8.4)
ωl ωl ωh

The frequency fractional order behavior of β(s) is defined over an interval around
the frequency ωcg of a depth of r. Moreover, for the nominal plant, previous studies
have shown that it is practical to satisfy such frequency constraints:
 ωcgnom
ωl = √
10 r √ (8.5)
ωh = 10ωcgnom r

Once the open-loop transfer function is calculated, the CRONE controller

transfer function is defined as the ratio of the open-loop transfer function β(s) over
the nominal plant transfer function H (s):

β(s)
CF (s) = (8.6)
H (s)

The rational form CR (s) is deduced using the recursive poles and zeros distribu-
tion of the function. For more details, please refer to Oustaloup (1991a).
190 A. Abid et al.

8.3 Robot Servo Model

A model for the behavior of the Dynamixel MX − 28AT position servo is

developed in Maximo et al. (2017), including phenomena such as viscous friction
and saturation. Such a model allows the correct simulation of the servo behavior
and the adequate design of the controller. For that purpose, constants for both servo
and DC motor are directly obtained from corresponding data sheets or from some
standard tests. The servo is built using an armature controlled Maxon DC motor
(Maximo et al. 2017). The equivalent electric circuit is depicted in Fig. 8.1, where
u is the input voltage at the terminals, e is the back electromotive force, i is the
current, R is the resistance and L is the inductance.
Further, the motor is characterised in its mechanical part by a torque constant Kt
and a speed constant Kω . The rotation angle of the motor shaft is θm and its moment
of inertia is Jm . Alike, the load axis angle of rotation is θl and its moment of inertia
is Jl .
Usually, models of the servo motor are nonlinear, that means at least one of the
states (i: current, θ : motor position) is an argument of a nonlinear function. In order
to present such a model as a transfer function, it has to be linearized. However for
the servo motor model, nonlinearities are small, so they can be neglected.
The open loop transmittance of the Dynamixel MX − 28AT is expressed by the
following transfer Eq. (8.7):
θ (s)
H (s) = (8.7)
U (s)
Using plant parameters, the transfer function is expressed as follows:

H (s) =
NηKt
* 2
+ (8.8)
s L(Jl +Jm N 2 η)s 2 +[R(Jl +Jm N 2 η)+Lbm N 2 η]s+bm N 2 ηR+ KtKNω η

Fig. 8.1 Electrical

architecture
8 Crone Controller Design for a Robot Arm 191

Table 8.1 presents the values of the motor parameters supplied by the manufac-
turer.

8.4 Control Strategy

The manufacturer proposed a P I D controller for the MX − 28AT . Unfortunately,

the servomotor electrical parameters L and R (Table 8.1) may vary with the
temperature, the saturation of magnetic circuits and frequency variations. To
encounter parameter uncertainties effects, this work proposes a robust fractional
order controller belonging to the second CRONE generation. The proposed control
schema is given in Fig. 8.2.
We assume that the variation of the inductance L can be neglected because the
angular velocity of the servo motor and thereby the frequencies are small. So, three
parametric states are considered:
• R = Rnom = 8.3 
• R = R1 = 150% of Rnom
• R = R2 = 250% of Rnom
Hence, the corresponding identified models are given by:

1.298 10−2 s + 1.326 10−2

H (s) =
1.123 10−1 s 3 + 1.371s 2 + 1.284s

Table 8.1 Parameters of the Parameters Notations Values

MX − 28AT
Resistance R 8.3 
Inductance L 2.03 ×10−3 H
Gear ratio N 193
Gear efficiency η 0.836
Speed constant Kω 93.1 rad/V
Torque constant Kt 0.0107 Nm/A
Inertia Jm 8.68 ×10−8 kg m2
Friction bm 8.87 ×10−8 Nms

Fig. 8.2 Control schema

192 A. Abid et al.

0
Magnitude (dB)

-100

R=R nom R=1.5 R nom R=2.5 R nom

-200
-90
Phase (deg)

-135

-180
10 -1 10 0 10 1
Frequency (rd/s)10 2 10 3 10 4

Fig. 8.3 Bode shapes of the plant transfer functions

8.655 10−2 s + 8.843 10−2

H1 (s) = (8.9)
1.123s 3 + 9.905s 2 + 8.948s
5.193 10−2 s + 5.306 10−2
H2 (s) =
1.123s 3 + 6.861s 2 + 5.838s

where H (s) is the nominal model corresponding to R = Rnom , H1 (s) is the model
corresponding to R1 = 150% of Rnom and H2 (s) corresponds to R2 = 250% of
Rnom .
Figure 8.3 shows the Bode diagrams for the three plants. It is noticed that plant
uncertainties are reduced to variations in the gain. Since, the use of the second
generation controller is justified.

8.4.1 Controller Design

The purpose is to command the servo in position, so a time response of 0.07 s is

adequate. The cross over frequency is fixed to 32 rad/s (Maximo et al. 2017). The
synthesis of the second generation controller is realized by means of the module
“CRONE Control System Design” from the Matlab Toolbox “Crone Toolbox”
(Lanusse et al. 2013).
Taking into consideration the specifications of the user guide, different variables
can be defined. Thus:
• nl = 2, in order to ensure zero steady-state error;
• nh = 2, in order to limit the input sensitivity;
• Mφ = 90◦ (phase margin);
• ωcgnom = 32 rad/s.
8 Crone Controller Design for a Robot Arm 193

Fig. 8.4 Rational CRONE

{140341248740.61},[1 10119]} +
controller
LTI1

1 tf{[0 28666],[1 0]} + 1

In LTI2 Out

+
tf{[17929331.2808],[1]}

LTI3

Plant output y(t)

1
1

0.5 0.5

0.02 0.04 0.06 0.08 0.1

0
0 H1 (s)
0.2 H2 (s)
0.4 0.6
H(s) 0.8
Reference 1

Control effort u(t)

2000
2000
1000
1000 0
0.02 0.04 0.06 0.08 0.1 0.12
0
0 0.2 0.4 0.6 0.8 1
Time(s)

Fig. 8.5 Plant output and control effort versus time

For practical reasons, the second generation CRONE controller is approximated

by a rational transmittance, obtained by fitting its frequency response. By export
on Simulink, and in order to counter numerical problems, the synthesized CRONE
controller is decomposed into parallel transmittances according to Fig. 8.4.

8.4.2 Simulation Results

By introducing the rational CRONE controller in the open loop plant, the following
plots resume the outcome (Fig. 8.5).
In fact, for the closed loop step responses, all three outputs have the same behav-
ior respecting the closed loop order: no overshoot. So this proves the robustness of
the fractional controller.
194 A. Abid et al.

8.5 Comparative Study

The manufacturer proposed a PID controller for the servo, so the input voltage is
given by:
 
Ki !
U (s) = Kp + Kd s + θref (s) − θ (s) (8.10)
s

where Kp , Kd and Ki are the controller parameters, θref (s) is the reference input
angle and θ (s) is the actual output angle. Gains of the controller used by the
manufacturer are given in the datasheet as follows (Maximo et al. 2017):
⎧
⎨ Kp = 4
K =0 (8.11)
⎩ d
Ki = 0

Plots of Fig. 8.6 show the three plants outputs for reference input changes of
±200 counts with both PID and CRONE controllers. According to simulation
results illustrated in Fig. 8.6, the CRONE controller proves to be robust against
parameter variations especially for the extreme case. Further, according to results
summarized in Table 8.2, the fractional controller is faster and permits a gain of
more than 25% in settling time comparing to the traditional controller. No overshoot
is detected in all cases even when the electrical parameter is in its maximum
uncertainty range, which is not the case for the PID.

8.6 Robust Control of a Robot Arm

The crone controller designed in Sect. 8.4 is introduced to control two servo-motors
coupled to design a robot arm used to move on an XY plane (Fig. 8.7). The arm has
two degrees of freedom and is characterized by the parameters given in Table 8.3.
The control strategy is presented in Fig. 8.8.
The controlled arm is modeled by the matrix form expressed in Eq. (8.12).

M(θ )θ̇ + C(θ, θ̇ )θ̇ + G(θ ) = τ (8.12)

where
• θ = [θ1 θ2 ]T ∈ R2 is the position vector,
• θ̇ = [θ˙1 θ˙2 ]T ∈ R2 is the speed vector,
• θ̈ = [θ¨1 θ¨2 ]T ∈ R2 is the acceleration vector,
• M(θ ) ∈ R2×2 is the inertia matrix, which is symmetric, uniformly bounded and
positive definite
• C(θ, θ̇ ) ∈ R2×2 , with C(θ, θ̇ )θ̇ represents the Coriolis and centrifugal forces,
8 Crone Controller Design for a Robot Arm 195

H(s)
200
PID
150 CRONE
position

200
100
180
50
1 1.1 (a)
0
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
time (s)
H1(s)

200
PID
150 CRONE
200
Position

100 180
160
50 1.05 1.1 1.15
(b)
0
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
Time (s)
H2(s)

200
PID CRONE

150
200
Position

20
100 190
10
180
50 0 1.05 1.1 1.15 1.2
0.05 0.1 0.15 0.2 0.25
(c)
0
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
Time (s)

Fig. 8.6 Servo position response. (a): R = Rnom nominal case. (b): R = 1.5Rnom . (c): R =
2.5Rnom .

Table 8.2 Simulation results of both controllers

CRONE PID
Overshoot (%) Settling time t5% [s] Overshoot (%) Settling time t5% [s]
H (s) 0 0.061 0 0.081
H1 (s) 0 0.062 0 0.086
H2 (s) 0 0.066 2.56 0.085
196 A. Abid et al.

Fig. 8.7 Two degrees of y

freedom robot arm

θ2 I2
d2

I1 M2

d1
θ1

M1

0
x

Table 8.3 Parameters of the Parameters Notations Values

arm
Arm length l1 , l2 10 cm
Gravity center distance d1 , d2 5 cm
Payload weight m 1 , m2 100 g
Inertia J1 , J2 0.27×10−2 kg m2

Fig. 8.8 Two degrees of freedom robot arm control strategy

• G(θ ) ∈ R2 denotes the gravity vector, is supposed null in our case study,
• τ ∈ R2 is the vector of actuator torques.
The inertia matrix coefficients are expressed as follows:

M11 = J1 + J2 + m2 l12 + d22 + 2l1 d2 cos θ2 + m1 d12

M12 = M21 = J2 + m2 d22 + l1 d2 cos θ2 (8.13)

M22 = J2 + m2 d22
8 Crone Controller Design for a Robot Arm 197

Table 8.4 Arm simulation results for position angle θ1

CRONE PID
Overshoot (%) Settling time t5% [s] Overshoot (%) Settling time t5% [s]
Rnom case 0 0.143 2.3 0.148
R1 case 1.7 0.151 5.8 0.217
R2 case 6.4 0.207 12 0.277

Table 8.5 Arm simulation results for position angle θ2

CRONE PID
Overshoot (%) Settling time t5% [s] Overshoot (%) Settling time t5% [s]
Rnom case 0 0.136 0.8 0.148
R1 case 0 0.142 2.9 0.151
R2 case 0 0.153 6.6 0.253

C(θ ) is expressed as:

 
−θ˙2 − θ˙1 − θ˙2
C(θ, θ̇ ) = m2 l1 d2 sin θ2 (8.14)
θ˙1 0

and G(θ ) is given by:

( )
(m1 d1 + m2 l1 )g cos q1 + m2 gd2 cos(q1 + q2 )
G(q) = (8.15)
m2 d2 g cos(q1 + q2 )

Simulations are run for the nominal case, and for two cases of uncertainty (R =
R1 = 150% of Rnom and R = R2 = 250% of Rnom ).
Arm responses controlled by the fractional controller are given in Fig. 8.9. As
in Fig. 8.5, to prove the efficiency and notably the robustness of the fractional
controller, PID-controller results are drawn simultaneously in Fig. 8.10.
Tables 8.4 and 8.5 compare accurately both controllers performances in position
control.
According to Figs. 8.9a and 8.10a (the nominal case), both controllers give good
results: both position angles θ1 and θ2 go towards desired positions perfectly without
any overshoot for the fractional case, and with likely small overshoots with PID
controller. Response times are likely equivalent for both controllers.
However, if the resistance value is changed, the fractional controller does
extremely well notably in cancelling the overshoot for the final desired position as
shown in Fig. 8.9b, c. Comparing Figs. 8.9c and 8.10c, the CRONE controller proves
its robustness even in the extreme case of uncertainty. Even if PID performances are
acceptable (a maximum overshoot of 12%), the CRONE controller makes proof of
accurate path tracking and robustness.
198 A. Abid et al.

600

500 θref
θ
400 2
θ
Position

300 1

500
200
450
100
400
0 1 1.1 1.2 1.3
(a)
−100
1 1.5 2 2.5 3 3.5 4 4.5 5 5.5 6
Time(s)

600
θref
400 θ
2
θ
Position

540 1
520
200 500
480
460
440
0 420
1 1.1 1.2 1.3
(b)
−200
1 1.5 2 2.5 3 3.5 4 4.5 5 5.5 6
Time(s)

600
θ
ref
400 θ
2
θ
Position

1
550
200
500

450
0
1 1.2
(c)
−200
1 1.5 2 2.5 3 3.5 4 4.5 5 5.5 6
Time(s)

Fig. 8.9 Arm responses with the CRONE controller: position angles θ1 and θ2 . (a): R = Rnom
nominal case. (b): R = 1.5Rnom . (c): R = 2.5Rnom

8.7 Conclusion

In this work, a fractional order controller is employed in a robotic application. In

fact, a second generation controller, based on CRONE approach, is designed first
for the position control of a servo MX − 28AT , used as actuator in a humanoid
research robot Darwin OP2, then to control a robot arm with two degrees of
8 Crone Controller Design for a Robot Arm 199

600
θref
400 θ2
θ1
Position

200 550
500
450
0
400
1 1.1 1.2 1.3 (a)
−200
1 1.5 2 2.5 3 3.5 4 4.5 5 5.5 6
Time(s)

600

500 θref

400 θ
2
θ1
Position

300 550
500
200
450
100 400
1 1.2 1.4
0
(b)
−100
1 1.5 2 2.5 3 3.5 4 4.5 5 5.5 6
Time(s)

600
θref
400 θ
2
θ1
Position

200 550
500
450
0
400
1 1.2 1.4
(c)
−200
1 1.5 2 2.5 3 3.5 4 4.5 5 5.5 6
Time(s)

Fig. 8.10 Arm responses with the PID controller: position angles θ1 and θ2 . (a): R = Rnom
nominal case. (b): R = 1.5Rnom . (c): R = 2.5Rnom

freedom. Simulation results validate the proposed solution in both application cases.
Further, a comparison study with the PID controller, traditionally used, proves better
temporal performances and distinguished advantage in term of robustness for the
fractional controller.
200 A. Abid et al.

References

Abid, A., Jallouli-Khlif, R., Melchior, P., & Derbel, N. (2015). Negative input shapers for explicit
fractional derivative systems. In IEEE International Conference on Modelling, Identification
and Control, Sousse.
Delavari, H., Lanusse, P., & Sabatier, J. (2013). Fractional order controller design for a flexible
link manipulator robot. Asian Journal of Control, 15(3), 783–795.
Jallouli-Khlif, R., Melchior, P., Levron, F., Derbel, N., & Oustaloup, A. (2010). Analytical impulse
response of third generation crone control. In D. Baleanu, Z. B. Guvenc, & J. A. Tenreiro
Machado (Eds.), New trends in nanotechnology and fractional calculus applications (Fractional
Control Systems, pp. 343–356). Dordrecht: Springer.
Lanusse, P., Malti, R., & Melchior, P. (2013). Crone control-system design toolbox for the control
engineering community. Philosophical Transactions of the Royal Society A: Mathematical,
Physical and Engineering Sciences, 371, 20120149.
Machado, J. A. T. (1997). Analysis and design of fractional-order digital control systems. Journal
of Systems Analysis, Modelling, Simulation, 27, 107–122.
Manabe, S. (1961). The non integer integral and its application to control systems. Electrotechnical
Journal of Japan, 6(3–4), 83–87.
Maximo, M., Ribeiro, C., & Afonso, R. (2017). Modeling of a position servo used in robotics
applications. XIII Simposoi Brasileiro de Automaçao Intelligente.
Moreau, X., Daou, R., & Christophy, F. (2018). Comparison between the second and third
generations of the crone controller: Application to a thermal diffusive interface medium.
Fractal and fractional, 2(1), 5.
Nouillant, C., Assadian, F., Moreau, X., & Oustaloup, A. (2002). Feedforward and crone feedback
control strategies for automobile ABS. International Journal of Vehicle Mechanics and
Mobility, 38(4), 293–315.
Oustaloup, A. (1991a). La commande CRONE. Paris: Hermes Editions.
Oustaloup, A. (1991b). The CRONE control. European Control Conference ECC’91, Grenoble.
Oustaloup, A., & Mathieu, B. (1999). La commande CRONE: du scalaire au multivariable. Paris:
Hermes Editions.
Oustaloup, A., Moreauand, X., & Nouillant, M. (1996). The crone suspension. Control
Engineering Practice, Journal of the International Federation of Automatic Control, 4(8),
1101–1108.
Oustaloup, A., Lanusse, P., Sabatier, J., & Melchior, P. (2013). Crone control: Principles,
extensions and applications. Journal of Applied Nonlinear Dynamics (JAND), 2(3), 207–223.
Petras, I., & Dorcak, L. (1999). The frequency method for stability investigation of fractional
control systems. Stability and Control: Theory and Applications, 2(1–2), 75–85. www.cer.co.
za/sacta/.
Podlubny, I. (1999). Fractional-order systems and pid-controller. IEEE Transactions on Automatic
Control, 44(1), 208–214.
Chapter 9
Cartesian Sliding Mode Control of an
Upper Extremity Exoskeleton Robot
for Rehabilitation

Brahim Brahmi, Maarouf Saad, Cristobal Ochoa-Luna,

Mohammad H. Rahman, and Abdelkrim Brahmi

Abstract Rehabilitation robots play an important role in rehabilitation treatment.

Unlike conventional rehabilitation approach, the rehabilitation robotics provides an
intensive rehabilitation motion with different modes (passive, active and active-
assisted) based on the ability of the exoskeleton robot to perform assistive motion
for a long period. However, this technology is still an emerging field. In this chapter,
we present a Cartesian adaptive control based on a robust proportional sliding
mode combined with time delay estimation for controlling a redundant exoskeleton
robot called ETS-MARSE subject to uncertain nonlinear dynamics and external
forces. The main objective of this research is to allow the exoskeleton robot to
perform both rehabilitation modes, passive and active assistive motions with real
subjects. The stability of the closed loop system is solved systematically, ensuring
asymptotic convergence of the output tracking errors. Experimental results confirm
the efficiency of the proposed control to provide an excellent performance despite
the presence of dynamic uncertainties and external disturbances.

Keywords Rehabilitation robots · Adaptive control · Time delay estimation ·

Active and passive motion

B. Brahmi () · M. Saad · A. Brahmi

Electrical Engineering Department, École de Technologie Supérieure, Montreal, QC, Canada
e-mail: [email protected]; [email protected];
[email protected]
C. Ochoa-Luna
School of Physical & Occupational Therapy, Centre for Interdisciplinary Research in
Rehabilitation of Greater Montreal, McGill University, Montreal, QC, Canada
e-mail: [email protected]
M. H. Rahman
Mechanical Engineering Department, University of Wisconsin-Milwaukee, Milwaukee, WI, USA
e-mail: [email protected]

© Springer Nature Singapore Pte Ltd. 2019 201

N. Derbel et al. (eds.), New Developments and Advances
in Robot Control, Studies in Systems, Decision and Control 175,
https://doi.org/10.1007/978-981-13-2212-9_9
202 B. Brahmi et al.

9.1 Introduction

Habitually, survivors of strokes, trauma and/or neurological disorders suffer paraly-

sis and loss of physical strength, often on one side of the body, such as the superior
extremity (De Morand 2014). The disability of the upper-limb can create difficulties
conducting daily live activities such as eating, dressing and cleanliness tasks.Such
disabilities can have a large impaction the victim’s behavior (Xie 2016). Currently,
effective treatment that helps stroke survivals regain their missing functional
capacity, acquire new skills, and improve their quality of life is a rehabilitation
treatment (Xie 2016). Physical therapy is a set of medical exercises performed by a
physiotherapist in order to increase the range of movement, functional capacity and
quality of subject’s life (De Morand 2014). Latterly, a new method of rehabilitation
has emerged. This approach is named robotics rehabilitation that has attracted a
lot of attention from the scientific community since robots are able to supplement
the treatments provided by conventional physical therapy.The importance of the
rehabilitation robots lies in their ability to provide intensive physiotherapy over
a long period of time and not be influenced by factors such as the therapist’s
unavailability or fatigue (Xie 2016).
Rehabilitation robots are new devices designed to overcome the limitations of
conventional physiotherapy approaches and to create new methods of rehabilitation
treatment (Xie 2016). Typically, the development of this type of robots is conducted
in order to adjust or to be identical to the human arm configuration. To rehabilitate
patients with upper-limb impairment, these exoskeletons habitually are worn on
the lateral side of the patients’ upper-limb. The degree of freedom (DOFs) of the
exoskeleton makes it a capable to reach several arm configurations in its workspace.
When the degrees of freedom of the robot are increased, it is impossible to obtain
the dynamic model of the robot accurately due to the complex mechanical structure
and parameters such as nonlinear friction forces, backlash and the complexity of the
actuators of the robot (Brahim et al. 2016a,b). Furthermore, the dynamic parameters
change due to different physiological conditions of the subjects, such as an external
disturbance caused by subject’s muscular activity (Brahim et al. 2016a,b). That
means the manipulator can be subjected to both parametric uncertainties and
unknown nonlinear functions. The uncertain nonlinear dynamic model and external
forces can turn into an unknown function that can affect the performance of the
exoskeleton In this case, we need a solid adaptive control to better achieve the
therapeutic activity by estimating the dynamic model uncertainties and the external
forces that negatively affect the overall system performance.
In fact, control of uncertain nonlinear dynamics is one of the new topics of the
nonlinear control area. Many control algorithms have been developed to overcome
the effect of these uncertainties which can be characterized into two categories.
The first one is adaptive control (Slotine and Li 1991), e.g., feedback linearization
(Park and Cho 2007), backstepping control (Shieh and Hsu 2008) and H∞ control
(Yim and Park 1999). However, the above controllers assume that the dynamic
model of the robot is linear in a set of physical parameters and derive a control
9 SMC of an Upper Extremity Exoskeleton Robot for Rehabilitation 203

law capable of ensuring the stability of the linear system. Actually, the manipulator
is highly nonlinear, which means the integration of this law may affect the stability
of the system in the presence of even small disturbances. The second one is the
deterministic robust control (Khalil and Grizzle 1996; Slotine and Li 1991). One of
the successful robust methods that have been used to control the uncertain dynamics
is the nonlinear sliding mode control based on high gains switching controller,
which forces the system to converge to a selected sliding surface. The robustness
of the controller is ensured when the switching gain is larger than the bounded
uncertainties (Fridman 1999; Young et al. 1999). To compensate for the chattering
problem caused by the high switching gains that provoke damage in the actuators
of the robot (Rahman et al. 2013), there are numerous approaches combined with
sliding mode control such fuzzy logic (Chen et al. 2009), neural network or/and
both (Wang et al. 2009). However, these approaches need heavy computations
which make the implementation very difficult. Time-delay control estimation (TDE)
(Brahmi et al. 2018a; Youcef-Toumi and Ito 1990) is a strong approach used to limit
the effect of the uncertainties. The TDE technique can be combined with sliding
mode to control the uncertainties and external forces of the robot (Baek et al. 2016).
The time delay estimation technique basically employs time-delayed information
from the previous state response of the system and the previous control input to
estimate the dynamic uncertainties of the robot. However, the TDE method suffers
from time delay error (TDR). This TDR can affect the robustness and the accuracy
of the robot performance.
It is interesting to remark that in most rehabilitation applications, the therapeutic
tasks are expressed in Cartesian space (Xie 2016). In some cases, the dynamics
of the robot is modeled in Cartesian space (Craig 2005). Then, a Cartesian sliding
surface is selected to control the uncertain nonlinear dynamics of the robot (Slotine
and Li 1991). On the other hand, the dynamics of the manipulator is developed
in joint space (Craig 2005). Then, a sliding joint surface is selected to achieve
the control. In this case, the transformation between the joint space and Cartesian
space is done by the Jacobian matrix (Craig 2005). Unfortunately, it is impossible
to determine the dynamic and kinematic parameters of the robot exactly, due to
the uncertainty to which the system is subjected. Therefore, the above mentioned
methods suffer from the lack of accuracy on tracking the desired trajectory, since the
error of the transformation between joint space and Cartesian space is not taken into
consideration. This error may affect the accurate performance of the exoskeleton
robot.
To address the above problems, we propose time delay estimation based on
Cartesian sliding mode controller that allows an exoskeleton robot dealing with
dynamic uncertainties. The adaptation of dynamic uncertainties is done by a new
time delay approach, where the time delay error (TDR) is taken into consideration.
The proposed controller provides a high level of robustness and accuracy by
designing a control law consisting of both sliding Cartesian surface and sliding joint
surface. This approach may limit the effect of the transformation kinematics error
on the performance of the system and reduces the chattering effect.The proposed
strategy is not limited to perform a passive rehabilitation but also qualified to
204 B. Brahmi et al.

perform an active rehabilitation. To complete this protocol, we rely necessarily

on the estimation of the Desired Movement Intention (DMI) of the subject using
indirect force control. This latter is done by Damped Least Square method (DLS),
which has been successfully used in multiple applications (Nakamura and Hanafusa
1986; Wampler 1986). The DLS approach aims to provide a compromise between
robustness of the solution and accuracy of the robot’s performance. Besides,
the proposed control is evaluated with healthy subjects.Anappropriate Lyapunov
function is given to solve the adaptation problem systematically, proving the closed-
loop stability and ensuring the asymptotic convergence of the output errors. The
principal benefit of the designed control is an accurate knowing of the dynamic
model, the external disturbances of the robot system are not needed, and afford
exceptional tracking performance regardless the existence of uncertainties.
The rest of the chapter is organized as follows. In the Sect. 9.2, the dynamics
of the ETS-MARSE is presented. The control approach is detailed in Sect. 9.3.
Simulation results are exhibited in Sect. 9.4; finally, the conclusion is presented in
Sect. 9.5.

9.2 Dynamics of the ETS-MARSE

9.2.1 Modeling of the Upper Limb ETS-MARSE Robot

Definition

The developed exoskeleton robot ETS-MARSE1 is a redundant robot consisted

of 7 degrees of freedom (DOFs), as shown in Fig. 9.1. It is created to provide
an assistive physical therapy motion to the injured upper limbs. The idea of the
designed exoskeleton is basically extracted from the anatomy of the upper limb of
the human to be ergonomic for their wearer along the physical therapy session.The
shoulder part consists of three joints, the elbow part comprises by one joint and the
wrist part consists of three joints. Each part from the previous parts is responsible
for performing a variety of upper limb motions as we show in Table 9.2. The design
of the ETS-MARSE has special features compared with the existing exoskeleton
robots. All special characteristics of the ETS-MARSE and comparison with similar
existing exoskeleton robots are summarized in (Rahman et al. 2013). The modified
Denavit-Hartenberg (MDH) parameters are given in Table 9.1. These parameters
are obtained from frames reference as shown in Fig. 9.1, and are used to obtain the
homogeneous transformation matrices.The workspace of the designed robot is given
in Table 9.2 and the details of the parameters and design of ETS- MARSE are given
in Rahman et al. (2013, 2015).

1 École de Technologie Supérieure – Motion Assistive Robotic-exoskeleton for Superior Extremity.

9 SMC of an Upper Extremity Exoskeleton Robot for Rehabilitation 205

Fig. 9.1 ETS-MARSE robot system (Rahman et al. 2013)

Table 9.1 Modified Joint (i) ai−1 ai di θi

Denavit-Hartenberg
parameters 1 0 0 ds θ1
2 − π2 0 0 θ2
π
3 2 0 de θ3
4 − π2 0 0 θ4
π
5 2 0 dw θ5
6 − π2 0 0 θ6 − π
2
7 − π2 0 0 θ7

9.2.2 Dynamics of the ETS-MARSE Robot

The dynamic equation of the ETS-MARSE robot can be shown in joint space, using
the Lagrangian method as follows:

M(θ )θ̈ + C(θ, θ̇ )θ̇ + B0 θ̇ + G(θ ) + d(t) = τ (9.1)

x = J (θ )θ̇ (9.2)
206 B. Brahmi et al.

Table 9.2 Workspace ETS-MARSE

Joints Motion Workspace
1 Shoulder joint horizontal flexion/extension 0◦ /140◦
2 Shoulder joint vertical flexion/extension 140◦ /0◦
3 Shoulder joint internal/external rotation −85◦ /75◦
4 Elbow joint flexion/extension 120◦ /0◦
5 Forearm joint pronation/supination −85◦ /85◦
6 Wrist joint ulnar/radial deviation −30◦ /20◦
7 Wrist joint flexion/extension −50◦ /60◦

where:
• θ , θ̇ and θ̈ ∈ R7 are respectively the joints position, velocity, and acceleration
vectors.
• ẋ ∈ R7 is the Cartesian velocity.
• M(θ ) ∈ R7×7 is the symmetric positive-definite inertia matrix, C(θ, θ̇ ) ∈ R7×7
is the Coriolis and centrifugal matrix, G(θ ) ∈ R7 is the gravity vector, and B0 ∈
R7×7 is the diagonal viscous friction matrix.
• d(t) ∈ R7 is the unknown bounded disturbance caused by the patient muscular
activity.
• Finally, τ ∈ R7 is the input torques vector, and J (θ ) ∈ R6×7 is the Jacobian
transformation matrix.
Without loss of generality, the inertia matrix can be rewritten as follows:

M(θ ) = M0 (θ ) + ΔM(θ ) (9.3)

where M0 (θ ) is the known part of inertia matrix, and ΔM(θ ) is the uncertain part.
Hence, the dynamic model expressed in (9.1) is rewritten as follows:

M0 (θ )θ̈ + U (δ, θ̈ , θ̇ , θ, t) = τ (9.4)

where,
• U (δ, θ̈ , θ̇ , θ, t) = ΔM(θ )θ̈ + C(θ, θ̇ )θ̇ + F (θ, θ̇ ) is the total unknown nonlinear
dynamic uncertainties of the robot.
• δ is the uncertain parameters of the dynamic model
• F (θ, θ̇ ) = B0 θ̇ + G(θ ) + d(t)
Let us denote M0 (θ ) = M0 and U (δ, θ̈ , θ̇ , θ, t) = U (t) for an easier handling of
the notation of the control design.
Several properties of system exoskeleton robot are given in (9.4) to prove the
stability such that:
Property 1 The inertia matrix M0 (θ ) is symmetric, bounded and positive definite
for all θ ∈ Rn (Lewis et al. 2003).
9 SMC of an Upper Extremity Exoskeleton Robot for Rehabilitation 207

Assumption 1 The joint position and joint velocity are measured.

Assumption 2 The pseudo-Jacobian matrix is non-singular.
Assumption 3 The desired trajectory is bounded.
Assumption 4 The function U (t) is a locally Lipschitz function.

9.3 Control Design

The objective of the suggested control is to ensure that the measured Cartesian
position x of the robot follows the desired trajectory xd along the designed
therapeutic task. The proposed controller is aimed to provide a high level of
accuracy even if the dynamic model of the robot is not completely known and in
the existence of external disturbances. All input and output signals in the closed-
loop system stay bounded. Along the tracking trajectory, an accurate adaptation of
the dynamic model was achieved. The closed-loop of the overall system is presented
in Fig. 9.2.
We can start now by the determination of the Cartesian position error and the
Cartesian velocity error as follows:

e = x − xd
(9.5)
ė = ẋ − ẋd

Fig. 9.2 Block diagram of the proposed controller

208 B. Brahmi et al.

where, xd ∈ R6 is the desired position and rotation of Cartesian trajectory and

x ∈ R6 is the measured position and rotation of the end-effector of the exoskeleton
robot.
Let us define the required Cartesian velocity as:

ẋr = ẋd − βex (9.6)

where x ẋr ∈ R6 and β is a scalar positive constant. The derivative of Eq. (9.6) with
respect to time yields:

ẍr = ẍd − β ëx (9.7)

Using Eq. (9.2), we define the sliding Cartesian vector as follows:

sx = ẋ − ẋr = J (θ )θ̇ − ẋr (9.8)

The derivative of Eq. (9.8) with respect to time yields:

ṡx = ẍ − ẍr = J (θ )θ̈ + J˙(θ )θ̇ − ẍr (9.9)

We can now determine the required joint velocity vector as follows:

θ̇r = J + (θ )ẋr (9.10)

where J + (θ ) = J T (θ )[J (θ )J T (θ )]−1 is the pseudo-jacobian matrix.

The derivative of Eq. (9.10) with respect to time is gives:

θ̈r = J + (θ )ẍr + J˙+ (θ )ẋr (9.11)

It is important also to define the sliding joint vector. By using Eqs. (9.2), (9.6)
and (9.10), we have:

s = θ̇ − θ̇r = J + (θ )[(ẋ − ẋd ) + βex ] = J + (θ )sx (9.12)

Differentiating Eq. (9.12) with respect to time, one obtains:

ṡ = θ̈ − θ̈r (9.13)

Substituting (9.13) into (9.4), the motion equation (9.4) can be rewritten as
follows:

M0 ṡ + M0 θ̈r + U (t) = τ (t) (9.14)

In order to facilitate the stability analysis, let us define ṡ from Eq. (9.14) such
that:
9 SMC of an Upper Extremity Exoskeleton Robot for Rehabilitation 209

ṡ = M0−1 [τ (t) − U (t)] − θ̈r (9.15)

In this chapter, we proposed a new Cartesian adaptive tracking control tracking

by using the Jacobian transpose matrix such that:

/(t) + ε(t)
τ = M0 [−J T (θ )(k1 ex + k2 ėx ) − J T (θ )k4 sx − k3 s + θ̈r ] + U (9.16)

with k1 ∈ R6×6 , k2 ∈ R6×6 , k4 ∈ R6×6 , and k1 ∈ R7×7 being diagonal positive

matrices. U /(t) is the estimation of the uncertain part. ε(t) is the time delay error
that is defined such that:

ε(t) = L |t − (t − td )| (9.17)

where L is the Lipschitz constant; td is the sampling time. Let us start by the
/(t). If we have a sampling time very small and
determination of the uncertain part U
Assumption 4 is verified, we can estimate the variation of the uncertain dynamics
by conventional time delay control using Eq. (9.4):

/(t)  U (t − td ) = τ (t − td ) − M0 θ̈(t − td )
U (9.18)

where θ̈ (t − td ), the delayed acceleration input computed using the following

approximation:

θ (t) − 2θ (t − td ) + θ (t − 2td )
θ̈ (t − td ) = (9.19)
td2

Let us define the term of uncertainties error using Lipschitz condition (Assump-
tion 4) as follows:

(t) = U (t) − U
U /(t) = U (t) − U (t − td ) ≤ L|t − (t − td )| ≤ ε(t) (9.20)

In what follows, we will choose appropriate L to always achieve U(t) = ε(t), by

helping an approximation technique (Reyes-Sierra and Coello 2005). Consider the
Lyapunov function candidate that stabilizes the overall system as follows:

1 T 1
V = s s + exT (k1 + βk2 )ex (9.21)
2 2
The derivative of the above Lyapunov function with respect to time is obtained
as:

V̇ = s T ṡ + exT (k1 + βk2 )ėx (9.22)

Substituting ṡ from Eq. (9.15) and control input from Eq. (9.16) into Eq. (9.22),
using Eqs. (9.17) and (9.20), we find:
210 B. Brahmi et al.

V̇ = −s T J T (θ )(k1 ex + k2 ėx ) − s T J T (θ )k4 sx − s T k3 s

/(t)) + M −1 ε(t)) + exT (k1 + βk2 )ėx
+s T (−M0−1 (U (t) − U (9.23)
0

Substituting (9.12) into (9.23), we have:

V̇ ≤ −sxT k1 ex − sxT k2 ėx − sxT k4 sx − s T k3 s + exT (k1 + βk2 )ėx (9.24)

We have from (9.5), (9.6) and (9.8):

sx = ėx + βex (9.25)

Substituting Eq. (9.25) into Eq. (9.24), we find:

V̇ ≤ −(ėx + βex )T k1 ex − (ėx + βex )T k2 ėx − sxT k4 sx

−s T k3 s + exT (k1 + βk2 )ėx
≤ −ėxT k1 ex − βexT k1 ex − ėxT k2 ėx − βexT k2 ėx − sxT k4 sx
−s T k3 s + exT k1 ėx + βexT k2 ėx (9.26)

Finally, we obtain:

V̇ ≤ −βexT k1 ex − ėxT k2 ėx − sxT k4 sx − s T k3 s (9.27)

It is clear from the V̇ , which implies that the exoskeleton system is stable.

9.3.1 Active Assistive Motion

This section provides a summary of the estimation method of the Desired Movement
Intention (DMI) of the subject. Let us start with the definition of the desired
trajectory in the active assistive motion. In this protocol, the desired trajectory is
updated as follows (Ochoa-Luna et al. 2015):

xd = x + Δxd (9.28)

where x ∈ R6 is the measured joint position, and xd ∈ R6 is the estimated of

DMI.This quantity of movement is estimated from the measured user’s force. If
Δxd −→ 0, this means that the exoskeleton’s wearer stops exercising forces on the
force sensor, making the exoskeleton to decrease its motion and progressively, and
whenever xd = x, the exoskeleton rests in its most recent position, Δxd . Through
such a mode, the exoskeleton robot is enabled to track the user’s movement inten-
tion, while the trajectory tracking adaptive control here guarantees the compensation
9 SMC of an Upper Extremity Exoskeleton Robot for Rehabilitation 211

of the uncertainties dynamics of the robot including the arm weight of the subject.
Based on jacobian matrix, it is possible to write the Cartesian displacement (Δxd )
as follows (Craig 2005):

Δxd = J (θ )Δθd (9.29)

where θd ∈ R7 is the joint angular displacement. To estimate the desired movement

intention Δxd from the user’s force, we can use the following equation (Khan et al.
2016):

Fm = J (η1 )Δθd (9.30)

where F ∈ R6 is measured user’s force is the measured user’s force. To solve

the Eq. (9.7), we used Damped least squares (DLS) or Levenberg-Marquardt
stabilization (Wampler 1986). This approach was employed firstly for avoiding the
singularity of inverse kinematics solution (Nakamura and Hanafusa 1986). Rather
than merely obtaining the minimum vector Δθd that provides a best solution to
Eq. (9.7), we determine the value of Δθd that minimizes the quantity:

min ||J Δθd − Fm ||2 + λ2 ||Δθd ||2 (9.31)

Δηd

where 0 < λ < 1 is the damping factor and can be determined to be positive
depending upon the accurate estimation specifications. Since the sum of Eq. (9.8)
can be written as:
2   2
2 J F 2
2 2
2 λI − 0 2 (9.32)

Consequently, the DLS solution is:

Δθd = J T (J J T + λ2 I )−1 F (9.33)

It is sufficient to substitute Δθd in Eq. (9.29) to obtain the estimated DMI.

9.4 Experimental and Comparative Study

9.4.1 Experiments Set-Up

Implementation was carried out on the ETS-MARSE system described below.The

system consists of three processing units, the first is a PC where the top-level
commands are sent to the robot using LabVIEW interface, i.e. the control scheme
selection, joint or Cartesian space trajectory, gain adjustments, etc. This PC also
212 B. Brahmi et al.

receives the data after the robot task is executed to analyze its performance. The
other two processing units are part of a National Instruments PXI platform. Firstly,
a NI-PXI 8081 controller card with an Intel Core Duo processor; in this card, the
main operating system of the robot and the top-level control scheme are executed.
In our case, the proposed control as well as the estimation based on TDE approach,
at a sampling time of 500 s. Finally, at input-output level, a NI PXI-7813R remote
input-output card with an FPGA (Field programmable gate array) executes the low-
level control; i.e. aPI current control loop (sampling time of 50 s) to maintain the
current of the motors required by the main controller. Also, in this FPGA, the
position feedback via Hall-sensors (joint position), and basic input-output tasks are
executed.Force sensor feedback is important to accurately control the movement
of the exoskeleton. A high linearity 6 axis force sensor [NANO17-R-1.8-M2-
M1PCI,ATI industrial Automation] is so chosen to obtain accurate real-time force
measurements.This sensor is mounted on the tip of the exoskeleton robot. The
joint of the ETS-MARSE is powered by Brushless DC motors (Maxon EC-45, EC-
90) combined with harmonic drives (gear ratio 120:1 for motor-1, motor-2, and
motor-4 and gear ratio 100:1 for motor-3 and motors 5 and7) (Brahmi et al. 2018b;
Rahman et al. 2013). The diagram of the architecture and overview of the ETS-
MARSEsystem with a human subject is shown in Fig. 9.3. The control gains are
chosen manually as follows: k1 = 18I6 , k2 = 80I6 , k4 = 5.7I6 , k3 = 20I7 , L = 0.6,
β = 1.5 and λ = 0.6.
An experimental session was created to show the effectiveness of the designed
control system. The physical therapy tasks are performed by three different healthy
subjects (average age: 27 ± 4:6 years; average height: 170 ± 8:75 cm; average
weight: 75 ± 18 kg). In the session, the subject was comfortably seated relax in
a chair as we show in Fig. 9.3. Experimental session was separated into three
scenarios. In the two scenarios, the subject-1 (age: 28 years; height: 177 cm; weight:
83 kg) performed the designed task as rectangular form (Initial position −→ Target-

ETS- MARSE
PXI 8108 PXI-7813R (FPGA)

PC
End-user Iref
Motor
commands current
q
Hall
sensors
data

BITDC (500 μs) Motors’ current PI

Robot OS control (50us)
Inverse Kinematics Position feedback

Fig. 9.3 The diagram of the experiment architecture and overview of the ETS-MARSEsystem
with a human subject
9 SMC of an Upper Extremity Exoskeleton Robot for Rehabilitation 213

A −→ Target-B −→ Initial position). The initial position of the exoskeleton

robot is given where elbow joint position is at 90◦ . This part is followed directly
by a comparison study through evaluating the proposed controller with another
subject (age: 38 years; height: 173 cm; weight: 94 kg) which permits to influence
the exoskeleton robot by different kinds of disturbances. Where, the external
disturbances here are represented by different physiological conditions of subjects,
such as non-linear biomechanical characteristics of the musculoskeletal system and
the different payload of the upper-limb for each subject. In the last scenario, we
evaluated the proposed control to achieve the active assistive motion performed by
the subject-3 (age: 30 years; height: 178 cm; weight: 76 kg).

9.4.2 Results of Passive Rehabilitation Motion

The experimental results with the ETS-MARSE robot in Cartesian space conducted
by Subject-1: (age: 28 years; height: 177 cm; weight: 83 kg) using the proposed
control strategy are shown in Figs. 9.4, 9.5 and 9.6. From Fig. 9.4, we notice that the
reference trajectory (red line) nearly interlocked with the measured trajectory (black
line). In this case, we can say that these results are reasonably good. Figure 9.5
exhibits the Cartesian tracking errors as functions of time, where they are clearly
converging and smaller along the reference trajectory. Figure 9.6 displays that the
control input is bounded without any remarkable chattering.

Robot End Effector Desired and Measured Trajectory

0
10 Desired Trajectory
Target-A Meaured Trajectory
20
Z(cm)

30
40
Target-B
50
Start Position

-10 60
0 50
40
10 30
20 20
Y(cm) 10
X(cm)

Fig. 9.4 Workspace performance of ETS-MARSE on 3D Cartesian space performed by subject-1:

(age: 28 years; height: 177 cm; weight: 83 kg)
214 B. Brahmi et al.

×10-3 Cartesian variable- Position Errors Cartesian variable- Orientation Errors

5 2

α (Deg)
[X]cm

1
0
0
-5 -1
0 2 4 6 8 10 12 0 2 4 6 8 10 12
Time (sec) Time (sec)
×10-3
2 5

β (Deg)
[Y]cm

0
0
-5

-2 -10
0 2 4 6 8 10 12 0 2 4 6 8 10 12
Time (sec) Time (sec)
×10-3 0.5
5
γ (Deg)
0
[Z]cm

0
-0.5

-5 -1
0 2 4 6 8 10 12 0 2 4 6 8 10 12
Time (sec) Time (sec)

Fig. 9.5 Cartesian Error tracking of passive motion performed by subject-1

Joint-1 Torque Joint-2 Torque

0.5 5
Nm

Nm

0 0

-0.5 -5
0 2 4 6 8 10 12 0 2 4 6 8 10 12
Time (sec) Time (sec)
Joint-3 Torque Joint-4 Torque
0.5 5
Nm

Nm

0 0

-0.5 -5
0 2 4 6 8 10 12 0 2 4 6 8 10 12
Time (sec) Time (sec)
Joint-5 Torque Joint-6 Torque
0.2 2
Nm

Nm

0 0

-0.2 -2
0 2 4 6 8 10 12 0 2 4 6 8 10 12
Time (sec) Time (sec)
Joint-7 Torque
0.1
Nm

-0.1
0 2 4 6 8 10 12
Time (sec)

Fig. 9.6 Control inputs of passive motion performed by subject-1

9 SMC of an Upper Extremity Exoskeleton Robot for Rehabilitation 215

Robot End Effector Desired and Measured Trajectory

10 Desired Trajectory
Measured Trajectory
20 Target-A
Z(cm)

30

40
Target-B
50

Start Position

-10
0 60
50
10 40
30
20 20
Y(cm) 10
X(cm)

Fig. 9.7 Workspace performance of ETS-MARSE on 3D Cartesian space performed by subject-2:

(age: 38 years; height: 173 cm; weight: 94 kg)

In order to show the efficiency and feasibility of the proposed scheme, we

evaluated the proposed controller by using the exoskeleton with another subject
(age: 38 years; height: 173 cm; weight: 94 kg), which means different payload of
the upper limb of the human.
Figures 9.7, 9.8 and 9.9 present the Cartesian tracking (red is the reference
trajectory, black is real trajectory) performed by Subject-2: (age: 28 years; height:
177 cm; weight: 83 kg) using the same gain’s values that employed in the first
trial. In fact, we remark from figures (Figs. 9.7, 9.8 and 9.9) that the proposed
control maintains a good performance. It is clear from Fig. 9.8 that the Cartesian
error is getting smaller with time. Figure 9.9 illustrates a smooth and bounded
control input. From the comparison of the two experimental results, we can
conclude that the proposed control algorithm strategy keeps providing a high level
of precision and robustness to the nonlinear uncertain dynamics and unknown.
So,these results confirm that the proposed control algorithm is proper to accomplish
the desired passive rehabilitation performance even than the nonlinear dynamics of
the exoskeleton robot are uncertain and the presence of external disturbances.

9.4.3 Results of Active Rehabilitation Motion

In this section, the subject exercises a free motion in 3D space. The objective of this
section is to evaluate the active rehabilitation motion and to show the effectiveness of
216 B. Brahmi et al.

×10-3 Cartesian variable- Position Errors Cartesian variable- Orientation Errors

10 0.5

α (Deg)
[X]cm

5 0

0 -0.5

-5 -1
0 2 4 6 8 10 12 0 2 4 6 8 10 12
Time (sec) Time (sec)
-3
×10 1
2

β (Deg)
0
[Y]cm

0 -1

-1 -2
0 2 4 6 8 10 12 0 2 4 6 8 10 12
Time (sec) Time (sec)
-3
×10 0.5
5

γ (Deg)
0 0
[Z]cm

-5 -0.5

-10 -1
0 2 4 6 8 10 12 0 2 4 6 8 10 12
Time (sec) Time (sec)

Fig. 9.8 Cartesian Error tracking of passive motion performed by subject-2

Joint-1 Torque Joint-2 Torque

0.5 5
Nm
Nm

0 -5
0 2 4 6 8 10 12 0 2 4 6 8 10 12
Time (sec) Time (sec)
Joint-3 Torque Joint-4 Torque
0.5 5
Nm

Nm

0 0
-0.5 -5
0 2 4 6 8 10 12 0 2 4 6 8 10 12
Time (sec) Time (sec)
Joint-5 Torque Joint-6 Torque
1 2
Nm

Nm

0 0
-1 -2
0 2 4 6 8 10 12 0 2 4 6 8 10 12
Time (sec) Time (sec)
Joint-7 Torque
0.2
Nm

-0.2
0 2 4 6 8 10 12
Time (sec)

Fig. 9.9 Control inputs of passive motion performed by subject-2

9 SMC of an Upper Extremity Exoskeleton Robot for Rehabilitation 217

Robot End Effector Desired and Measured Trajectory

0
10
20 Desired Trajetory
Measured Trajectory
30
Final Position
Z(cm)

40
50 Start Position

60
70

-20

0
60
20 50
40
30
20
Y(cm) 40 10
X(cm)

Fig. 9.10 Workspace performance of ETS-MARSE on 3D Cartesian space performed by subject-

3(age: 30 years; height: 178 cm; weight: 76 kg)

the estimation DMI of the subject. In this case, the exoskeleton robot is completely
passive and the subject is completely active. The Damped Least Square (DLS)
algorithm is used in order to provide the estimation of the desired movement
intention (DMI). Figures 9.10, 9.11 and 9.12 present the results of this experiment.
Figure 9.10 presents the performance of the subject-3 (age: 30 years; height:
178 cm; weight: 76 kg) with helping of ETS-MARSE exoskeleton robot (red line is
desired and the blue line is the measured trajectory) using Damped Least Square
(DLS) algorithm. It is clear from these plots (Figs. 9.10, 9.11 and 9.12) that the
control strategy achieved the desired performance with small tracking errors and
acceptable control input. From the good performance of the exoskleton-subject-
3 (Fig. 9.11), we can infer that the Damped Least Square (DLS) algorithm was
qualified to estimate accurately the desired movement intention Δxd . It is easy to
conclude from these plots (Figs. 9.10, 9.11 and 9.12) two points. The first one is that
the proposed control approach achieved the desired active motion performance with
high characteristics (small tracking errors and acceptable control input). The second
one is the algorithm of DLS permits to estimate the DMI accurately.
218 B. Brahmi et al.

Cartesian variable- Position Errors Cartesian variable- Orientation Errors

0.02 2

α (Deg)
[X]cm

0.01
0
0

-0.01 -2
0 5 10 15 20 25 0 5 10 15 20 25
Time (sec) Time (sec)
×10-3
5 1

β (Deg)
[Y]cm

0 0

-5 -1

-10 -2
0 5 10 15 20 25 0 5 10 15 20 25
Time (sec) Time (sec)
0.01 0

γ (Deg)
0 -0.2
[Z]cm

-0.01 -0.4

-0.02 -0.6
0 5 10 15 20 25 0 5 10 15 20 25
Time (sec) Time (sec)

Fig. 9.11 Cartesian Error tracking of active motion performed by subject-3

Joint-1 Torque Joint-2 Torque

2 10
Nm

Nm

0 0

-2 -10
0 5 10 15 20 25 0 5 10 15 20 25
Time (sec) Time (sec)
Joint-3 Torque Joint-4 Torque
1 10
Nm

Nm

0 0

-1 -10
0 5 10 15 20 25 0 5 10 15 20 25
Time (sec) Time (sec)
Joint-5 Torque Joint-6 Torque
1 1
Nm
Nm

0 0

-1 -1
0 5 10 15 20 25 0 5 10 15 20 25
Time (sec) Time (sec)
Joint-7 Torque
0.5
Nm

-0.5
0 5 10 15 20 25
Time (sec)

Fig. 9.12 Control inputs of active motion performed by subject-3

9 SMC of an Upper Extremity Exoskeleton Robot for Rehabilitation 219

9.5 Conclusions

In this chapter, we have proposed a Cartesian adaptive control strategy based on time
delay estimation and a robust proportional sliding mode controller which permits
to a redundant exoskeleton robot dealing with nonlinear uncertain dynamics and
to perform a passive and active rehabilitation treatment. The key advantage of the
designed controller is that precise information about the dynamic of ETS-MARSE
robot is not required. The robustness and the highest precision of the controller have
been achieved along a desired therapeutic task. The stability has been ensured for the
ETS-MARSE exoskeleton robot in closed loop, based on a Lyapunov theorem. The
experimental results have confirmed the efficiency and feasibility of the proposed
algorithm. In future work, we will seek to overcome the limitations of this approach.
In particular, the value of the delayed acceleration for the controller; where the
estimation of this variable may deteriorate its accuracy. We are also leaning to use
this controller with a more complex scenario.

References

Baek, J., Jin, M., & Han, S. (2016). A new adaptive sliding-mode control scheme for application
to robot manipulators. IEEE Transactions on Industrial Electronics, 63(6), 3628–3637.
Brahim, B., Maarouf, S., Luna, C. O., Abdelkrim, B., & Rahman, M. (2016a). Adaptive iterative
observer based on integral backstepping control for upper extremity exoskelton robot. In 8th
International Conference on Modelling, Identification and Control (ICMIC) (pp. 886–891).
Piscataway: IEEE.
Brahim, B., Rahman, M. H., Saad, M., & Luna, C. O. (2016b). Iterative estimator-based nonlinear
backstepping control of a robotic exoskeleton. World Academy of Science, Engineering and
Technology, International Journal of Mechanical, Aerospace, Industrial, Mechatronic and
Manufacturing Engineering, 10(8), 1313–1319.
Brahmi, B., Saad, M., Lam, J. T. A. T., Luna, C. O., Archambault, P. S., & Rahman, M. H. (2018a).
Adaptive control of a 7-DOF exoskeleton robot with uncertainties on kinematics and dynamics.
European Journal of Control, 42, 77–87.
Brahmi, B., Saad, M., Luna, C. O., Rahman, M. H., & Brahmi, A. (2018b). Adaptive tracking
control of an exoskeleton robot with uncertain dynamics based on estimated time delay control.
IEEE/ASME Transactions on Mechatronics, 23, 575–585.
Chen, P., Chen, C.-W., & Chiang, W.-L. (2009). GA-based modified adaptive fuzzy sliding mode
controller for nonlinear systems. Expert Systems with Applications, 36(3), 5872–5879.
Craig, J. J. (2005). Introduction to robotics: Mechanics and control. Upper Saddle River:
Pearson/Prentice Hall.
De Morand, A. (2014). Pratique de la rééducation neurologique. Issy-les-Moulineaux: Elsevier
Masson.
Fridman, L. (1999). The problem of chattering: An averaging approach. In Variable structure
systems, sliding mode and nonlinear control (pp. 363–386). London: Springer.
Khalil, H. K., & Grizzle, J. (1996). Nonlinear systems. New Jersey: Prentice Hall.
Khan, A. M., Yun, D.-W., Ali, M. A., Zuhaib, K. M., Yuan, C., & Iqbal, J. (2016). Passivity
based adaptive control for upper extremity assist exoskeleton. International Journal of Control,
Automation and Systems, 14(1), 291–300.
220 B. Brahmi et al.

Lewis, F. L., Dawson, D. M., & Abdallah, C. T. (2003). Robot manipulator control: Theory and
practice. Boca Raton: CRC Press.
Nakamura, Y., & Hanafusa, H. (1986). Inverse kinematic solutions with singularity robustness for
robot manipulator control. Journal of Dynamic Systems, Measurement, and Control, 108(3),
163–171.
Ochoa-Luna, C., Habibur Rahman, M., Saad, M., Archambault, P. S., & Bruce Ferrer, S. (2015).
Admittance-based upper limb robotic active and active-assistive movements. International
Journal of Advanced Robotic Systems, 12(9), 117.
Park, C.-W., & Cho, Y.-W. (2007). Robust fuzzy feedback linearisation controllers for Takagi-
Sugeno fuzzy models with parametric uncertainties. IET Control Theory & Applications, 1(5),
1242–1254.
Rahman, M. H., Saad, M., Kenné, J.-P., & Archambault, P. S. (2013). Control of an exoskeleton
robot arm with sliding mode exponential reaching law. International Journal of Control,
Automation and Systems, 11(1), 92–104.
Rahman, M. H., Rahman, M. J., Cristobal, O., Saad, M., Kenné, J.-P., & Archambault, P. S. (2015).
Development of a whole arm wearable robotic exoskeleton for rehabilitation and to assist upper
limb movements. Robotica, 33(1), 19–39.
Reyes-Sierra, M., & Coello, C. A. C. (2005). A study of fitness inheritance and approximation
techniques for multi-objective particle swarm optimization. In IEEE Congress on Evolutionary
Computation, Edinburgh (pp. 65–72).
Shieh, H.-J., & Hsu, C.-H. (2008). An adaptive approximator-based backstepping control approach
for piezoactuator-driven stages. IEEE Transactions on Industrial Electronics, 55(4), 1729–
1738.
Slotine, J.-J. E., & Li, W. (1991). Applied nonlinear control. Englewood Cliffs: Prentice-Hall.
Wampler, C. W. (1986). Manipulator inverse kinematic solutions based on vector formulations and
damped least-squares methods. IEEE Transactions on Systems, Man, and Cybernetics, 16(1),
93–101.
Wang, L., Chai, T., & Zhai, L. (2009). Neural-network-based terminal sliding-mode control of
robotic manipulators including actuator dynamics. IEEE Transactions on Industrial Electron-
ics, 56(9), 3296–3304.
Xie, S. (2016). Advanced robotics for medical rehabilitation. Cham: Springer.
Yim, J., & Park, J. H. (1999). Nonlinear H ∞ control of robotic manipulator. In Proceedings of
IEEE International Conference on Systems, Man, and Cybernetics, SMC’99. Proceedings 1999
(pp. 866–871). Piscataway: IEEE.
Youcef-Toumi, K., & Ito, O. (1990). A time delay controller for systems with unknown dynamics.
Journal of Dynamic Systems, Measurement, and Control, 112(1), 133–142.
Young, K. D., Utkin, V. I., & Ozguner, U. (1999). A control engineer’s guide to sliding mode
control. IEEE Transactions on Control Systems Technology, 7(3), 328–342.
Chapter 10
Canonical Particle Swarm Optimization
Algorithm Based a Hybrid Vehicle

Mohamed Elhedi Hmidi, Ines Ben Salem, and Lilia El Amraoui

Abstract This paper deals with the modeling and optimization of a PID 2 DOF
controller design for a Hybrid vehicle. Such a Particle Swarm Optimization (PSO)
based PID 2Dof controller is investigated in order to stabilize both the velocity
of studied vehicle. The aim goal of this paper is to improve the effectiveness of
synthesized control using the strategy of a canonical PSO optimization to tune
its weighting matrices instead to configure it by a trials-errors method. This work
reminds firstly to describe all aerodynamic forces and moments of the hybrid vehicle
within an inertial frame and a dynamical model is obtained thanks to the Lagrange
formalism. A 2Dof PID controller is then designed for the Velocity stabilization
of the studied vehicle. Several PSO updating strategies are proposed to enhance
the stability and the rapidity of our studied system through minimizing a definite
cost function of’ controller’s weighting matrices. The obtained results are carried
out in order to show the effectiveness and robustness of the different PSO updating
strategies based the 2Dof PID Controller.

Keywords Hybrid vehicle · Modeling · PID two degree of freedom control ·

Driving cycle · Particle swarm optimization · PSO velocity stabilization

10.1 Introduction

In recent years, research and developments related to the vehicle have attracted
much attention. Improvements in the autonomy, flexibility, and adaptability of these
vehicles have been proposed during the thermal phase (Mensing and Trigui 2011).
The hybrid vehicle is now one of the ways to reduce the energy consumption.

M. E. Hmidi () · I. B. Salem · L. El Amraoui

Research Unit Signals and Mechatronics System, National Engineering School of Carthage
(ENI-Carthage), University of Carthage, Carthage, Tunisia
e-mail: [email protected]; [email protected]

© Springer Nature Singapore Pte Ltd. 2019 221

N. Derbel et al. (eds.), New Developments and Advances
in Robot Control, Studies in Systems, Decision and Control 175,
https://doi.org/10.1007/978-981-13-2212-9_10
222 M. E. Hmidi et al.

This vehicle hybrid consists of Electric assembly and thermal with a tilting
mechanism to toggle between these phase of driving.
This cycle of driving of the vehicle combines the advantages of electric and
thermal vehicles allowed to experiment the best effects of the electric phase and
to minimize the energy consumed in thermal phase.
This is very useful, compared to others vehicles classical, However, these
structures design bring its own problems since the degradation in stability is usual-ly
observed in the thermal phase of driving.
The history of hybrid vehicles (HV) dates back to the late nineteenth century
(Trigui and Jeanneret 2003). However, because of their limited autonomy, the
industry has further promoted the evolution and use of this new vehicle type, taking
into ac-count the availability and price of oil. Following oil price inflation and taking
into account environmental constraints, a great deal of research has been carried out
to market a new generation of HVs equipped with cleaner sources of renewable
energy. Different approaches have been studied to model the vehicles in order to
increase their autonomy while minimizing the fuel consumption of the main source.
In recent years, these hybrid vehicles have seen a great evolution in terms of the
optimization of energy consumption, the modeling and especially the driving control
design (Romero et al. 2016). This explains the interest of researchers to study their
driving dynamics and improve controller performances of this kind of hybrid vehicle
thanks to different metaheuristic optimization algorithms. We find in the literature
several works that focus to enhance controller’s efficiencies thanks to tuning its
decision parameters. Among these works (Amir Ghoreishi and Basiri 2011; Hamidi
2012) the authors presented an LQR controller designed that adjusted through its
weighting matrices using both PSO and GA algorithms. An Intelligent Fuzzy Logic
Control proposed by Zoric et al. (2014) based on PSO strategy to improve the
robustness of parameters. The work of Prasad and Tyagi (2012) proposed a new
algorithm (ABC) provided to adjust the weighting matrices of LQG controller for
the antenna system.
Thus, a dynamic model of this type of hybrid vehicle is established thanks to
the Lagrange formalism (Romero et al. 2016) We are interested in the dynamic
modeling of the thermal phase based on the nonlinear model of this system studied
obtained from a sub model to make the system easy to control. Control methods
of these nonlinear systems have been received much attention and become one of
the most important topics in this research fields, Many control schemes such as PID
classical controller, Two-Degree-of-Freedom PID Controllers have been employed
for the velocity control and stabilization of different structures of the driving vehicle.
For a conventional PID regulator structure, there are several versions, for
example, the serial and parallel structure and other structures depending on the
degree of freedom. The two main objectives of this controller are the follow-
up of the Set-point and the rejection of the load disturbances. Using the 2DOF
structure introduces other parameters that need to be defined precisely. The History
of regulators is already long and it may be interesting to recall some important steps,
the first centrifugal type regulators appeared around 1750 to regulate the speed of
10 Canonical Particle Swarm Optimization Algorithm Based a Hybrid Vehicle 223

windmills, followed in 1788 by the famous speed controller of a steam engine Of

James Watt.
The PID regulators with two degrees of freedom (2DOF) include a setpoint
weighting on proportional terms and derivatives. A 2DOF PID controller is capable
of quickly rejecting disturbances without increasing the overshoot to the set-point
level. These controllers are also useful for attenuating the influence of variations of
the reference signal on the control signal.
This control approach is a promising design technique that shows many appli-
cations in various engineering fields. Vilanova (2012) proposed a Model Reference
Robust Tuning of PID Controllers. The PID 2DOF control strategy is also widely
applied in the classical and modern vehicle field, due to its control methodology.
It has been applied successfully to the regulation of disturbance rejections, it also
remains at a given set-point and the follow-up of the controls whose controlled
variable is good in monitoring the desired value (Araki 2003). So, in this document,
such a PID 2DOF control strategy is used for controlling the velocity of the thermal
phase of a hybrid vehicle, as well as the complete modeling of the traction chain
also presented.
The remainder of this article is organized as follows. Section 10.2 presents the
nonlinear mathematical model of the studied driving phase based on the Lagrange
formalism. In Sect. 10.3, we present the formulation of the design problem of the
2DOF PID controller and provide the calculated control laws. Section 10.4 describes
the demonstrative simulations that are performed to validate the proposed control
approach. Finally, the conclusions are drawn from Sect. 10.5.

10.2 Dynamic System Model

10.2.1 Vehicle Model

The model of the thermal phase of the hybrid vehicle is a rigid body, in this model
Let us consider an Internal Combustion Engine powered Vehicle (ICEV) and let us
make the hypothesis is that it is possible to measure the speed v(t) of the vehicle
during a trip.
Using the fundamental principle of dynamics that is written:

→ →
Ma = F (10.1)

→ →
where a is the vehicle acceleration, F the sum of streaks forces on the vehicle. We
→ →
notice Fx the tractive force abandoned the wheels supplied by the motor and Fext all
→
external forces. So the main resistance forces are the aerodynamic forceFA , rolling
→ →
resistance FR and the force of gravity FG . In Fig. 10.1, they are shown by their
standards denoted respectively FA , FR and FG and their directions shown by the
224 M. E. Hmidi et al.

Z
FA

x Fc

FR m3

Fx
Hybrid Vehicle q

Fig. 10.1 Schematic of the system architecture

arrows. In the case of a rise with an angleθ , all those forces opposed to the direction
→
of movement of the vehicle. The tractive force Fx is itself in the direction of path.
And each can be expressed as follows.
Aerodynamic Force is calculated as:

1
FA = ρSCx vr2 (10.2)
2
Rolling Resistance force is formulates using:

F R = Mg CR (10.3)

The gravity force using:

FG = Mg sin θ (10.4)

with S, the frontal surface Cx , the aerodynamic drag coefficient CR , the rolling
resistance coefficient g, the gravity acceleration and θ , the road slope angle.

10.2.2 Force of the Tires Provided by the Engine

The engine is directly connected to the transmission and to the wheels. It participates
parallel to the movement of the vehicle, and transmits a mechanical power to the
wheels.
The conventional vehicle is only equipped with a conventional combustion
engine running on gasoline. The mechanical braking device is loading slow the
vehicle during braking. The torque engine Te at the output of the motor shaft is
delivered to the input of the gearbox via the clutch. Considering Ie the engine
rotational inertia, we the acceleration angular of engine, so the torque to the output
of the clutch is:
10 Canonical Particle Swarm Optimization Algorithm Based a Hybrid Vehicle 225

.
Tc = Te − Ie w (10.5)
e

At the output of the gearbox, the torque engine at the input is amplified by the
gear ratio, then we noted:
.
Td = (Tc − It ω)Nt (10.6)
e

where Nt the value of the transmission ratio of the box and It the rotational inertia
of the transmission, at the same principle, the torque Ta is needed to turn the wheels
and providing the transmission of force:
.
Ta = (Td − Id ω)Nf (10.7)
d

where Id is the rotational inertia of the drive shaft, and Nf the differential gear ratio.
Finally combining all the above equations, the torque supplied by the engine
through the transmission and given by:
 . 
Ta = [Te − (Ie + It ) ω]Nt − Id ωd Nf (10.8)
e

In the existence of the braking torque, the actual torque Tx created by this
equation:
.
TX = rFx + Iω w = Ta − Tb (10.9)
ω

We conclude, the tractive force to the wheels delivered by the combustion engine
can be written as:
1# $ . $
Fx = ηTe N − {Iet N 2 + Id Nf2 w −Tb (10.10)
r ω

where: Iet = Ie + It , N 2 = Nt2 Nf2 and N = Nt Nf .

10.2.3 Equation of Movement

The tractive force is knowing Fx provided by the engine and the different forces
opposing the motion of the vehicle, using the fundamental principle of dynamics,
so we get the equation of motion of the vehicle approximated by the equation:
.
M ν = Fx − FR − FA − FG (10.11)

with the absence of longitudinal sliding, the angular velocity of the wheel, is given
by the following equation:
226 M. E. Hmidi et al.

. .
ν =rw (10.12)
ω

where r is the wheel radius, and Ww is the angular velocity of the wheel.
In conclude therefore the final equation of motion of the vehicle:

ηTe N # $ ν .
.
Mν = − Iet N 2 + Id Nf2 + Iω 2 − H (10.13)
r r
So we note:
 
Iet N 2 + Id Nf2 + Iω . ηTe N − Tb
M+ ν= −H (10.14)
r2 r

Finally we find:
. r
ν= × (ηTe N − Tb − rH ) (10.15)
Mr 2 + Iet N 2 + Idω

The longitudinal modeling of the vehicle and therefore described by the follow-ing
system of equations:
⎧.
⎨ x = νx
. r (10.16)
⎩ νx = × (ηTe N − Tb − rH )
Mr 2 + Iet N 2 + Idω

10.3 Control Design Strategies

10.3.1 PID Control

The field of hybrid vehicle control the main objective of all controllers is the
reduction of error in the majority of cases. In this case, the PID controller is one
of the solutions, because of its simplicity and its best performance in the majority
of cases. Optimizing the PID controller is the main task for better performance.
The PID control system is characterized by its three correction terms, proportional,
integral and derivative (Haugen 2012; Taguchi 2000), as shown by the PID structure
of Fig. 10.2.
Defining u(t) as the controller output of the PID, is expressed by this following
equation:
,
d
u(t) = Kp.e(t) + Ki e(t) + Kd e(t) (10.17)
dt
10 Canonical Particle Swarm Optimization Algorithm Based a Hybrid Vehicle 227

Fig. 10.2 PID controller structure of hybrid vehicle

Fig. 10.3 2DOF PID controller structure of hybrid vehicle

where Kp , Ki ,Kd are respectively the controller gain, the integral time constant and
the derivative time

10.3.2 PID2-DOF Control Problem Formulation

Since the hybrid vehicle model thermal phase is strongly coupled and non-linear, the
proposed PID 2DOF control system is a multi-objective problem. A general form of
the 2DOF PID controller is shown in Fig. 10.3, where the controller consists of two
compensators A1 (s) and A2 (s).
The PID 2DOF control laws design for the Thermal Phase model is achieved
according to the procedure proposed by Kumar and Vandana (2015).
228 M. E. Hmidi et al.

10.3.3 Velocity Controller

Since the controller PID 2DOF is a controller with two inputs, an output of the form
A(s), as shown in Fig. 10.4. The transfer function of each input to the output is itself
a PID controller. This allows us to introduce the following equations which are used
to control the speed of our vehicle can be written respectively as:
 
Ki Kd
A1 (s) = c1 Kp + + c2 (10.18)
s 1 + Tf s
 
Ki Kd
A2 (s) = Kp + + (10.19)
s 1 + Tf s

where: Kp , Ki ,Kd are respectively the controller gain, the integral time constant and
the derivative time, constant, c1 and c2 the set-point weights, and Tf the derivative
filter constant.
The relationship between the two inputs (r and y) and the output (u) of the 2DOF
PID controller can be represented in a parallel manner. The two forms are different,
one can therefore express the proportional, integral and derivative actions of the
controller.
The Independent gains or “parallel” 2DOF PID control algorithm is expressed
by the equation follows:

Ki Kd s
u = Kp (c1 r − y) + (r − y) + (c2 r − y) (10.20)
s 1 + Tf s

where, u is the Manipulated variable, y is the controlled variable or output, and r is

the set-point.

Fig. 10.4 Particle’s position

and velocity update
10 Canonical Particle Swarm Optimization Algorithm Based a Hybrid Vehicle 229

10.4 Proposed Particle Swarm Optimization Algorithms

10.4.1 Basic Concept

The Particle Swarm Optimization (PSO) metaheuristic is a swarm intelligence

Algorithm based computation algorithm originally proposed by Eberhart and Shi
(2012), This new stochastic and global optimization technique is inspired by the
collaborative behaviour of biological populations and organisms such as the birds’
flocks and fish’s schools.
The information exchange between different population individuals allows solv-
ing various complex optimization problems in various engineering fields. These
applications can be summarized around domains of robotics, image and signal
processing, electronic circuits design, communication networks, but more especially
the domain of plant control design.

10.4.2 Canonical PSO Algorithm

The PSO algorithm uses a swarm consisting of np particles, i.e., x 1 , x 2 , . . . , x np

randomly distributed in the considered initial search space, to find an optimal
solution of a generic optimization problem (10.22). Each particle, that represents
a potential solution, is characterised by a position and a ve-locity given byxki :=
 T  T !!
xki,1 , xki,2 , . . . , xki,m and vki := vki,1 , vki,2 , . . . , vki,m where: (i, k) ∈ 1, np ×
[[1, kmax ]].
At each algorithm iteration, the ith particle position, evolves based on the
following update rules:
i
xk+1 = xki + vk+1
i
(10.21)
 
g
i
vk+1 = wvki + c1 r1,k
i
pki − xki + c2 r2,k
i
pk − xki (10.22)

where: w is the inertia factor, c1 and c2 are the cognitive and the social scaling
factors respectively, r1,ki and r i are random numbers uniformly distributed in the
2,k
i g
interval [[0, 1]] ,pk is the best previously obtained position of the ith particle and pk
is the best obtained position in the entire swarm at the current iteration k. Hence,
the principle of a particle displacement in the swarm is graphically shown in the
Fig. 10.4, for a two dimensional design space.
In the PSO formalism, a particle can leave the search space initially defined
during the motion of the swarm. Therefore, it should define constraints on the
problem decision variables, given as follows:

j i,j j
xmin ≤ xk ≤ xmax , ∀ (i, j, k) ∈ [[1, nP ]] × [[1, m]] × [[1, kmax ]] (10.23)
230 M. E. Hmidi et al.

Set-Point Position

Torque
Velocity
Engine Thermal Phase of
Hybrid Vehicle
Braking Dynamic model
Torque acceleration

Gearbox absorbed
power

Fig. 10.5 Vehicle hybrid, thermal phase block

where i, j and k represent the indexes on the particle in the swarm, the components
of the decision vector and the algorithm iterations, respectively (Fig. 10.5).
The following interval confinement mechanism
** is introduced
++ on the decision
j j
variables when a particle leaves the range xmin , xmax defined according to the
optimization problem:
#  $
i,j i,j i,j j j
xk = min max xk + vk , xmin , xmax (10.24)

In this case, Eq. (10.21) replaces the above PSO motion equation (10.24) and the
new velocity of the corresponding particle is updated accordingly as follows:
i,j i,j
ṽk = −μvk , 0 ≤ μ < 1 (10.25)

In order to improve the exploration and exploitation capacities of the proposed PSO
algorithm, we choose for the inertia factor a linear evolution with respect to the
algorithm iteration as given in Eberhart and Shi (2012):
 
wmax − wmin
wk+1 = wmax − k (10.26)
kmax

where wmax = 0.9 and wmin = 0.4 represent the maximum and minimum inertia
factor values, respectively, wmax is the maximum iteration number Finally, a pseudo
code of this swarm intelligence based algorithm is given for a minimization problem
in Canonical PSO Algorithm:
• Initialize a population of nP ART particles having random positions and velocities
on D dimensions of the search space
• At every iteration K and for each ith particle xki evaluate the considered
optimization fitness function on the D decision variables.
10 Canonical Particle Swarm Optimization Algorithm Based a Hybrid Vehicle 231

   
• Compare particle’s fitness evaluation fki = f xki with its pbestki = f pki . If
fki ≤ pbestki then pbestki = fki and pki = xki
• Identify the particle in the neighborhood with the best success so far and as-sign
its index to the variable g.
• Change the velocity and position of the particle according to the motion equations
(10.24) and (10.25).
If the last criterion is met (satisfied fitness or a maximum number of iterations), the
algorithm terminates with the solution, otherwise, go to Step 2:
#  $
x ∗ = arg min f xki , ∀i, k
xki

10.5 Obtained Results

The simulations of our dynamic model of the hybrid vehicle thermal phase are
first carried out by a system involved in Simulink/Matlab. Our model of the hybrid
vehicle thermal phase is detailed by the following block diagram (Fig. 10.6):

10.5.1 Simulation of the Model with the PID and the PID
2-DOF Controller

The control objectives of such a thermal phase of hybrid vehicle, whose physical
parameters are given in Table 10.1, are the closed-loop stabilization of its velocity
(Fig. 10.7)

2.5
Gearbox

1.5

0.5
0 10 20 30 40 50
Time[s]

Fig. 10.6 The gearbox

232 M. E. Hmidi et al.

Table 10.1 Control Parameters Values

parameters of the PSO
algorithms Dimension of search space d = 15
Population size np = 30
Number of generations kmax = 100
Cognitive coefficient c1 = 0.5
Social coefficient c2 = 0.5
Inertia weight w = 0.9

60
[Setpoint]

40

20

0
0 5 10 15 20 25 30 35 40 45 50
Time[s]

Fig. 10.7 The driving scenario

2000

1500
Position x(m)

1000

500

0
0 5 10 15 20 25 30 35 40 45 50
Time[s]

Fig. 10.8 Position of hybrid vehicle

The inputs of the model are the set-point, the gearbox and the torque engine
(Fig. 10.8).
The simulation results show the controlled position, and acceleration dynamics as
well as the thermal phase of hybrid vehicle power-train control inputs. High perfor-
mances in transient response and steady-state precision are guaranteed (Table 10.2).
This result shows the superiority in terms of exploration and exploitation capabil-
ities, solutions quality and fastness convergence of the PSO variants (Fig. 10.9). All
proposed PSO algorithms produce 2DOF PID controllers that stabilize the velocity
dynamics of the thermal phase of a hybrid vehicle as shown in Fig. 10.10.
It can be observed from these simulation results that the estimated and real
system states are close and similar to the velocity dynamics. In addition, the system
10 Canonical Particle Swarm Optimization Algorithm Based a Hybrid Vehicle 233

Table 10.2 Model Parameters Description Values

parameters of the thermal
phase vehicle m Mass of vehicle 1269 kg
r Tire radius 0.269 m
Iet Motor inertia 0.2630 m
Id Inertia drive 0.115 kg m3
Iw Inertia of wheel 2.8 kg m3
SCx Aerodynamic 0.720
ρ Air density 1.205
CR Rolling resistance 0.020 kg m3
g Acceleration of the gravity 9.81 m/s2

400

200
Vehicle acceleration

-200

-400

-600

-800
0 5 10 15 20 25 30 35 40 45 50
Time[s]

Fig. 10.9 Velocity of hybrid vehicle

60

50

Reference
40
estimated state PSO
Vehicle Speed [km/h]

real state PID 2DOF

30

20
42
10
40

0 38

-10 36

-20 34
20 22 24
-30
0 5 10 15 20 25 30 35 40 45 50
Time[s]

Fig. 10.10 System responses using PID 2DOF controller based PSO
234 M. E. Hmidi et al.

became responds to its set points with a stable and rapid manner, due to the
tuned weighting matrices through the different algorithms. On the other hand, the
estimation errors are negligible and a good reconstitution of the system state is
obtained. The stabilization objectives of the 2DOF PID controller are made with
a satisfied tracking performance and the effectiveness of the proposed 2DOF PID
control approach is guaranteed.

10.6 Conclusion

In this paper, we established a nonlinear dynamical model of a Hybrid vehicle in the

thermal phase using the Lagrange formalism, extensively adopted in the lit-erature.
All aerodynamic forces and moments of the studied vehicle hybrid are described
within an inertial frame. Such an established dynamical model is then used to design
a 2DOF PID controller for the stabilization of the velocity of the vehicle. Parameters
design of the proposed 2DOF PID control approach, i.e., the weighting matrix
are obtained thanks to the PSO algorithm through weighted sum multi-objective
optimization problem. Finally, some demonstrative simulation results are obtained
under the MATLAB/Simulink environment in order to show the effectiveness of
tuned weighting parameters stabilization approach

References

Amir Ghoreishi, S., Nekoui, M. A., & Basiri, S. O. (2011). Optimal design of LQR weighting
matrices based on intelligent optimization methods. International Journal of Intelligent
Information Processing, 2, 63–74.
Araki, M., & Taguchi, H. (2003). Two- degree of freedom PID controllers. International Journal
of Control, Automation, and Systems, 1, 401–411.
Eberhart, R. C., & Shi, Y. (2012). Comparing inertia weights and constriction factors in particle
swarm optimization. Congress on Evolutionary Computation, 1, 84–88.
Hamidi, J. (2012). Control system design using particle swarm optimization (PSO). International
Journal of Soft Computing and Engineering, 1.
Haugen, F. (2012). The good gain method for simple experimental tuning of PID controllers.
Norwegian Society of Automatic Control.
Kumar, M., Vandana, P., & Patel, V. (2015). Two degree of freedom pid controller for speed control
of DC motor. American International Journal of Research in Science, Technology, Engineering,
Mathematics, 39, 94–97.
Mensing, F., Trigui, R., & Bideaux, E. (2011). Vehicle trajectory optimization for application in
eco-driving. In IEEE Vehicle Power and Propulsion Conference, Belfort.
Prasad, B., Tyagi, B., & Gupta, H. O. (2012). Modeling and simulation for optimal control
of nonlinear inverted pendulum dynamical system using pid controller, LQR. In Sixth Asia
Modeling Symposium, Kuala Lumpur.
Romero, J. A., Lozano-Guzman, A. A., Betanzo-Quezada, E., & Arroyo Contreras, G. M. (2016).
Cargo securement methods and vehicle braking performance. International Journal of Vehicle
Performance, 4, 353–373.
10 Canonical Particle Swarm Optimization Algorithm Based a Hybrid Vehicle 235

Taguchi, H., & Araki, M. (2000). Two-degree-of-freedom pid controllers their functions and
optimal tuning. IFAC Digital Control, 33, 5–7.
Trigui, R., Jeanneret, B., & Badin, F. (2003). Systemic modelling of hybrid vehicles in order to
predict dynamic performance and energy consumption. VEHLIB Library of Models.
Vilanova, V. A. R. (2012). Conversion formulae and performance capabilities of two-degreeof-
freedom PID control algorithms. In: 17th IEEE International Conference on Emerging
Technlogies and Factory Automation (ETFA), Kraków (pp. 17–21).
Zoric, N. D., Simonovic, A. M., Mitrovic, Z. S., Stupar, S. N., Obradovic, A. M., & Lukic, N.
S. (2014). Free vibration control of smart composite beams using particle swarm optimized
self-tuning fuzzy logic controller. Journal of Sound and Vibration, 333, 5244–5268.
Chapter 11
Digital Stabilizing and Control for
Two-Wheeled Robot

Bachir Nail, Abdellah Kouzou, and Ahmed Hafaifa

Abstract In this chapter a design control is proposed for a two-wheeled robot sys-
tem based on a combined between the matrix fraction description (MFD) theory and
digital PID controller, the model of the system is presented in nonlinear differential
equations and by using the Taylor series expansion method, the dynamics model is
linearized to be linear multivariable with two inputs (voltages) and two outputs (the
rotational speed and tilt angle) model. after the check of the block controllability
applied on the model, the transformation of the model into block controllable form
have been done. In order to stabilized the behavior dynamics of robot model, a set
of block poles are chosen diagonally were consists from stable and fast eigenvalues,
are assigned by matrix gain K. To make the outputs of the robot system track its set
references and to ensure a good controlling, two PID controllers are proposed, one
for the rotational speed output and the second for the tilt angle, the parameters of the
PID controllers (Kp , Ki and Kd ), are tuning and selected based on Ziegler-Nichols
method, through the simulations results performed illustrated the effectiveness of
the proposed design controller and it is more suitable and robustness against the
disturbances which were injected and shown that the robot can be confirmed as
mobile platform for transporting human and goods.

Keywords Two-wheeled robot · Multivariable system · Matrix fraction

description (MFD) · Block poles · PID controller · Ziegler-Nichols

Nomenclature

τm Motor torque(Nm)
TmL ,TmR Applied torque from motors to wheels

B. Nail () · A. Kouzou · A. Hafaifa

Applied Automation and Industrial Diagnostics Laboratory, Faculty of Sciences and Technology,
University of Djelfa, Djelfa, Algeria
e-mail: [email protected]; [email protected]; [email protected]

© Springer Nature Singapore Pte Ltd. 2019 237

N. Derbel et al. (eds.), New Developments and Advances
in Robot Control, Studies in Systems, Decision and Control 175,
https://doi.org/10.1007/978-981-13-2212-9_11
238 B. Nail et al.

τa Applied torque(Nm)
Hf L ,Hf R Friction forces
R Nominal resistance(Ω)
r Radius of wheels
L Rotor inductance(H)
θ Angle of body offset fromvertical
Nms
kf frictional constant
rad
θ̇ Angular velocity 
Nm
km Torque constant
A
θ̈ Angular acceleration 
Vs
ke Back emf coefficient
rad
θw Rotation angle of the wheels
Va Applied voltage (V )
w Angular velocity of shaft
Ve Back emf voltage (V )
α,ẇ Angular acceleration of shaft
i Armature current (A)
θ̇w Angular velocity of wheels
IR Rotor inertia (Kgm2 )
ẋ Linear speed
Mw Wheel mass
ẍ Linear acceleration
Mp Robot’s chassis mass
Tm Torque of the driving wheels
Iw Moment of inertia of the wheels
Tr Torque of the reaction wheels
Ip moment of inertia of the body
VDw Input voltage of the driving wheel
HL ,HR ,PL ,PL Reaction forces
VRw Input voltage of the reaction wheel

11.1 Introduction

There are various definitions for designating a mobile robot, but in general it can be
presented as a machine capable of autonomously carrying out a displacement from
one point to another in an environment not perfectly known a priori. To do this, the
robot must have the functionality of perception, decision on the movement to be
carried out, and realization of this movement.
Nowadays, mobile robotics is experiencing a real expansion with more and more
attractive applications, of which we can cite as examples: planetary exploration,
11 Digital Stabilizing and Control for Two-Wheeled Robot 239

exploration or intervention in environments hostile to humans (radioactive zone,

polar, submarine depths, presence of fire. . . ), or simply difficult to access, in the
industry, applications mainly concern transport or distribution, whether in factories,
mines, hospitals or workshops. An example is conveying in a often congested
environment, which consists of storing or shipping spare parts, medicines, waste, or
automatic soil cleaning. Another example is the automatic conveyance of vehicles
used in road transport.
The design of automated vehicles, or mobile robots, especially with wheels is a
rapidly expanding field of research. These robots are used in industry as a means
of transport, inspection or operation and are particularly suited to interventions in
hostile environments confined areas, high construction, underwater depths, etc.
Mobile robots with wheels are characterized by their simplicity of design.
However, these systems have raised several difficult problems that have been the
subject of several research studies. These are problems of control, planning, stability
and trajectory tracking.
Many important papers and chapters published in recent years talk about the
modeling, stability and tracking in the two wheeled robot among them we find: A
transformable wheel-legged mobile robot: Design, analysis and experiment (TaoSun
et al. 2017), Mechanical Design and Dynamic Modeling of a Two-Wheeled Inverted
Pendulum Mobile Robot (Jingtao et al. 2007), Posture self-stabilizer of a biped
robot based on training platform and reinforcement learning (Weiguo and Liyang
2017), modelling and control of two-wheeled robots (Ronald et al. 2013), A robust
adaptive fuzzy variable structure tracking control for the wheeled mobile robot:
Simulation and experimental results (Mauricio et al. 2017), MIMO Fuzzy Control
for Autonomous Mobile Robot (Mac et al. 2016), Experimental Validation of an
Under-actuated Two-Wheeled Mobile Robot (Oryschuk et al. 2009), Intelligent
block spectral factors relocation in a quadrotor unmanned aerial vehicle (Bekhiti
et al. 2017), Development of a Self-Balancing Human Transportation Vehicle for
the Teaching of Feedback Control (Lin and Tsai 2009), Adaptive Neural Network
Control of a Self-Balancing Two Wheeled Scooter (Tsai et al. 2010), Modeling and
LQR Control of a Multi-DOF Two-wheeled Robot (Shigong et al. 2014), Wireless
Controlled Two Wheel Balancing Robot (Yong and Foong 2011), Autonomous Dual
Wheel Self Balancing Robot Based on Microcontroller (Adeel et al. 2013), Two-
wheeled self balanced pendulum workspace improvement via under-actuated robust
nonlinear control (Guilherme et al. 2015), Adaptive neural network control of a self-
balancing two-wheeled scooter (Tsai et al. 2010), Motion control for a two-wheeled
vehicle using a self-tuning PID controller (Ren et al. 2008), Maneuverability and
path following control of wheeled mobile robot in the presence of wheel skidding
and slipping (Low and Wang 2010), Equiangular navigation and guidance of a
wheeled mobile robot based on range-only measurements (Teimoori and Savkin
2010), Bio-inspired algorithm for wheeled robot’s navigation (Kherici and Ali
2011), Design and parametric control of co-axes driven two-wheeled balancing
robot (Memarbashi and Chang 2011) and others researches, which are indicate the
importance of this robot.
240 B. Nail et al.

The current concern of robotics is to realize robots capable of reacting fast

without impairing desired performance, stability and precision. One of the first
solutions to the problem of control, in the history of robotics, consisted in using
linear PID-type fixed-gain drives.
The advantage of such a method is its simplicity of implementation and low
installation cost. It has given good performance because they were slow movements
and actuators with a high reduction ratio. The effects of the load (robot arm with the
mass of the load) on the motors are practically negligible, in comparison with the
inertia of the engines, since they are divided by a high reduction ratio.
In this chapter a combination between the stability and the tracking are suggested
to control the two-wheeled robot system, and the main contributions in this chapter
is organized as the following steps:
The Sect. 11.1 an introduction, the Sect. 11.2 the problem statement the difficul-
ties in the model behavior. . . etc., in the Sect. 11.3 presented the nonlinear dynamics
model of the two-wheeled robot, and the linearized model based on their equilibrium
point using the Taylor series expansion (Jingtao et al. 2007; Shigong et al. 2014),
and the discretizing (digitalizing) the obtained continuous linearized model (Ogata
1994), the Sect. 11.4 checking the block controllability of the dynamic model (Chen
1984; Kailath and Li 1980; Nail et al. 2016b), and transforming the system to
block controllable form (Chen 1984; Kailath and Li 1980; Nail et al. 2016b), in
the Sect. 11.5 the stabilization by assigning block poles choosing diagonally, where
calculating the feedback matrix gain K based on MFD (Bekhiti et al. 2017) using
the Vandermonde matrix (Bekhiti et al. 2017), in Sect. 11.6. The tuning of the
two digital PID controllers via Ziegler-Nichols method rules (Ogata 1994, 2010)
in order to obtained best set-points tracking, and finally in Sect. 11.7 a conclusion
summarized the advantages and the limitations of proposed controller.

11.2 Problem Statement

The problem arise in this work is to achieve and ensure best control to two-wheeled
robot with minimal cost and with simple mathematical tools, by using the matrix
fraction description theory which is used generally in the linear systems, on the
other hand the robot model is a non-linear behavior, also the choice of the best block
poles in which forms gives the best performances, also the interactions between the
inputs/outputs of the multivariable model which are influenced negatively in the
selection of the two PID controllers parameters.

11.3 Model of Two-Wheeled Self-Balancing Robot System

In this chapter we have used the final model of the two-wheeled inverted pendulum
and the linear model for a DC motor, where the model behavior is multivariable
11 Digital Stabilizing and Control for Two-Wheeled Robot 241

unstable dynamics, with a very high non-linearity and coupling, for more detail see
Jingtao et al. (2007) and Shigong et al. (2014).
The nonlinear dynamics model for a two wheeled inverted pendulum and the
body is:
⎧  α
⎪
⎪ αβg sin θ − β + cos θ Tm − βTr − α 2 θ̇ 2 sin θ cos θ
⎪
⎪ r
⎪
⎪ θ̈ =
⎨ βγ − α 2 cos2 θ
 (11.1)
⎪
⎪ γ
⎪
⎪ θ̇ 2 αγ sin θ − α 2 g sin θ cos θ + Tm α cos θ + + Tr α cos θ
⎪
⎪
⎩ ẍ = r
βγ − α 2 cos2 θ

2Iw
where: α = Mp l, β = 2Mw + + Mp , γ = Ip + Mp l 2
r2

11.3.1 Linearization of the Plant

The system plant has to be linearized about the equilibrium point, which in the case
of inverted pendulum is the upright position θ = 0◦ . Hence the system is only valid
in the region of operation about this equilibrium point. This linearization is done by
assuming θ as a small angle in the Taylor series expansion.
 2
dθ
Therefore: sin θ ≈ θ , cos θ ≈ 1 and ≈0
dt
the resulting equations are given below:
⎧  α
⎪
⎪ αβgθ − β+ Tm − βTr
⎪
⎪ r
⎪
⎪ θ̈ =
⎨ βγ − α 2
 (11.2)
⎪
⎪
⎪
⎪ −α 2 gθ + α + γ Tm + αTr
⎪
⎪
⎩ ẍ = r
βγ − α 2

km ke km
T =− w+ Va (11.3)
R R
the relationship between the input voltage to the motor and the output torque is given
by Eq. (11.3) and knowing that the motors have the same specifications leads to:
⎧
⎨ θ̈ = a1 θ + a2 w + b1 VDw + b2 VRw
(11.4)
⎩
ẍ = a3 θ + a4 w + b3 VDw + b4 VRw
242 B. Nail et al.

Where:
αβg  α
a1 = , km ke 2β +
βγ − α 2 a2 = r
 − α )γ
α2g R(βγ 2
a3 = − ,
 − αα
βγ 2 km ke 2α +
a4 =− r
km β + R(βγ − α 2 )
b1 = − r , βkm
 −γα ) = −
R(βγ 2 b2
R(βγ − α 2 )
km α + αkm
b3 = r , b4 =
R(βγ − α 2 ) R(βγ − α 2 )

. km ke km
For the reaction wheel: M = Tr = IR ẇ. Therefore: ẇ = − w+ VRw .
IR R IR R
Thus from Eq. (11.4) the state space model is represented by a dynamic and
control matrices A and B respectively, where its formulas are given as follows:
⎛  γ ⎞ ⎛  γ ⎞
2 km ke 2α + km α +
⎜0 − α g 0− r ⎟ ⎜ r αk m ⎟
⎜ ⎟ ⎜ ⎟
⎜ βγ − α 2 R(βγ − α 2 ) ⎟ ⎜ R(βγ − α 2 ) R(βγ − α 2 ) ⎟
⎜ ⎟ ⎜ ⎟
⎜ ⎟ ⎜ ⎟
⎜ ⎟ ⎜ ⎟
⎜0 0 1 0 ⎟ ⎜ 0 0 ⎟
⎜ ⎟ ⎜ ⎟
⎜ ⎟ ⎜ ⎟
⎜  α ⎟,⎜  α ⎟
⎜ km ke 2β + ⎟ ⎜ k ⎟
⎜ αβg ⎟ ⎜ m β+ βkm ⎟
⎜0 r ⎟ ⎜− r − ⎟
⎜ βγ − α 2 0
R(βγ − α 2 ) ⎟ ⎜ R(βγ − α 2 ) R(βγ − α 2 ) ⎟
⎜ ⎟ ⎜ ⎟
⎜ ⎟ ⎜ ⎟
⎜ ⎟ ⎜ ⎟
⎝ km ke ⎠ ⎝ km ⎠
0 0 0 − 0
IR R IR R

where: x1 = ẋ is the linear speed, x2 = y1 = θ is the tilt angle of the body, x3 = θ̇

is the angular velocity of the body and x4 = y2 = w is the rotational speed of the
reaction wheel.
The robot state space equations should be transformed to the discrete time for the
digital implementation control:

X(k + 1) = Ad X(k) + Bd u(k)

(11.5)
Y (k) = Cd X(k) + Dd u(k)

where

X ∈ R n , Y ∈ R p , u ∈ R m , Ad ∈ R n×n , Bd ∈ R n×m , Cd ∈ R p×n , and Dd ∈ R p×m .

The conversion will be done by the following formulas based on Ogata (1994):
11 Digital Stabilizing and Control for Two-Wheeled Robot 243

, ts
Ad = e A.ts
≈ I + A.ts , Bd = eA.δ Bdδ, Cd = C. (11.6)
0

We assume that our sampling time ts is very short compared with mechanical
dynamics of the system.

11.4 Transformation to Block Canonical Forms

The system presented in Eq. (11.5) is block controllable of index l if:

• Ωc = [Bd , Ad Bd , Ad 2 Bd , . . . , Ad l−1 Bd ] has full rank
n
• l= is an integer
m
If both conditions are satisfied, then the change of coordinates Xc (t) = Tc X(t)
transforms the system into the following Block controller form:

Ẋc (t) = Ac Xc (t) + Bc u(t)

(11.7)
Y (t) = Cc Xc (t) + Dc u(t)

where
⎧
⎪
⎪ Ac = Tc ATc −1
⎪
⎪ ⎛ ⎞
⎪
⎪ Om Im · · · Om
⎛ ⎞ ⎪
⎪
⎪
⎪ ⎜ Om Om · · · Om ⎟
Tc1 ⎪
⎪ ⎜ ⎟
⎜ ⎟ ⎪
⎨ ⎜ ⎟
⎜ Tc1 A ⎟ −1 = ⎜ ... ..
⎟
Tc = ⎜ .. ⎟ , Tc1 = Bc Ωc ,
T
⎜ . . . . O m ⎟
⎝ . ⎠ ⎪
⎪ ⎝ Om Om . . . Im ⎠
⎪
⎪
Tc1 Al−1 ⎪
⎪
⎪
⎪ −Al −Al−1 · · · −A1
⎪
⎪  T
⎪
⎪ B = T c = Om Om · · · Im
B
⎪
⎩
c
 
Cc = CTc −1 = Cl Cl−1 · · · C1

The system described by general state space Eq. (11.5) with n = 4 : state number,
m = 2 : input number, and p = 2 : output number satisfy the condition n/m = l =
2 = integer and controllable system then can be transformed to block controller
form using the following similarity transformation xc = Tc x Where:
⎛ ⎞
Tc1 ! !−1
Tc = ⎝ ⎠ , Tc1 = Om Im Bd Ad Bd (11.8)
Tc1 Ad

For more detail we orient the reader to see Nail et al. (2018a). Referring to Nail
et al. (2018b), we will end up with the following model
244 B. Nail et al.

⎧
⎪
⎪ Ac = Tc Ad Tc −1 , Bc = Tc Bd , Cc = Cd Tc −1
⎪
⎪ ⎛ ⎞
⎪
⎪ 0 0 1 0
⎪
⎪ ⎜
⎪
⎪ ⎜ 0 0 0 1⎟⎟
⎪
⎨ Ac = ⎝ 0.1289 0.5944 0.6803 −1.0812 ⎠    
00 10
, Om = , Im =
⎪
⎪ 0.5865 0.0576 −0.5911 0.9884 00 01
⎪
⎪  T
⎪
⎪ Bc = Om Im
⎪
⎪  
⎪
⎪ 0.2637 −0.0013 0.0025 0.5102
⎪
⎩ Cc =
−0.1070 0.0524 0.0802 −0.1209

The right matrix fraction description (RMFD) form is:

y(q −1 ) = N(q −1 )D −1 (q −1 )u(q −1 ) (11.9)

The left matrix fraction description (LMFD) form is:

y(q −1 ) = A−1 (q −1 )B(q −1 )u(q −1 ) (11.10)

Where

D(q −1 ) = I + D1 q −1 + D2 q −2

N(q −1 ) = N0 + N1 q −1 + N2 q −2

A(q −1 ) = I + A1 q −1 + A2 q −2

B(q −1 ) = B0 + B1 q −1 + B2 q −2

The coefficient of D(q −1 ), N (q −1 ) are well known and directly obtained from
Ac , Cc see Nail et al. (2016a,b, 2017):
   
−0.6803 1.0812 −0.1289 −0.5944
D1 = , D2 =
0.5911 −0.9884 −0.5865 −0.0576
   
0.0025 0.5102 0.2637 −0.0013
N0 = O2×2 , N1 = , N2 =
0.0802 −0.1209 −0.1070 0.0524

From Eqs. (11.9) and (11.10) we can write:

A(q −1 )N (q −1 ) = B(q −1 )D(q −1 ) (11.11)

Expanding the last equation we can write in more compact form:

11 Digital Stabilizing and Control for Two-Wheeled Robot 245

⎛ ⎞ ⎛ ⎞−1 ⎛ ⎞
B1T I2 O2 O2 O2 N1T
⎜ ⎟ ⎜ ⎟ ⎜ ⎟
⎜ ⎟ ⎜ ⎟ ⎜ ⎟
⎜ BT ⎟ ⎜ D T I −N T O ⎟ ⎜ NT ⎟
⎜ 2 ⎟ ⎜ 1 2 2 ⎟ ⎜ 2 ⎟
⎜ ⎟ ⎜ 1
⎟ ⎜ ⎟
⎜ ⎟=⎜ ⎟ ⎜ ⎟ (11.12)
⎜ T ⎟ ⎜ T T ⎟ ⎜ ⎟
⎜ A1 ⎟ ⎜ D2 D1 −N2T −N1T ⎟ ⎜ O2 ⎟
⎜ ⎟ ⎜ ⎟ ⎜ ⎟
⎝ ⎠ ⎝ ⎠ ⎝ ⎠
AT2 O2 D2T O2 −N2T O2

Solving this last equation yield to:

   
−0.9939 1.1919 0.1146 −0.4768
A1 = , A2 =
0.7845 −0.6749 −0.7148 −0.0035
   
0.0025 0.5102 0.0569 −0.1510
B0 = O2×2 , B1 = , B2 =
0.0802 −0.1209 −0.0330 0.3280

11.5 Dynamics Stabilization Using Block Poles Placement

This algorithm is an extension of classical pole placement using the theory of matrix
fraction description, in this application we reinstruct to give the necessary steps and
for more detail see our contribution (Nail et al. 2016a, 2017).
 Right block poles
Right block poles chosen diagonally based on selected eigenvalues:
   
−.5 0 −.35 0
R1 = , R2 =
0 − .2 0 − .7

 Right block Vandermonde matrix

⎛ ⎞
1 0 1 0
  ⎜ ⎟
I I ⎜ 0 1 0 1⎟
⎜ ⎟
VR = =⎜ ⎟
R1 1 R 2 1 ⎜ −0.5 0 −0.35 0⎟
⎝ ⎠
0 −0.2 0 −.7

 Constructing from a complete set of right block poles

Consider a complete set of right poles {R1 , R2 } for the desired matrix polynomial
Ds (λ), If Ri is a right poles of Ds (λ) so:
246 B. Nail et al.

Ri 2 + Ds1 Ri 1 + D2 = O2 (11.13)

Ds1 Ri 1 + Ds2 = −Ri 2 (11.14)

Replacing i from 1 to 2, we get the following:

* +
[Dsd2 , Dsd1 ] = − R12 , R22 VR −1

Where VR is the right block Vandermonde matrix.

 State Feedback matrix gain
Consider the linear discret-time invariant dynamic system described by Eq. (11.21).
Applying the state feedback u = −Kx(t) to this system, where K is a 2 × 4 gain
matrix, after using the block controller form transformation for the system, we get
u = −Kc xc (t).
!
K = Kc Tc = Kc(2) , Kc(1) (11.15)

Tc and Kci ∈ R m×m For i = 1, 2, the closed loop system is shown below:

ẋc = (Ac − Bc Kc ) xc
(11.16)
yc = Cc xc

Where
( )
Om Im
(Ac − Bc Kc ) = (11.17)
− (A2 + Kc2 ) − (A1 + Kc1 )

The characteristic matrix polynomial of this closed loop system is:

D (λ) = Im λ2 + (A1 + Kc1 ) λ1 + (A2 + Kc2 ) (11.18)

From a set of desired eigenvalues, we construct the set of Block poles, the desired
characteristic matrix polynomial of the block poles in the form, by putting

Dsd (λ) = D(λ) (11.19)

We get the coefficients Kci = Dsdi − Ai as follows:

Kci = Dsdi − Ai for i = 1, 2 (11.20)

11 Digital Stabilizing and Control for Two-Wheeled Robot 247

The gain matrix given by the following formula:

K = Kc T c (11.21)

Finally and after numerical application the feedback static matrix gain is given as
follows:
 
0.3039 0.5944 1.5303 −1.0812
K=
0.5865 0.1976 −0.5911 1.8884

11.6 Design Digital Controller for the Tracking

How to tuning the two PID Control of Plant. If a mathematical model of the plant
can be derived, then it is possible to apply various design techniques for determining
parameters of the controller that will meet the transient and steady-state specifica-
tions of the closed-loop system. However, if the plant is so complicated that its
mathematical model cannot be easily obtained, then an analytical or computational
approach to the design of a PID controller is not possible. Then we must resort to
experimental approaches to the tuning of PID controllers. The process of selecting
the controller parameters to meet given performance specifications is known as
controller tuning. Ziegler and Nichols suggested rules for tuning PID controllers
(meaning to set values kp , ki and kd ) based on experimental step responses or based
on the value of kp that results in marginal stability when only proportional control
action is used. Ziegler-Nichols rules, which are briefly presented in the following,
are useful when mathematical models of plants are not known. Such rules suggest
a set of values of kp , ki and kd that will give a stable operation of the system.
However, the resulting system may exhibit a large maximum overshoot in the step
response, which is unacceptable. In such a case we need series of fine tunings until
an acceptable result is obtained. In fact, the Ziegler-Nichols tuning rules give an
educated guess for the parameter values and provide a starting point for fine tuning,
rather than giving the final settings for kp , ki and kd in a single shot (Ogata 1994,
2010).
Ziegler-Nichols method Rules is used for Tuning the two PID Controllers shown
in the schematic diagram in Fig. 11.1 we determining the values of the proportional
gain kp integral time ki and derivative time kd of the two PID based on the
transient response characteristics of a given linearized plant. Such determination
of the parameters of PID controllers or tuning of PID controllers can be made
by engineers on-site by experiments on the plant (Ogata 1994, 2010). Figure 11.1
shown the schematic bloc diagram of the two digital PID controllers connected with
the stabilized linearized dynamics plant by the feedback matrix gain. Table 11.1
shown the parameters of the two PID controllers.
248 B. Nail et al.

Ref 1 u1 y1
P ID1

Ref 2 u2 y2
P ID2
−K

Fig. 11.1 Schematic bloc diagram of the proposed controller

Table 11.1 Digital PID Kp Ki Kd

controllers parameters
PID controller 1 0.95 200.00 0.001
PID controller 2 −1.38 −276.61 0.100

11.6.1 Simulation Results and Interpretations

From the obtained results as shown in the next figures, we see that the angular
speed of the motor tracks its reference trajectory in Fig. 11.2 with very small error
in Fig. 11.3, no overshooting, no static error is obtained at both transient and steady
state regimes, the convergence is ensured based on the smooth error and tracking.
In Fig. 11.5, the real and the desired tilt angle signals are coincided in all time with
negligible static error in Fig. 11.6. From the Figs. 11.4 and 11.7, the control signals
voltages not exceed the norm voltage of the two DC motors, which are drive the
two-wheeled robot.

11.7 Conclusion

In this chapter digital controller is proposed and implemented in two-wheeled robot

system in order to control two inputs: angular speed and tilt angle, based on MFD
theory, which its effect has been clear in the stabilization of the behavior and the
dynamics of the two-wheeled robots, and also by the feedback matrix gain K which
is designed by the assigning of two robust stable block poles in diagonal form,
and for the tracking, two digital PID controllers are tuning using Ziegler-Nichols
method to acquired the best parameters, the obtained results are well satisfactory
11 Digital Stabilizing and Control for Two-Wheeled Robot 249

1
Ref
Real
0.8
Angular speed (rad/s)

0.6

0.4

0.2

0
0 2 4 6 8 10
Time (s)

Fig. 11.2 The signals of the real and the desired rotational speed

0.06

0.04
Error angular speed (rad/s)

0.02

-0.02

-0.04

-0.06
0 2 4 6 8 10
Time (s)

Fig. 11.3 The error between the real and the desired rotational speed
250 B. Nail et al.

10

5
Control u 1 (volte)

-5

-10
0 2 4 6 8 10
Time (s)

Fig. 11.4 The input signal control u1

1.5
Ref
Real
1

0.5
θ (rad)

-0.5

-1

-1.5
0 2 4 6 8 10
Time (s)

Fig. 11.5 The signals of the real and the desired tilt angle
11 Digital Stabilizing and Control for Two-Wheeled Robot 251

0.15

0.1

0.05
Error θ (rad)

-0.05

-0.1

-0.15
0 2 4 6 8 10
Time (s)

Fig. 11.6 The error between the real and the desired tilt angle

-2
Control u2 (volte)

-4

-6

-8

-10
0 2 4 6 8 10
Time (s)

Fig. 11.7 The input signal control u2

252 B. Nail et al.

and the proposed controller schema and led his role effectively and reliability, as
further works we suggest the follows:
• For the stabilization adaptive block poles is more suitable in the presence of
incertitude in the parameters.
• For the tracking fractional order PID is with self tuning is more robust in the
presence of incertitude in the parameters..

References

Adeel, U., Alimgeer, K. S., Inam, O., Hameed, A., Qureshi, M., & Ashraf, M. (2013). Autonomous
dual wheel self balancing robot based on microcontroller. Journal of Basic and Applied
Scientific Research, 3, 843–848.
Bekhiti, B., Dahimene, A., Hariche, K., & Nail, B. (2017). Intelligent block spectral factors
relocation in a quadrotor unmanned aerial vehicle. International Journal of Systems, Control
and Communications, 8, 303–334.
Chen, C. (1984). Linear system theory and design. New York: Holt, Reinhart and Winston.
Guilherme, R., Manuel, O., Vicente, M., & Francisco, R. (2015). Two-wheeled self-balanced
pendulum workspace improvement via underactuated robust nonlinear control. Control
Engineering Practice, 44, 231–242.
Jingtao, L., Xueshan, G., & Qiang, H. (2007). Mechanical design and dynamic modeling of a two-
wheeled inverted pendulum mobile robot. In IEEE International Conference on Automation
and Logistics, Jinan, China (pp. 1614–1619).
Kailath, T., & Li, W. (1980). Linear systems. Englewood Cliffs: Prentice Hall.
Kherici, N., & Ali, Y. M. B. (2011). Bio-inspired algorithm for wheeled robot’s navigation.
International Journal of Artificial Intelligence and Soft Computing, 2, 353–366.
Lin, S.-C., & Tsai, C.-C. (2009). Development of a self-balancing human transportation vehicle
for the teaching of feedback control. IEEE Transactions on Education, 52, 157–168.
Low, C. B., & Wang, D. (2010). Maneuverability and path following control of wheeled mobile
robot in the presence of wheel skidding and slipping. Journal of Field Robotics, 27, 127–144.
Mac, T., Copot, C., Keyser, R., Tran, T., & Vu, T. (2016). Mimo Fuzzy control for autonomous
mobile robot. Journal of Automation and Control Engineering, 4, 65–70.
Mauricio, B., Douglas, B., & Nardnio, M. (2017). A robust adaptive fuzzy variable structure
tracking control for the wheeled mobile robot: Simulation and experimental results. Control
Engineering Practice, 64, 27–43.
Memarbashi, H. R., & Chang, J.-Y. (2011). Design and parametric control of co-axes driven two-
wheeled balancing robot. Microsystem Technologies, 17, 1215–1224.
Nail, B., Kouzou, A., Hafaifa, A., & Djeddi, C. (2016a). Parametric identification of multivariable
industrial system using left matrix fraction description. Journal Automation & Systems
Engineering, 10, 221–230.
Nail, B., Kouzou, A., Hafaifa, A., Kamel, H., & Bekhiti, B. (2016b). Parametric output feedback
stabilization in mimo systems: Application to gas turbine power plant. In 8th International
Conference on Modelling, Identification and Control (ICMIC), Medea, Algeria (pp. 971–976).
Nail, B., Kouzou, A., Hafaifa, A., & Chaibet, A. (2017). Power plant system identification based
on extended least square kronecker operator. Journal of the Technical University-Sofia Plovdiv
Branch, Bulgaria Fundamental Sciences and Applications, 23, 15–20.
Nail, B., Kouzou, A., & Hafaifa, A. (2018a). Robust block roots assignment in linear discrete-
time sliding mode control for a class of multivariable system: Gas turbine power plant
application. Transactions of the Institute of Measurement and Control. https://doi.org/10.1177/
0142331218774615.
11 Digital Stabilizing and Control for Two-Wheeled Robot 253

Nail, B., Kouzou, A., Hafaifa, A., & Chaibet, A. (2018b). Parametric identification and stabilization
of turbo-compressor plant based on matrix fraction description using experimental data.
Journal of Engineering Science & Technology, 13(6), 1850–1868.
Ogata, K. (1994). Discrete-time control systems. Englewood Cliffs/Upper Saddle River: Prentice-
Hall.
Ogata, K. (2010). Modern control engineering (Prentice-Hall Electrical Engineering Series).
Upper Saddle River: Instrumentation and Controls Series.
Oryschuk, P., Salerno, A., Al-Husseini, A. M., & Angeles, J. (2009). Experimental validation of
an underactuated two-wheeled mobile robot. IEEE/ASME Transactions on Mechatronics, 14,
252–257.
Ren, T.-J., Chen, T.-C., & Chen, C.-J. (2008). Motion control for a two-wheeled vehicle using a
self-tuning pid controller. Control Engineering Practice, 16, 365–375.
Ronald, C., Karl, S., & Roger, H. (2013). Review of modelling and control of two-wheeled robots.
Annual Reviews in Control, 37, 89–103.
Shigong, J., Fuquan, D., Ling, L., & Xueshan, G. (2014). IEEE International Conference on
Robotics and Biomimetics Robio Modeling and LQR Control of a Multi-dof Two-Wheeled
Robot (pp. 1964–1969).
Sun, T., Xiang, X., Su, W., Wu, H., & Song, Y. (2017). A transformable wheel-legged mobile
robot: Design, analysis and experiment. Robotics and Autonomous Systems, 98, 30–41.
Teimoori, H., & Savkin, A. V. (2010). Equiangular navigation and guidance of a wheeled mobile
robot based on range-only measurements. Robotics and Autonomous Systems, 58, 203–215.
Tsai, C.-C., Huang, H.-C., & Lin, S.-C. (2010). Adaptive neural network control of a self-balancing
two-wheeled scooter. IEEE Transctions on Industrial Electronics, 57, 1420–1428.
Weiguo, W., & Liyang, G. (2017). Posture self-stabilizer of a biped robot based on training
platform and reinforcement learning. Robotics and Autonomous Systems, 98, 42–55.
Yong, C., & Foong, C. (2011). Wireless controlled two wheel balancing robot. International
Journal Network and Mobile Technologies, 2, 88–98.
Chapter 12
Mobile Robot Path Planning Based
on Optimized Fuzzy Logic Controllers

L. Cherroun, M. Boumehraz, and A. Kouzou

Abstract The path planning task is defined as the process to compute the motion
sequence allowing the robot to move from the start position to the final destination
autonomously without human actions. The path planning is one of the popular tasks
encountered by imprecision and uncertainties and it has been studied using fuzzy
logic systems (FLS). The construction of a well performing fuzzy controller is
not always easy. The problem of finding appropriate membership functions and
fuzzy rules is a difficult task. However, the design of fuzzy rules is often reliant
on heuristic experience and it lacks systematic methodology, therefore these rules
might not be correct and consistent. The design can prove to be long and delicate
due to the important number of parameters to determine, and can lead then to a
solution with poor performance. To cope with this difficulty, many researchers have
been working to find and apply learning algorithms for fuzzy controller design.
These automatic methods enable to extract information when the knowledge is
not available. The most popular approach to optimize fuzzy logic controllers may
be a kind of supervised learning where the training data is available. However,
in real applications, extraction of training data is not always easy and become
impossible when the cost to obtain training data is expensive. For these problems,
reinforcement learning (RL) is more suitable than supervised learning. A control
strategy with a learning capacity can be carried out by using Q-learning for
tuning fuzzy logic controllers; which the robot receives only a scalar signal likes
a feedback. This information makes to adjust the robot behavior in order to improve
their performances. The basic idea in Q-learning algorithm of RL is to maximize
the received rewards after each interaction with the environment. In this chapter,

L. Cherroun () · A. Kouzou

Laboratory of Applied Automatic and Industrial Diagnostics (LAADI), University of Djelfa,
Djelfa, Algeria
e-mail: [email protected]; [email protected]
M. Boumehraz
LMSE Laboratory, University of Biskra, Biskra, Algeria
e-mail: [email protected]

© Springer Nature Singapore Pte Ltd. 2019 255

N. Derbel et al. (eds.), New Developments and Advances
in Robot Control, Studies in Systems, Decision and Control 175,
https://doi.org/10.1007/978-981-13-2212-9_12
256 L. Cherroun et al.

Q-learning algorithm is used to optimize Takagi-Sugeno fuzzy logic controllers for

autonomous path planning of a mobile robot. These optimized fuzzy controllers
are used for the different robot tasks: goal seeking, obstacle avoidance and wall-
following. The obtained results of this optimization method present significant
improvements of the robot behaviors.

Keywords Mobile robot · Behavior · Fuzzy controller · Optimization ·

Q-learning · Fuzzy Q-learning

12.1 Introduction

The existence of robots in various types became very significant in the industrial
and service sectors. Due to the growing interest of the service robots, they can
achieve their tasks in environments contains multiple obstacles. Autonomous mobile
robot navigation and path planning are very important issues in the field of
unmanned robot control. The ability to autonomous navigation is one of the standing
challenging aspects in mobile robotic field. It is a difficult task, which requiring
a complete modeling of the environment and intelligent controllers (Cuesta and
Ollero 2005). Autonomous path planning is considered as the task of determining
a collision-free path that enables the robot to move through an obstacle course,
starting from an initial position to reach a goal position in unknown spaces.
Path planning must respect the constraints kinematics of the robot without human
intervention. It can be classified into two categories: global path planning and local
path planning (Khatib 1986; Latombe 1991). Global path planning methods are
usually conducted off-line in a completely known environments (Cuesta and Ollero
2005; Latombe 1991), where an exact environment model is required for planning
the path. Although these approaches have an exact solution, their time complexity
grows with the geometry complexity and grows exponentially with the number of
degrees of freedom in the vehicle’s motion (Latombe 1991). This fact has led to
the emergence of numerous heuristic or approximating approaches, which rely on
either calculating the potential Fields (Borestein and Koren 1989) or performing a
search through a state space model (Cuesta and Ollero 2005). As a pre-specified
environment is required for these methods to plan the path, they fail when the
environment is not fully known.
On the other hand, the local path planning techniques, also known as the obstacle
avoidance methods, are more efficient in autonomous mobile robot navigation in
unknown or partially known environment (Algabri et al. 2015; Khatib 1986; Seraji
and Howard 2002; Wang and Liu 2008). These reactive methods use the on-line
sensory information to tackle the uncertainty.
Fuzzy logic systems (FLSs) are the robust approaches used for autonomous
mobile robot navigation and path planning (Antonelli and Fusco 1986; Seraji and
Howard 2002; Wang and Liu 2008; Yang et al. 2005). These systems have been
applied in many engineering problems, especially in control systems and robotic
12 Mobile Robot Path Planning Based on Optimized Fuzzy Logic Controllers 257

(Abadi and Khooban 2015; Al Yahmedi and Fatmi 2011; Cherroun and Boumehraz
2013a; Vadakkepa et al. 2004). The modeling process is based on the behavior of the
operator and their knowledge to handle uncertainties. However the performances of
FLSs depend on the formulation of the fuzzy rules and the numerical specification
of all linguistic terms. The conception of a powerful FLS is not always easy. Each
design process can prove to be long and difficult due to the number of parameters
of membership functions of inputs-outputs, the fuzzy rules, consequent parameters
(Cherroun and Boumehraz 2013b; Jallouli et al. 2010; Passino and Yurkovich 1998;
Timothy 2004).
Tuning and adjusting parameters of FLC are interesting issues that are studied
by many researchers. The most applied are based on learning algorithms and
optimization method as the use of Neural Networks (NN) in (Algabri et al. 2014a;
Bakdi et al. 2016; Pothal and Parhi 2015). Using Genetic Algorithms (GA) as in
Begum et al. (2008) for simultaneous localization and mapping of mobile robots, or
for robot navigation as the paper (Zhao et al. 2015). FLC Optimization is too done by
using particle swarm optimization algorithms (PSO) as in Algabri et al. (2014a), and
by applying Ant Colony optimization (ACO) in Hsu et al. (2013). The optimization
task can be done using hybrid approaches between the previous methods. Other soft
computing techniques for mobile robot navigation are presented in Algabri et al.
(2015).
Reinforcement learning (RL) (Kaelbling et al. 1996; Sutton and Barto 1998) is
an automatic tuning method used by some authors to optimize FLC parameters:
premisse and conclusion parts (Berenji and Khedkar 1992; Jouffle 1998; Nemra and
Rezine 2008). This hybrid system is attractive in the field of robotic as an optimized
fuzzy navigators for the different robot behaviors as in Joo and Deng (2005); Juang
and Hsu (2009), and Cherroun and Boumehraz (2013b). Or by combination with
neural networks for robot obstacle avoidance as presented in the paper (Mihai and
Gheorghe 2016).
These automatic methods enable to extract information when the experts’ prior
knowledge is not available.
A control strategy with a learning capacity can be applied using the reinforce-
ment learning and fuzzy logic approaches; which the robot receives only a scalar
signal likes a feedback. This reinforcement makes it possible the navigator to
adjust its strategy in order to improve their performances (Glorennec and Jouffle
1997; Kaelbling et al. 1996; Sutton and Barto 1998). RL is considered as an
automatic modification of the robot behavior in its working environment (Cherroun
and Boumehraz 2013b; Nemra and Rezine 2008). The reinforcement learning is a
method of optimal control, when the agent starts from an ineffective solution which
gradually improves according to the experience gained to solve a sequential decision
problem. The basic idea in reinforcement learning is that, the agent is placed in
its environment and it can observe the results of their actions. The objective of
the agent is to maximize the received rewards (Sutton and Barto 1998). To use
reinforcement learning, several approaches are possible (Jouffle 1998). The first
consists in discrediting the problem using Q-tables, and the second method is based
on working at continuous spaces by the estimation of the quality functions.
258 L. Cherroun et al.

In this chapter, reinforcement learning is used to optimize fuzzy logic controllers

(FLCs) for mobile robot path planning. The obtained fuzzy navigators are able to
perform a successful autonomous path planning.
The present chapter is organized as follows: Sect. 12.2 gives the necessary
background of Fuzzy Logic Systems (FLS). Reinforcement learning algorithms are
presented in Sect. 12.3 by explaining Q-learning and the optimization algorithm of
fuzzy systems. Section 12.4 discusses the application of optimized fuzzy controllers
for the different tasks of a mobile robot with the obtained results. Section 12.5
presents conclusions and future research directions of the work presented in this
chapter.

12.2 Fuzzy Logic System

Fuzzy logic system is defined as a mathematical function used to imitate the capacity
of reasoning and decision making from the human operator, using uncertainties
and/or unclear information (Passino and Yurkovich 1998). In fuzzy logic applica-
tions, the model of a system is ensured by linguistic rules (fuzzy rules) between the
input and output variables. A fuzzy logic system consists basically of four blocks:
fuzzification, rule-base, inference and defuzzification (Passino and Yurkovich 1998;
Timothy 2004). In Fig. 12.1, we present the basic structure of a FLS. We can define
each block as follows.

12.2.1 Fuzzification

This part converts the real inputs (X) into information easily used to activate and
apply fuzzy rules. The fuzzy sets of all inputs must be defined. This interface maps
inputs of the associated fuzzy sets in order to calculate the membership degree of
each input variable μA (x).

Rule-Base

Input X Inference Output Y

Fuzzification Mechanism Defuzzification

Fig. 12.1 Structure of fuzzy logic system

12 Mobile Robot Path Planning Based on Optimized Fuzzy Logic Controllers 259

12.2.2 Rule-Base

The linguistic description of the expert to quantify the input-output relationship is

done by a set of “If-Then” rules. For a fuzzy logic system with p inputs (x1 ∈
X1 , . . . xp ∈ Xp ) and one output (y ∈ Y ). Using M fuzzy rules, the rth rule for
r = 1 . . . M has the following forms for a FLS:

If (x1 is F1r ) and (x2 isF2r ) . . . and (xp is Fpr ) Then (y is Gr ) (12.1)

The fuzzy rule-base of the robot controller can be described as follows:

R1 : If (goal is Far) Then (speed is Big) and (steering is zeros)

R2 : If (goal is right) Then (speed is Medium) and (steering is right)
:
Rn : If (goal is near) Then (speed is zeros) and (steering is zeros)

12.2.3 Inference Mechanism

The role of this block is to interpret and apply the operator knowledge to control the
system based on If-Then fuzzy rules. The inference mechanism combines the input
fuzzy sets and output fuzzy sets by compositions of the relationships.
For a FLS with (x1 ∈ X1 , . . . xp ∈ Xp ) inputs and one output (y ∈ Y ) using M
rules as mentioned in Eq. (12.2). The firing strength of the ith rule is calculated by:

F i (x) = μF i (x1 )  μF i (x2 )  . . .  μFpi (xp ) (12.2)

1 2

12.2.4 Defuzzification

converts the results of the inference mechanism into real-applied control values (Y ).
The output is calculated by Eq. (12.3) using the centroid method:

.
M
F i (x)y i
i=1
y= (12.3)
.M
F i (x)
i=1

In fuzzy logic applications, the two most popular fuzzy logic systems used in control
fields are the Mamdani and Takagi-Sugeno (TS) models (Passino and Yurkovich
260 L. Cherroun et al.

1998; Timothy 2004). The difference between TS fuzzy logic controller and a
Mamdani fuzzy logic consist essentially of the definition of outputs (consequent
part of rules).
In this chapter, we used optimized Takagi-Sugeno Fuzzy Logic Controllers
(TSFLC) for the design of mobile robot behaviors. All inputs and outputs variables
(angle, distances, speed and steering. . . ) have degree of membership functions. Each
functionality is designed using an optimized zero order TSFLC.

12.3 Reinforcement Learning (RL)

Reinforcement learning (RL) is a machine learning method. It makes possible the

solution of the problem in a finite time based on its experimental learned knowledge.
In reinforcement learning approach, an agent learns to optimize an interaction with
a dynamic environment through trial and error process. The RL structure is shown in
Fig. 12.2. The agent receives a scalar value or reward with every action it executes.
The objective of the agent (in our case, it is a mobile robot) is to learn a strategy for
selecting actions such that the expected sum of discounted rewards is maximized
(Kaelbling et al. 1996; Sutton and Barto 1998). In the standard reinforcement
learning model, the robot is connected to its environment via perception and action.
At any given time step t, the robot perceives the state st of the environment and
selects an action at . The environment returns by a scalar reinforcement signal rt and
changing into state st+1 (Sutton and Barto 1998). The robot should choose actions
that tend to increase the long run sum of values of the reinforcement signal. It can
learn to do this overtime by systematic trial and error process.
The agent task is to find an optimal policy, π : {S, A} −→ [0, 1] which maps
states to actions in order to maximize the received reinforcements. In general, the
robot’s actions determine not only its immediate rewards, but also the next state of
the environment. As a result, when applying the generated action, the robot must
take the next state into account. Generally, the function value is defined at the form
of PDM by:
∞ 
Vπ (s) = Eπ (Rt st = s) = Eπ γ k rt+k st = s (12.4)
k=1

where γ ∈]0, 1[ is a factor to regulate the importance of future returns.

Fig. 12.2 Reinforcement

learning structure Agent/Robot

State St Reinforcement rt Action at

rt+1
Environment
st+1
12 Mobile Robot Path Planning Based on Optimized Fuzzy Logic Controllers 261

The most algorithms of reinforcement learning use a quality function noted

Q-function, representing the value of each pair state-action to obtain an optimal
behavior (Sutton and Barto 1998). It gives for each state, the future return if the
agent pursues this policy π :

Qπ (s, a) = Eπ (Rt st = s, at = a) (12.5)

The optimal quality is calculated as:

Q (s, a) = max Qπ (s, a) (12.6)

π

We obtain then:

Q (s, a) = E(rt+1 + γ V  (st+1 )st = s, at = a) (12.7)

12.3.1 Q-Learning Algorithm (QLA)

Q-Learning is an artificial intelligent technique that has been successfully used for
solving complex Markovian Decision Process (MDP) that model realistic system.
It is a technique to solve specific problems by using the reinforcement learning
approach. Q-Learning is an incremental version of dynamic programming for
solving multistage decision problems. It is the popular of temporal difference
algorithms proposed by Watkins and Dayan (1992). It is based on three main
functions: an evaluation function, a reinforcement function and an update function
(Cherroun and Boumehraz 2013b). The idea of Q-learning is to learn a Q-function
that maps the current state st and action at to Qπ (st , at ), that predicts the total future
discounted reward that will be received from the current action at . In that, it learns
the optimal policy function incrementally after each transition (st , at , rt , st+1 ).
The update of Q (s, a) is done by observation of the instantaneous transitions
and their rewards associated by the following equation (Kaelbling et al. 1996; Sutton
and Barto 1998):

Q(st , at ) =←− Qπ (st , at ) + α[rt + γ max Q(st+1 , at ) − Q(st , at )] (12.8)

a∈A(s)

where α ∈ [0, 1] is a learning rate.

The quality functions are stored in the form of table: the qualities of all actions
for a given state. Firstly, when the table does not contain sufficient data, a random
component is added in order to restrict the eligible actions with the small number
of the actions already tested. As the table fills, this random component is reduced
in order to allow the exploitation of received information and for obtaining a good
performance.
262 L. Cherroun et al.

In the next section, we will present a method has as advantages two properties:
extending the Q-learning algorithm to continuous state-action spaces, and consid-
ered as an optimization method of FLC parameters (conclusion parts).

12.3.2 Optimization of FLS Using Q-Learning Algorithm

(FQLA)

Fuzzy inference systems (FIS) are promising solutions for representing the quality
functions with continuous spaces of states and actions (Glorennec and Jouffle 1997;
Nemra and Rezine 2008). The idea of this optimization is to propose several
conclusions for each rule, and to associate each conclusion by a quality function
which will be evaluated during the time. The training process permit to obtaining
the best rules that maximizing the future reinforcements (Jouffle 1998). This fuzzy
version of the Q-learning algorithm is named fuzzy Q-learning algorithm (FQLA)
summarized in Table 12.1. The initial proposed rule-base using a zero order Takagi-
Sugeno model is composed therefore of rules and N conclusions such as Eq. (12.9)
from Glorennec and Jouffle (1997), Nemra and Rezine (2008), and Cherroun and
Boumehraz (2013b):

If s is Si Then y = a[i, 1] with q[i, 1] = 0

or y = a[i, 2] with q[i, 2] = 0
(12.9)
:
or y = a[i, N] with q[i, N] = 0

where q(i, j ) with i = 1 . . . m and j = 1 . . . N, are potential solutions whose

values are initialized to 0. During the learning process, the conclusion part of each

Table 12.1 Optimization of FLSs using Q-learning

1. Choose the FIS structure
2. Initialize randomly q[i, j ], i = 1, . . . m (m: rule number)
j = 1, . . . N (N : Number of proposed conclusions)
3. t = 0, observe the state st
4. For each rule i, compute wi (st )
5. For each rule i, choose a conclusion with the EEP
6. Compute the action A(st ) and correspondence quality Q[st , A(st )]
7. Apply the action A(st ). Observe the new state st
8. Receive the reinforcement rt
9. For each rule i, compute wi (st )
10. Compute a new evaluation of the state value
11. Update parameters q[i, j ] using this evaluation
12. t ← t + 1, Go to 5
12 Mobile Robot Path Planning Based on Optimized Fuzzy Logic Controllers 263

rule is selected by means of an exploration-exploitation policy noted (EEP), where

EEP (i) ∈ {1, . . . , N } such as the ε greedy:
3
3 arg max Q(st , a) with ε probability
3
Aε−greedy =3 a∈A(st ) (12.10)
3 Random action A(st ) with 1 − ε probability

In this case, the inferred action is given by:

m
A(s) = wi (s) a[i, EEP (i)] (12.11)
i=1

and the quality of this action will be:

m
/ A(s)] =
Q[s, wi (s) q[i, EEP (i)] (12.12)
i=1

12.4 Application of Optimized Fuzzy Controllers and Results

Some typical cases are simulated in this section to illustrate the effectiveness of
the designed fuzzy controllers. In which, the robot task is to move from a current
position (x0 , y0 ) to a desired goal (xg , yg ) in various unknown environments (free
or with encounter obstacles).

12.4.1 Goal Seeking Task

The role of the goal seeking behavior is to align the mobile robot with the direction
of the desired goal to reach their coordinates (xg , yg ) by considering all types of
environments (containing or without obstacles).
The application of optimized FLC is summarized in the implementation of a
Takagi-Sugeno fuzzy controller for mobile robot navigation. The initialized rule-
base is adjusted on-line using a reinforcement signal (reward or punishment). In
this optimization algorithm, the robot has an initial rule-base which defines the
possible situations with the proposed actions for the desired robot task (optimization
of conclusion parts).

12.4.1.1 Steering Controller

For each time step, the robot must define the state in which it is by calculating
the angle between its orientation and that of the goal noted (θrg ), and the distance
between the robot and the target noted (drg ) defining the position error. Using these
264 L. Cherroun et al.

Fig. 12.3 The MFs of drg m(drg)

N M F

drg (m)
0 0.3 0.8 3

Fig. 12.4 The MFs of θrg μ( qrg )

NB NS Z PS PB

qrg (rad)

–π –π/2 –π/6 0 π/6 π/2 π

two variables as inputs, the optimized FLC must generate the appropriate control
actions allowing the robot to move toward the desired goal. In this first application,
we assume the robot motion with a fixed velocity.
The used membership functions to define the states of this behavior are shown in
Figs. 12.3 and 12.4. This subdivision defines 15 fuzzy rules with the following fuzzy
labels: (N: Near, M: Medium, F: Far, Z: Zero, PS: Positive Small, PB: Positive Big,
NB: Negative Big, NS: Negative Small).
The fuzzy rule-base of the robot goal seeking behavior is as follows:

if s is Si Then α = αi1 with the quality q[i, 1]

or α = αi2 with the quality q[i, 2] (12.13)
or α = αi3 with the quality q[i, 3]

for: i = 1 . . . 15.
For each rule, three conclusions are proposed: α1 = − π5 , α2 = 0, and α1 = π5
(corresponding to the steering angle α of the front robot wheel). After a training
time, the robot should choose for each fuzzy-rule the best conclusion corresponding
to the best Q-function q[i, j ]N
j =1 .
During the optimization process, the robot is guided by the following reinforce-
ment signals:
3
3 4 if the robot reaches the target
3
3 3 if drg decreases and θrg = 0
3
3
3 2 if drg and θrg decrease
r=3 (12.14)
3 −1 if drg decreases and θrg increases
3
3 −2 if drg increases
3
3 −3 if a collision is occurred with the environment
12 Mobile Robot Path Planning Based on Optimized Fuzzy Logic Controllers 265

Starting from the actual perceived state, the robot must take a decision on the action
to be carried out (the steering angle). According to the result obtained during the
execution of the selected action, it is either punished to decrease the probability of
execution of the same action in the future, or rewarded, to support this behavior
in the similar situations. These applied actions are chosen by the exploration-
exploitation policy (EEP ) shown in Eq. (12.10) in order to explore the robot state
spaces.

12.4.1.2 Optimized FLC for a Fixed Position

Firstly, we tested the optimized FLC with fixed positions of the robot and the
goal. Figure 12.5a, b present the robot paths during the learning process (episodes
1 and 40). As seems, the robot can’t reach its goal, but after a sufficient time
of optimization (number of episodes), the robot moves toward the desired goal
successfully (it can reach their final coordinates) by executing continuous actions
(steering angle). This result is depicted in Fig. 12.6. So, the conclusion part of
the fuzzy rule-base is optimized by the selection of the best conclusion for each
state condition. The optimized fuzzy controller is effective and well performed to
accomplish this task.
The average rewards are maximized to a fixed value indicating the best optimiza-
tion of parameters as shown in Fig. 12.7.

Fig. 12.5 (a) Episode 1. (b) Episode 40

266 L. Cherroun et al.

Fig. 12.6 Mobile robot path Navigation Environment

20

18

16
Goal
14

12

Y (m)
10

8
Robot Path
6
Start
4

0
0 2 4 6 8 10 12 14 16 18 20
X (m)

Fig. 12.7 Average rewards Average of Reinforcements for each Episode

2.5

1.5

0.5

−0.5

−1

−1.5
0 100 200 300 400 500
Episode

12.4.1.3 Optimized FLC for Random Positions (15 Fuzzy Rules)

In order to generalize the robot navigation for all possible situations, we used
the previous fuzzy controller with 15 fuzzy rules to generate the steering angle
α. The same parameters are used (number of membership functions: 3 × 5 and 3
conclusions, reinforcement values,. . . ). The optimization process is made with a
random initial positions of the robot and the goal for each episode. Each episode
starts with a random position of the robot and finished when it is reached the target
or strikes the limits of its environment. After an optimization task, the robot behavior
12 Mobile Robot Path Planning Based on Optimized Fuzzy Logic Controllers 267

Fig. 12.8 Maximization of Average of Reinforcements for each Episode

the robot rewards 2.5

1.5

0.5

−0.5

−1

−1.5

−2

−2.5
0 200 400 600 800 1000 1200 1400
Episode

is improved during the time. The average rewards are maximized indicating the best
selected conclusions for each episode (Fig. 12.8). Figure 12.9a–d shows the obtained
robot paths. As depicted, in all cases, the robot moves toward the goal effectively by
executing continuous actions.

12.4.1.4 Optimized FLC (15 Fuzzy Rules-3 Conclusions)

In our contribution, we have applied this optimization method in two steps: the first
to optimize the FLC to generate the appropriate robot steering angle (α), and the
second to learn the optimal motion velocity (vr ).

Velocity Controller

When the robot is able to reach his final goal with a fixed speed (as done previously
by the steering controller with the same parameters), we have applied the second
algorithm (FQLA) to optimize the FLC of velocity. The premisse parts of the fuzzy
rules are the same used for the calculation of the steering angle, but the difference
is in the conclusion part. The proposed fuzzy rule-base is in the form given by the
following equations:

if s is Si Then vr = vi1 with the quality q[i, 1]

or vr = vi2 with the quality q[i, 2] (12.15)
or vr = vi3 with the quality q[i, 1]

for: i = 1 . . . 15.
268 L. Cherroun et al.

a Navigation Environment b Navigation Environment

20 20

18 18
Start Goal
16 16

14 14

12 12
Y (m)

Y (m)
10 10

8 8

6 Goal 6
Start
4 4

2 2

0 0
0 2 4 6 8 10 12 14 16 18 20 0 2 4 6 8 10 12 14 16 18 20
X (m) X (m)
c Navigation Environment d Navigation Environment
20 20

18 18

16 16

14 14

12 12
Start Goal
Y (m)

Y (m)

10 10
Goal Start
8 8

6 6

4 4

2 2

0 0
0 5 10 15 20 0 2 4 6 8 10 12 14 16 18 20
X (m) X (m)

Fig. 12.9 Goal seeking using optimized FLC (15 fuzzy rules-3 conclusions)

To learn the robot velocity action, the mobile robot receives the following
reinforcement values as punishments and rewards:
3
3 −1 if drg and θrg decrease
3
3 −1 if θrg decreases and the velocity and drg increase
r = 33 (12.16)
3 1 if the robot reaches the target with low velocity
3 0 elsewhere

In Fig. 12.10a, b, we illustrate the average values of the obtained reinforcements in

each episode to learn the two control values (steering angle and the linear velocity).
As shown on these two curves, the behavior of the robot improves over time. The
reinforcements are maximized illustrating the convergence of the optimized FLCs to
accomplish this behavior. The trajectories of the mobile robot for this functionality
with different start and end situations after learning are shown in Fig. 12.11a–d.
12 Mobile Robot Path Planning Based on Optimized Fuzzy Logic Controllers 269

a Average of Reinforcements b Average of Reinforcements

3 0.04

0.02
2
0
1
−0.02

0 −0.04

−0.06
−1
−0.08
−2
−0.1

−3 −0.12
0 500 1000 1500 2000 2500 3000 0 100 200 300 400 500 600 700 800 900
Episode Episode

Fig. 12.10 Average values of the received reinforcements: steering and velocity

We observe that the robot is always able to move toward the desired goal from
their initial positions autonomously. The robot stops when it reaches the goal by
executing continuous actions generated by the optimized FLC.

12.4.1.5 Optimized FLC with 35 Fuzzy Rules

In the second test, we will optimize a Takagi-Sugeno FLCs for the goal seeking
behavior by adjusting the conclusion parts for these two control actions (Steering
controller and velocity controller). The two fuzzy systems are defined by 35 fuzzy
rules with 5 proposed conclusions for each rule.

Steering Controller

The premisse part of the FLC is defined by the membership functions depicted
in Figs. 12.12 and 12.13 for the distance (drg ) and the angle (θrg ) respectively.
The used fuzzy labels are: Z(Zeros), S(Small), M(Medium), B(Big), VB(Very
Big), NB(Negative Big), NM(Negative Medium), NS(Negative Small), PS(Positive
Small), PM(Positive Medium) and PB(Positive Big). For the proposed steering !
angle values, it is suggested to uniformly distributed in the interval − π5 , π5 . The
fuzzy rule-base used to define this FLC is as follows:

if s is Si Then α = αi1 with the quality q[i, 1]

or α = αi2 with the quality q[i, 2]
or α = αi3 with the quality q[i, 3] (12.17)
or α = αi4 with the quality q[i, 4]
or α = αi5 with the quality q[i, 5]
270 L. Cherroun et al.

a Navigation Environment b Navigation Environment

20 20

18 18

16 16
Start
14 Goal 14

12 12
Y (m)

Y (m)
10 10

8 8

6 6

4 4 Goal
Start
2 2

0 0
0 5 10 15 20 0 5 10 15 20
X (m) X (m)
c Navigation Environment d Navigation Environment
20 20

18 18

16 16

14 Start 14
Y (m)

Y (m)

Goal
12 12

10 10

8 8
Goal
6 6
Start
4 4

2 2

0 0
0 5 10 15 20 0 5 10 15 20
X (m) X (m)

Fig. 12.11 Goal seeking using optimized FLC (15 fuzzy rules-3 control actions)

Fig. 12.12 Membership m(drg)

functions of drg

Z S M B VB

drg(m)

0.1 3.5 8 13 18 Dmax

where: α1 = − π5 , α2 = − 10 π
, α3 = 0, α4 = 10
π
and α5 = π5 .
As the same application done previously, this FLC is tuned on-line firstly for the
basic control action (steering angle). After that, when the robot is able to reach the
goal, this controller is optimized too to learn the required motion velocity. In both
12 Mobile Robot Path Planning Based on Optimized Fuzzy Logic Controllers 271

Fig. 12.13 Membership m(qrg)

functions of θrg

NB NM NS Z PS PM PB

qrg (rad)

-1.6 -1.5 -0.75 -0.25 0 0.25 0.75 1.5 1.6

optimization phases, the robot chooses for each fuzzy rule the optimal conclusion
with the best quality function q[i, j ]N
j =1 .

Velocity Controller

The fuzzy rule-base used to define this second FLC to learn the velocity is shown
in Eq. (12.18). This controller contains 35 fuzzy rules and 5 proposed velocities
distributed from 0 as a stop action to vmax. This last corresponds the maximum
motion velocity.

if s is Si Then vr = vi1 with the quality q[i, 1]

or vr = vi2 with the quality q[i, 2]
or vr = vi3 with the quality q[i, 3] (12.18)
or vr = vi4 with the quality q[i, 4]
or vr = vi5 with the quality q[i, 5]

The inferred control actions for robot movement are calculated by:

35
α(s) = wi (s) α[i, k] (12.19)
i=1
35
vr (s) = wi (s) vr [i, k] (12.20)
i=1

where: k = 1 . . . 5, wi (s) is he firing strength of the ith rule in the state s.

After the two optimization passes, the FLCs are able to generate the correct
control actions (steering and velocity) to move toward the desired destination. The
indicators of optimization are the mean of the received reward maximized in each
episode as depicted in Fig. 12.14a, b for the two actions respectively. The robot
behavior is improved during the time to obtain the optimal parameters of the steering
and the velocity Controllers. Figure 12.15a–d present the robot paths for different
272 L. Cherroun et al.

a Received Reinforcements b Received Reinforcements

1.5 0.04

1 0.02

0
0.5
−0.02
0
−0.04
−0.5
−0.06
−1
−0.08
−1.5
−0.1

−2 −0.12

−2.5 −0.14
0 1000 2000 3000 4000 5000 6000 0 500 1000 1500
Episode Episode

Fig. 12.14 Average values of the received reinforcements: steering and velocity

a Navigation Environment b Environnement De Navigation

20 20

18 18

16 16
Start
14 Goal 14

12 12
Y (m)

Y (m)

10 10

8 8
Start
6 6
Goal
4 4

2 2

0 0
0 5 10 15 20 0 5 10 15 20
X (m) X (m)

c Navigation Environment d Navigation Environment

20 20

18 18

16 16

14 14 Start
Goal
12 12
Y (m)

Y (m)

10 10

8 8

6 6

4 4 Goal
Start
2 2

0 0
0 5 10 15 20 0 5 10 15 20
X (m) X (m)

Fig. 12.15 Goal seeking using optimized FLC (35 fuzzy rules-5 control actions)
12 Mobile Robot Path Planning Based on Optimized Fuzzy Logic Controllers 273

initial positions of the robot and the goal. The robot is able to reach the target
successfully by executing the optimal actions. The obtained behavior is acceptable
for this task.
As a comparison with the previous controller, the convergence of this algorithm
(35 rules-5conclusions) to optimize the conclusion parts needs more time than the
first (15 fuzzy rules-3 conclusions). But the robot behavior is more suitable.

12.4.2 Obstacle Avoidance Task

Obstacle avoidance (OA) is one of the basic missions of a mobile robot. It is a

significant task that must have all the robots. This behavior permits these vehicles
to move in unknown environments without collisions. The obstacle avoidance tends
to steer the robot in such a way as to avoid collision with the nearest obstacles and
objects that happens to be in the vicinity of the robot. It consists to keep a safe
distance that the robot can move effectively.
To realize this task, it is assumed that the robot is able to determine the distance
to an obstacle noted (D) and the angle between the obstacle and the orientation of
the robot noted (θr0 ). These two variables define an area around the nearest obstacle.
This area is considered critical. If the robot is located in this space, it is penalized
and considered that a collision is occurred. The distance and the angle error to an
obstacle are fuzzified using the membership functions presented in Figs. 12.16 and
12.17 respectively. Where: N: Near, M: Medium, and F: Far, Z: Zeros, NB: Negative
Big, NM: Negative Medium, NS: Negative Small, PS: Positive Small, PM: Positive
Medium and PB: Positive Big.
This fuzzy controller is based on 21 fuzzy rules. The reinforcement signals are
used to select one conclusion from the set of the proposed conclusions: α1 = − π5 ,
α2 = 0 and α3 = π5 . In the tuning process, the robot receives these values as rewards
and punishments:

Fig. 12.16 MFs of the m (D)

distance to obstacles (D)
N M F

Distance (m)

d1 d2 d3
274 L. Cherroun et al.

Fig. 12.17 MFs of θr0 m(qrg)

NB NM NS Z PS PM PB

qro (rad)

-1.6 -1.5 -0.75 -0.25 0 0.25 0.75 1.5 1.6

a b Navigation Environment
15

10
Y (m)

Start
Obstacle
5

0
0 5 10 15
X (m)

Fig. 12.18 (a) Episode 1. (b) Path after optimization

3
3 −3 if the robot sticks an obstacle
3
3 −2 if an obstacle is near area
3
3
3 2 if an obstacle is detected on right or on left,
3
3 or it exists in a free area
3
r = 3 −2 if D decreases and θr0 = 0 (12.21)
3
3 −1 if D and θr0 decrease
3
3 0 if D decreases and θr0 increases
3
3 1.5 if the distance to an obstacle increases
3
3 −1.5 if the robot approaches the limits of the environment

12.4.2.1 Optimized FLC for a Fixed Position

For a fixed situation of the robot and the obstacle to avoid in their environment. In
the first episode, the robot can’t avoid the collision (so the FLC is not yet adjusted) as
depicted in Fig. 12.18a. After a sufficient time of optimization (learning), the robot
is able to avoid collision with the detected obstacle effectively and autonomously as
shown in Fig. 12.18b. The mean values of the received reinforcement are maximized
12 Mobile Robot Path Planning Based on Optimized Fuzzy Logic Controllers 275

Fig. 12.19 Reinforcements Reinforcements

2

1.5

0.5

−0.5

−1

−1.5
0 100 200 300 400 500 600
Episodes

to a constant value (Fig. 12.19) indicating the optimal FLC parameters and the
improvement of the robot behavior.

12.4.2.2 Optimized FLC for Random Positions

With random initial positions of the robot in its environment in each episode.
Figure 12.20a–d presents the paths of the mobile robot to learn the obstacle
avoidance behavior using an optimized FLC with 21 fuzzy rules and 3 proposed
conclusions. After the optimization process, the fuzzy controller can guide the robot
to avoid collisions with the obstacle effectively by generating correct and optimal
control actions (steering with a fixed velocity).
The optimization task is verified by the maximization of the average reinforce-
ment at each episode as depicted in Fig. 12.21.

12.4.2.3 Obstacle Avoidance Using Optimized FLC

Steering Controller

In this application, the optimized fuzzy controller uses as inputs the distances to
obstacles in the three directions of mobile robot (on front dF , on the right dR and
on the left dL ). Figure 12.22 represents the fuzzy sets used to fuzzify the readings
of sensors around the robot with two linguistic variables, Near (N) and Far (F). To
achieve this behavior, the rule-base consists of 8 basic situations. The optimization
phase permits to tuning the FLC parameters in order to accomplish the obstacle
avoidance task. For each fuzzy rule, we associate a quality function q[i, j ]Jj=1 for
276 L. Cherroun et al.

a Navigation Environment b Navigation Environment

15 15

10 10
Y (m)

Y (m)
5 5

0 0
0 5 10 15 0 5 10 15
X (m) X (m)

c Navigation Environment d Navigation Environment

15 15

10 10
Y (m)

Y (m)

5 5

0 0
0 5 10 15 0 5 10 15
X (m) X (m)

Fig. 12.20 Obstacle avoidance (21 rules-3 conclusions)

Fig. 12.21 Reinforcement Average of Reinforcements

values (random positions) (21 2
rules-3 conclusions) 1.5

0.5

−0.5

−1

−1.5

−2
0 200 400 600 800 1000 1200 1400
Episode
12 Mobile Robot Path Planning Based on Optimized Fuzzy Logic Controllers 277

Fig. 12.22 MFs of the

m(di)
distances: dF , dR and dL
N F

Distance (m)
dm ds Dmax

the interpretation j of the rule i, where: i = 1 . . . 8 and j = 1 . . . 3. The used fuzzy

rule-base to tune the steering angle is illustrated as:

if s is Si Then α = αi1 with the quality q[i, 1]

or α = αi2 with the quality q[i, 2] (12.22)
or α = αi3 with the quality q[i, 3]

where: i = 1 . . . 8, j = 1 . . . 3, α1 = − π5 , α2 = 0 and α3 = π
5.
The global quality of the inferred output is given by:
8
Q(s, α) = wi (s) q[, j  (i)] (12.23)
i=1

The reinforcement’s values presented in the following equation are used to the
optimization of the steering angle.
3
3 −4 if a collision is occured
3
r = 33 −1 if di decreases and di < 12 Dmax (12.24)
3 0 elsewhere

This signal punished the activated conclusion permits the robot to enter in collision
with the obstacles. In this case, the rewards values are used to define the best
conclusion part from the three proposed conclusions: α1 , α2 and α3 as steering
angles of the front wheel, where di , for i = 1 . . . 3 are the distances in three
directions, dm is the minimum distance, ds is distance of security and Dmax is the
maximum reading of the sensors.

Velocity Controller

After adjusting the steering values, we apply the same algorithm (FQLA) to generate
the appropriate motion velocity, we used the following reinforcement values:
278 L. Cherroun et al.

Fig. 12.23 Robot path in

episode 1

3
3 −4 if the robot sticks an obstacle
3
3 −1 if di > Dmax and vr is low, with i = 1 . . . 3
r = 33 (12.25)
3 −1 if di < Dmax and vr increases, with i = 1 . . . 3
3 0 elsewhere

The rule-base employed to learn the motion velocity is:

if s is Si Then vr = vi1 with the quality q[i, 1]

or vr = vi2 with the quality q[i, 2] (12.26)
or vr = vi3 with the quality q[i, 1]

In the first steps (during learning), the robot can’t avoid collisions with obstacles
as depicted in Fig. 12.23. After an optimization phase, the robot obtains the best
behavior to reach the goal (by selecting the conclusion that maximize the total
rewards). The robot can avoid the obstacles and moves in the direction of the target.
If there is a near obstacle (detected by their sensors), it chooses the turn right action
as depicted in Fig. 12.24a, b. The obtained behavior is more acceptable and the robot
path is considered as optimal. For this situation, the average values of the received
reinforcements are maximized for the two actions: Steering and velocity as seem in
Fig. 12.25a, b. This maximization indicates the best optimization process.
12 Mobile Robot Path Planning Based on Optimized Fuzzy Logic Controllers 279

a Navigation Environment b Navigation Environment

20 20
Goal
18 18
16 16
Start Obstacle 02
14 14
Obstacle
12 12
Y (m)

Y (m)
10 10
Obstacle 01
8 8
6 Goal 6
4 4
2 2 Start
0 0
0 5 10 15 20 0 5 10 15 20
X (m) X (m)

Fig. 12.24 Robot paths after optimization

a Average of Reinforcements for the Steering b Average of Reinforcements for the Velocity
−0.1 0.1
0
−0.2
−0.1
−0.3 −0.2

−0.4 −0.3
−0.4
−0.5 −0.5
−0.6 −0.6
−0.7
−0.7
−0.8
−0.8 −0.9
0 50 100 150 200 250 300 350 400 450 0 500 1000 1500 2000
Episode Episode

Fig. 12.25 Received reinforcements: steering and velocity (8 rules-3 conclusions)

12.4.3 Wall-Following Task

The objective of the wall following behavior is to keep the robot at a safe close
distance to the nearest right or left wall. This task is accomplished using sensor
readings about its environment presenting distances to obstacles and walls in various
directions (front, right or left). The fuzzy sets used to fuzzify the three distances are
illustrated in Fig. 12.26.

12.4.3.1 Steering Controller

In this application, we use an on-line optimization of this fuzzy rule-base with an

imprecise knowledge using the following reinforcement signals:
280 L. Cherroun et al.

Fig. 12.26 MFs of the

m(di)
distances to walls
N F

Distance (m)
dm ds Dmax

3
3 −2 if a collision is occurred
3
r = 33 −1 if di < d1 i = 1 . . . 3 (12.27)
3 0 elsewhere

The fuzzy rule-base used to tune the steering angle fuzzy controller is as follows:

if s is Si Then vr = vi1 with the quality q[i, 1]

or vr = vi2 with the quality q[i, 2] (12.28)
or vr = vi3 with the quality q[i, 1]

where: i = 1 . . . 8, j = 1 . . . 3, α1 = − π5 , α2 = 0 and α3 = π
5.

12.4.3.2 Velocity Controller

The rule-base employed to learn the robot velocity is:

if s is Si Then vr = vi1 with the quality q[i, 1]

or vr = vi2 with the quality q[i, 2] (12.29)
or vr = vi3 with the quality q[i, 1]

After a learning process, the optimization results are shown in Fig. 12.27a, b
presenting the mobile robot paths in unknown environment. The robot behavior is
improved and the robot is able to navigate in its environment without collision with
walls. It can keep a safe distance to the nearest left wall autonomously. The robot
wall-following behavior is satisfactory in all cases.
The mean values of the reinforcements are maximized as depicted in Fig. 12.28a,
b indicating the optimal FLC parameters for the two control actions: steering and
velocity.
12 Mobile Robot Path Planning Based on Optimized Fuzzy Logic Controllers 281

a Navigation Environment b Navigation Environment

20 20
18 18
16 16
14 14 Walls
12 12
Y (m)

Y (m)
Walls
10 10
8 8
6 6
4 4
2 2
0 0
0 5 10 15 20 0 5 10 15 20
X (m) X (m)

Fig. 12.27 Wall-following with different forms

Fig. 12.28 Average values of the received reinforcements: steering and velocity

12.5 Conclusions

In this chapter, we presented a study of navigation methods for an autonomous

mobile robot using optimized fuzzy logic controllers. The Q-learning algorithm is
a powerful tool to obtain an optimal fuzzy behavior which requires only one scalar
signal as a feedback indicating the quality of the applied action. Fuzzy controller
promises an efficient tool for mobile robot path planning. It is tolerant to noise and
error in the information coming from the sensory system and uncertainties, and most
importantly; it is a factual reflection of the behavior of human expertise based on If-
Then inference rules. Fuzzy Controller can be effectively tuned via reinforcement
learning using FQLA. The idea of fuzzy Q-Learning algorithm consists at fuzzy
inference systems optimization using a reinforcement signal. This signal makes
the navigator able to adjust his strategy in order to improve their performances.
This technique is based on the optimization of conclusion parts of FLC in order to
maximize the return function. The consequent value is selected from predefined
values set which is kept unchanged during learning. This used method has two
282 L. Cherroun et al.

properties: extending the Q-learning algorithm to continuous state-action spaces and

considered as a FLC parameters optimization using Q-learning algorithm.
The optimization of premisse parameters and the number of rules will improve
the performances of the studied methods. The interest will be given to reduce the
time of optimization-learning.

References

Abadi, N. M., & Khooban, M. H. (2015). Design of optimal Mamdani-type fuzzy controller for
nonholonomic wheeled mobile robots. Journal of King Saud University-Engineering Sciences,
27, 92–100.
Algabri, M., Mathkour, H., & Ramdane H. (2014a). Mobile robot navigation and obstacle-
avoidance using ANFIS in unknown environment. International Journal of Computer Appli-
cations, 91, 36–41.
Algabri, M., Ramdane, H., Mathkour, H., Al-Mutib, K., & Alsulaiman, M. (2014b). Optimization
of fuzzy logic controller using PSO for mobile robot navigation in an unknown environment.
Applied Mechanics and Materials, 541, 1053–1061.
Algabri, M., Mathkour, H., Ramdane, H., & Alsulaiman, M. (2015). Comparative study of soft
computing techniques for mobile robot navigation in an unknown environment. Computers in
Human Behavior, 50, 42–56.
Al Yahmedi, A. S., & Fatmi, A. (2011). Fuzzy logic based navigation of mobile robots. In A. V.
Topalov (Ed.), Recent advances in mobile robotics (pp. 287–310). Croatia: InTech. ISBN:978-
953-307-909-7.
Antonelli, G. C., & Fusco, S. G. (2007). A fuzzy logic based approach for mobile robot path
tracking. IEEE Transactions on Systems, Man and Cybernetics: Systems and Humans, 15, 211–
221.
Bakdi, A., Hentouta, A., Boutamiab, H., Maoudja, A., Hachourb, O., & Bouzouiaa, B. (2016).
Optimal path planning and execution for mobile robots using genetic algorithm and adaptive
fuzzy-logic control. Robotics and Autonomous Systems, 73, 95–109.
Begum, M., Mann, G. K., & Gosine, R. G. (2008). Integrated fuzzy logic and genetic algorithmic
approach for simultaneous localization and mapping of mobile robots. Applied Soft Computing,
8, 150–165.
Berenji, H. R., & Khedkar, P. (1992). Learning and tuning fuzzy logic controllers through
reinforcements. IEEE Transactions on Neural Network, 3, 724–740.
Borestein, J., & Koren, Y. (1989). Real-time obstacle avoidance for fast mobile robot. IEEE
Transactions on Systems, Man, and Cybernetics, 19, 1179–1186.
Cherroun, L., & Boumehraz, M. (2013a). Fuzzy behavior based navigation approach for mobile
robot in unknown environment. Journal of Electrical Engineering, 13, 284–291.
Cherroun, L., & Boumehraz, M. (2013b). Fuzzy Logic and reinforcement learning based
approaches for mobile robot navigation in unknown environment. Mediterranean Journal of
Measurement and Control, 9, 109–117.
Cuesta, F., & Ollero, A. (2005). Intelligent mobile robot navigation. Heidelberg: Springer.
Glorennec. P. Y., & Jouffle. L. (1997). Fuzzy Q-learning. In 6th IEEE International Conference on
Fuzzy Systems, Barcelona (pp. 659–662).
Hsu, C. C., Hou, R. Y., & Wang, W. Y. (2013). Path planning for mobile robots based on improved
Ant Colony optimization. In IEEE International Conference on Systems, Man, and Cybernetics
(SMC), Manchester (pp. 2777–2782).
Jallouli, M., Rekik. C., Chtourou. M., & Derbel. N. (2010). Optimised fuzzy Logic controller for
a mobile robot navigation. International Journal of Modelling, Identification and Control, 9,
400–408.
Joo, M., & Deng, C. (2005). Obstacle avoidance of a mobile robot using hybrid learning approach.
IEEE Transactions on Industrial Electronics, 52, 898–905.
12 Mobile Robot Path Planning Based on Optimized Fuzzy Logic Controllers 283

Jouffle, L. (1998). Fuzzy inference system learning by reinforcement methods. IEEE Transactions
on Systems, Man, and Cybernetics-Part C, 28, 338–355.
Juang, C. F., & Hsu, C. H. (2009). Reinforcement ant optimized fuzzy controller for mobile-robot
wall-following control. IEEE Transactions on Industrial Electronics, 56, 3931–3940.
Kaelbling, L. P., Littman, M. L., & Moore. A. W. (1996). Reinforcement learning: A survey.
Journal of Artificial Intelligence Research, 4, 237–285.
Khatib, O. (1986). Real-time obstacle avoidance for manipulators and mobile robots. International
Journal of Robotics Research, 5, 90–98.
Latombe, J. C. (1991). Robot motion planning. Norwell: Kluwer.
Mihai, D., & Gheorghe, M. (2016). Neural networks based reinforcement learning for mobile
robots obstacle avoidance. Expert Systems with Applications, 62, 104–115.
Nemra, A., & Rezine, H. (2008). Genetic reinforcement learning algorithms for on-line fuzzy
inference system tuning, application to mobile robotic. In P. Pecherková, M. Flídr, & J. Duník
(Eds.), Robotics, automation and control (pp. 228–256). Vienna: InTech. ISBN:978-953-7619-
18-3.
Passino, K. M., & Yurkovich, S. (1998). Fuzzy control. Menlo Park: Addison Wesley.
Pothal, J. K., & Parhi, D. R. (2015). Navigation of multiple mobile robots in a highly clutter terrains
using adaptive neuro-fuzzy inference system. Robotics and Autonomous Systems, 72, 48–58.
Seraji, H., & Howard, A. (2002). Behavior-based robot navigation on challenging terrain: A fuzzy
Logic approach. IEEE Transactions on Robotics and Automation, 18, 308–321.
Sutton, R. S., & Barto, A. G. (1998). Reinforcement learning: An introduction. Cambridge: The
MIT Press.
Timothy, J. R. (2004). Fuzzy logic with engineering applications (2nd ed.). Hoboken: Wiley.
Vadakkepa, P., Miin, O. C., Peng, X., & Lee, T. H. (2004). Fuzzy behavior-based control of mobile
robots. IEEE Transactions on Fuzzy Systems, 12, 559–564.
Wang, M., & Liu, J. N. K. (2008). Fuzzy logic based real-time robot navigation in unknown
environment with dead ends. Robotics and Autonomous Systems, 56, 625–643.
Watkins, C., & Dayan, P. (1992). Q-Learning. Machine Learning, 8, 279–292.
Yang, S. X., Moallem, M., & Patel, R. V. (2005). A layered goal-oriented fuzzy motion planning
strategy for mobile robot navigation. IEEE Transactions on Systems, Man, and Cybernetics-
Part B: Cybernetics, 35, 1214–1224.
Zhao, R., Lee, D. H., & Hong, K. L. (2015). Mobile robot navigation using optimized fuzzy
controller by genetic algorithm. International Journal of Fuzzy Logic and Intelligent Systems,
15, 12–19.
Chapter 13
Design and Implementation of a Reactive
Navigation System for a Smart Robot
Using Udoo Quad

Mohamed Njah and Ridha El-Hamdi

Abstract This chapter describes a process of design and implementation of a

reactive navigation system for a smart mobile robot. Equipped with a webcam and
distance sensors, the autonomous robot will explore an arena to locate a number
of sites in a limited time all while avoiding the arena boundary and any obstacles it
might encounter. A fuzzy behavior-based control scheme with adaptive membership
functions has been taken as a proposed reactive navigation system. The tests of the
proposed method were performed in a real robot using a UDOO Quad board, which
is a single-board computer equipped with two CPU, and experiments demonstrated
that this embedded system was able to successfully complete the autonomous
navigation task in a real arena.

Keywords Reactive navigation system · Behavior-based control · Mobile robot ·

Fuzzy logic · Adaptive membership function · UDOO Quad · OpenCV

13.1 Introduction

Considering the tremendous progress in technology, autonomous robots are attract-

ing more and more attention worldwide, and there are a lot of ongoing research and
development activities in both industry and academia. There are important benefits
in using autonomous mobile robot system to perform different task in AI area (Arena
et al. 2004; Assis et al. 2016; Hashikawa and Morioka 2015; Siegwart et al. 2011).
For example, robots may be used in prospecting missions in deep underwater sites,
hostile environments, or outer space, where human operators could be at risk (Lin
and Tzeng 2014).

M. Njah () · R. El-Hamdi

Control and Energy Management laboratory (CEM), Digital Research Center of Sfax, Technopole
of Sfax, Sakiet Ezzit, Tunisia
e-mail: [email protected]; [email protected]

© Springer Nature Singapore Pte Ltd. 2019 285

N. Derbel et al. (eds.), New Developments and Advances
in Robot Control, Studies in Systems, Decision and Control 175,
https://doi.org/10.1007/978-981-13-2212-9_13
286 M. Njah and R. El-Hamdi

Navigating a mobile robot in a known environment requires little or no inter-

action with the environment. However, a robot may have to operate in partially
known or completely unknown environments, like planetary exploration (Mathers
et al. 2012). In such circumstances the robot should be able to react appropriately to
unforeseen changes in the environment taking into account noise and the inaccuracy
of sensor information: it results in a great advantage for the implementation of the
reactive navigation system to adopt the best behavior-based architecture to take each
time the right decisions on the basis of its perception of the external environment
and to avoid possible damage of the robot.
The work reported in this chapter is motivated by the need to develop an efficient
reactive navigation system for an automated robot able to explore an arena to locate
a number of sites in a limited time all while avoiding the arena boundary and any
obstacles it might encounter.
The rest of this chapter is organized as follows: Sect. 13.2 gives a summary
of work in the field of reactive navigation. In Sect. 13.3, the proposed reactive
navigation system is presented. It is based on Fuzzy Inference Systems (FIS) with
an adaptive membership function scheme so as to avoid revisiting the same areas
twice. For solving the technical challenge of integrating such system elements and
the webcam processing, we use UDOO Quad to implement the software behavior-
based control system. Details of the design and its implementation, including
examples of experimental results, are presented in Sect. 13.4. The last section
contains conclusions and directions for future work.

13.2 Review of Related Studies

The reaction of autonomous mobile robot to unforeseen changes in the environ-

ment is one of the most difficult issues in the control of intelligent autonomous
robot movements. Reactive navigation systems typically have low computational
complexity, often on the order of O(n) (Murphy 2000). Several approaches have
been proposed to address the problem of the optimal design of reactive navigation
systems. The Potential Field algorithm has been a popular reactive approach for low
speed, highly manoeuvrable mobile robots (Siegwart and Nourbakhsh 2004). The
attractiveness of this approach is largely due to the simplicity in summing attractive
and repulsive forces to generate a desired steering direction and speed. However, in
its simple form, Potential Fields is sensitive to large, or closely spaced, obstacles.
In addition, if multiple obstacles are in close proximity, Potential Fields can exhibit
oscillatory navigation. Even worse, the robot can stop at local minima, where all the
attractive/repulsive forces sum to zero (Koren and Borenstein 1991).
In Xu et al. (2015) proposed a deployment algorithm based on the electrostatic
field theory for mobile wireless sensor networks. According to this algorithm, the
obstacles and nodes in the deployment area are taken as the charged particles and
the particles will move due to the Coulomb’s force from other particles or obstacles.
At the beginning of the deployment, there is a relatively large Coulomb’s force and
13 Reactive Navigation System for a Smart Robot 287

acceleration between nodes due to their higher densities. The velocity of the moving
nodes is increasing quickly, making the nodes spread over the deployment region.
When nodes hit the boundary, the repulsive force from the boundary will exert on
them and slow down the velocity until they will reach equilibrium.
A new potential field technique is proposed by Zhang (2013) to eliminate largely
the needless obstacle avoidance. The disadvantage of this algorithm is the collision
when the velocity of the hindrance increases since the repulsive force is too minor
to escape the moving obstacle. This shortcoming may be largely overcome by using
dynamically expanding perception zones to help track objects of immediate interest.
The Dynamic Expanding Zones (DEZ) reactive navigation algorithm (Putney 2006)
is believed to represent a fundamental contribution to the body of knowledge in
the area of high-speed reactive navigation. This method was implemented on the
Virginia Tech DARPA Grand Challenge vehicles (Putney 2006).
Kunwar and Benhabib (2008) have proposed a time-optimal interception method
for the autonomous robotic interception of moving targets in the presence of
dynamic obstacles. The interception manoeuvre is computed using a five-state
Extended Kalman Filter (EKF). The proposed algorithm, then, uses the novel
Advanced Predictive Guidance Law (APGL) to obtain the required acceleration
commands for rendezvous with the target. In the presence of obstacles, the algorithm
uses the novel Obstacle-Avoidance Navigation Law (OANL), which first predicts
the likelihood of a collision and, then defines a collision cone for each potentially
colliding obstacle (within close proximity of) the pursuer. Based on this information
the algorithm directs the relative velocity between the pursuer and the obstacle
outside the collision cone. Tang et al. (2013) suggested a new reactive obstacle
avoidance method that minimizes the tasks complexity and improves the reactivity,
but it cannot gain the shortest route. This technique is based on the “situated-activity
paradigm” and a “divide and conquer” approach, which navigate the robot to travel
among unidentified obstacles and towards a goal without hitting.

13.3 Fuzzy Behavior-Based Control Scheme

As we have mentioned in the introduction, a fuzzy behavior-based control scheme

with adaptive membership functions has been taken as a reactive navigation system.
Figure 13.1 illustrates the block diagram of the proposed system.
The robot is equipped with:
• A USB Webcam, which provides the localization of the sites identified.
• Two independent continuous rotation servo for steering and driving.
• Four ultrasonic distance sensors, which provide information on the distance
between the robot and obstacles or the side of the arena.
The robot receives a continuous flow of information, from the webcam and
distance sensors, about occurring events and generates new commands in response
to the incoming events, while previously planned motions are being executed.
 288

Fig. 13.1 Block diagram of the proposed reactive navigation system

M. Njah and R. El-Hamdi
13 Reactive Navigation System for a Smart Robot 289

Fig. 13.2 The schematic diagram of the autonomous ground robot

In addition to sites and obstacles information, the estimation of the robot’s current
position (Xc (t), Yc (t) and θ (t) as shown in Fig. 13.2) is communicated to the fuzzy
behavior via the position estimator block.
In this chapter, a simple way is used to estimate pose. With knowledge of the
wheel diameters, these values can be converted to translation and rotation values.
This is easy to implement but the odometry is not representative of the exact distance
travelled since errors in position quickly accumulate due to wheel slippage and due
to the open-loop nature there is no way to correct these errors. In this work, the fuzzy
behavior uses the information provided by the position estimator block in order to
avoid revisiting the same areas twice. For this reason, basic wheel odometry is used
as the only pose estimator in the robot system (it is not necessary to use a more
sophisticated algorithm to know the exact position and orientation of the robot).
The concept of fuzzy behavior is similar to the conventional fuzzy logic control
in that it performs an inference mapping from some input space to some output
space. As shown in Fig. 13.3, the developed fuzzy behavior has 16 inputs, 2 outputs
and 76 rules.
The first two inputs are the X,Y-coordinate of the nearest site detected in the
video stream from the USB webcam. The used camera coordinates (x_camera,
y_camera) is shown in Fig. 13.2 and the video stream image is 300 × 300. Fig-
ure 13.4 shows membership functions used to represent site position.
290 M. Njah and R. El-Hamdi

Fig. 13.3 System Fuzzy Behavior: 16 inputs, 2 outputs, 76 rules

Fig. 13.4 Membership functions used to represent site position

The fuzzy variables used to control obstacle position are d1 , d2 , d3 and d4 .

Figure 13.5 shows membership functions used to represent distance to obstacle
provided by the first ultrasonic sensor (associated with the input variable d1 ). Note
that the other three input variables (d2 , d3 and d4 ) have the same membership
functions.
In order to avoid revisiting the same areas twice, the proposed reactive system
has to keep track of which areas of the environment have already been explored.
The challenge here resides in defining adaptive membership functions that change
according to the robot movement.
The design of a fuzzy model that incorporates dynamical membership functions
whose parameters are adjusted online is related to the idea of variable universe
of discourse (VUD), which can tune online universes of discourse using a set of
13 Reactive Navigation System for a Smart Robot 291

Fig. 13.5 Membership functions used to represent distance to obstacle

Fig. 13.6 Example of adaptive membership functions used to represent information provided by
the position estimator block

nonlinear scale factors. The VUD fuzzy controllers have been applied in many
fields of intelligent controlling which was applied to a nonlinear control system
and obtained fine control effects (Long et al. 2010; Wang et al. 2017).
The developed fuzzy behavior uses nine input variables to represent information
provided by the position estimator block. Figure 13.6 shows an example of adaptive
membership functions used to represent information provided by the position
292 M. Njah and R. El-Hamdi

Fig. 13.7 Membership functions used to represent the output variable “PWM1”

estimator block. The idea is to divide the y axis of the arena coordinates (x_arena,
y_arena) into many regions of equal lengths W (shown in Fig. 13.2). For each
region represented by a membership function MF i, we associate an input variable
X_MFi with three adaptive membership functions as depicted in Fig. 13.6. The
fuzzy behavior rules (76 rules) are constructed such that the proposed reactive
system takes each time the right decisions on the basis of the perception of the
external environment of the robot. The decisions of the fuzzy behavior are the
robot’s right and left wheel’s velocities.
Figure 13.7 shows membership functions used to represent the first output
decision, with 0 being full-speed in one direction, 180 being full speed in the other,
and a value near 90 being no movement. The same membership functions are used
to the second output decision.

13.4 System Design Overview

This section describes results of the design and implementation of the proposed
reactive system architecture in the real autonomous robot. Electronic part of the
system consists of an RGB Led, a power supply module, a bidirectional voltage-
level translator 3.3–5 V, a single board computer (Udoo Quad) and a 7.4 V LiPo
battery. Figure 13.8 provides a simple overall system block diagram of the robot. It
shows also how the various components connect together.
The main processing system is a powerful board computer (called Udoo Quad1 ),
based on quad core ARM cortex-A9 CPU (Freescalei.MX6) with great performance
both on Android and Linux OS, and a dedicated ARM processor (Atmel SAM3X8E)
for the General Purpose Input/Output (GPIO) (UDOO 2013). Figure 13.9 shows the
main components of the Udoo quad board.
The Freescale i.MX6 is used to perform complex operations such as the fuzzy
behavior-based control scheme and the vision processing used to identify any sites

1 UDOO website: http://www.udoo.org/

13 Reactive Navigation System for a Smart Robot 293

Fig. 13.8 Overall system block diagram

294 M. Njah and R. El-Hamdi

Fig. 13.9 The main components of the Udoo quad board

within the camera’s field of vision. Whereas applications that manage external pins,
which can be used to read digital information from sensors, or to control lights
or motors, are developed torun on the SAM3X8E microcontroller. Both processors
(Freescale i.MX6 and Atmel SAM3X8E) can listen and send data via shared serial
line port embedded on the board as depicted in Fig. 13.10. Real prototype of the
robot is shown in Fig. 13.11.
Udoo Quad has no internal storage or built-in boot code, operating system (OS)
and storage are on micro SD. UDOObuntu is the Linux-based distribution used in
this work.
To perform the software implementation, several software development tools
have been used. These tools are:
• The OpenCV-Python,2 which is a library of Python bindings designed to solve
computer vision problems. This library is used to carry out the computer vision
processing.
• The Scikit-fuzzy,3 which is a collection of fuzzy logic algorithms intended for use
in the SciPy Stack, written in the Python Programming language. This package
is used to carry out the fuzzy behavior-based control scheme.
• The Arduino IDE for UDOO,4 which is a custom version of the original
Arduino IDE 1.5.X. This software is used to program the SAM3X8E embedded
microcontroller from Udoo itself.

2 OpenCV website: http://opencv.org

3 Scikit-fuzzy
website: http://pythonhosted.org/scikit-fuzzy/overview.html
4 Arduino IDE for UDOO: http://udoo.org/download/files/arduino-1.5.4-for__UDOO.tar.gz
13 Reactive Navigation System for a Smart Robot 295

TX0/RX0

Freescale imx6 Atmel Sam 3x

Programming Serial
Switch Switch
Control Control

Native USB
UART 2
UART

OTG

OTG bus

J18

UART serial

usb-to-serial
converter

mini-usb serial mini-usb OTG

Fig. 13.10 Communication between I.MX6 and SAM3X (UDOO 2013)

Fig. 13.11 Autonomous mobile robot prototype

296 M. Njah and R. El-Hamdi

Fig. 13.12 Example of code using the OpenCV-Python library under UDOO buntu distribution

Figure 13.12 shows an example of code using the OpenCV-Python library under
the used Linux-based distribution (UDOObuntu).
Figure 13.13 shows the autonomous robot prototype in the real arena. The arena
contains:
• Blue dots: they stand for sites that the robot has to locate.
• Obstacles to avoid
Figure 13.14a, b show the camera’s field of vision obtained every 0.2 s. In these
figures, it is possible to show how the real robot take each time the right decisions
on the basis of its perception of the external environment. Once more than one site
is detected in the camera’s field of vision (Fig. 13.14b), the decisions of the fuzzy
13 Reactive Navigation System for a Smart Robot 297

Fig. 13.13 The autonomous robot prototype in the real arena

behavior is to make the robot move toward the nearest site detected. This is done
one the detected site is in the middle of the camera’s field of view.

13.5 Conclusion

In this chapter, a novel reactive navigation system for an automated robot is

presented. The proposed reactive system is based on a fuzzy behavior-based control
scheme with an adaptive membership function. The goal of such fuzzy behavior is
to increase a robot’s knowledge of its workspace by selecting appropriate control
actions (robot’s right and left wheel’s velocities).
To implement the proposed reactive system in the real autonomous robot we
have adapted the technology available nowadays based on a powerful single board
computer, called Udoo Quad. We have shown that the autonomous robot prototype
equipped with a webcam and distance sensors is able to take each time the right
decisions on the basis of its perception of the external environment, which will lead
it to visit previously unseen areas and locate a number of sites, while avoiding the
arena boundary and any obstacles it might encounter.
We are presently working towards implementing such approach using Robot
Operating System (ROS) Platform.
298 M. Njah and R. El-Hamdi

Fig. 13.14 (a): Camera’s field of vision obtained every 0.2 s (from bottom to top). Only one site
is detected. (b): Camera’s field of vision obtained every 0.2 s. Two sites are detected
13 Reactive Navigation System for a Smart Robot 299

References

Arena, P., Di Giamberardino, P., Fortuna, L., La Gala, F., Monaco, S., Muscato, G., Rizzo, A.,
& Ronchini, R. (2004). Toward a mobile autonomous robotic system for Mars exploration.
Planetary and Space Science, 52, 23–30.
Assis, L. S., Soares, A. S., Coelho, C. J., & Baalen, J. V. (2016). An evolutionary algorithm for
autonomous robot navigation. International Conference on Computational Science, 80, 2261–
2265.
Hashikawa, F., & Morioka, K. (2015). An assistance system for building intelligent spaces based
on mapsharing among a mobile robot and distributed sensors. International Journal on Smart
Sensing and Intelligent Systems, 8(1), 1–25.
Lin, H.-I., & Tzeng, H. Jr. (2014). Search strategy of a mobile robot for radiation sources in an
unknown environment. In International Conference on Advanced and Intelligent Systems (pp.
56–60).
Long, Z., Liang, X., & Yang, L. (2010). Some approximation properties of adaptive fuzzy systems
with variable universe of discourse. Information Sciences, 180(16), 2991–3005.
Mathers, N., Goktogen, A., Rankin, J., & Anderson, M. (2012). Robotic mission to Mars: Hands-
on, minds-on, web-based learning. Acta Astronautica, 80, 124–131.
Murphy, R. (2000). Introduction to AI robotics. Cambridge: MIT Press.
Koren, Y., & Borenstein, J. (1991). Potential field methods and their inherent limitations for mobile
robot navigation. In IEEE International Conference on Robotics and Automation (pp. 1398–
1404).
Kunwar, F., & Benhabib, B. (2008). Advanced predictive guidance navigation for mobile robots:
A novel strategy for rendezvous in dynamic settings. International Journal on Smart Sensing
and Intelligent Systems, 1(4), 858–890.
Putney, J. S. (2006). Reactive navigation of an autonomous ground vehicle using dynamic
expanding zones. Thesis, Faculty of the Virginia Polytechnic Institute and State University.
Siegwart, R., & Nourbakhsh, I. N. (2004). Introduction to autonomous mobile robots. Cambridge:
MIT Press.
Siegwart, R., Nourbakhsh, I. R., & Scaramuzza, D. (2011). Introduction to autonomous mobile
robots (2nd ed.). Cambridge: MIT Press.
Tang, S. H., Ang, C. K., Nakhaeinia, D., Karasfi, B., & Motlagh, O. (2013). A reactive collision
avoidance approach for mobile robot in dynamic environments. Journal of Automation and
Control Engineering, 1(1), 16–20.
UDOO Starting Manual. http://udoo.org/download/files/Documents/UDOO_Starting_Manual_
beta0.4_11_28_2013.pdf.
Wang, R., Liu, Y. J., Yu, F. S., Wang, J. Y., & Yang, J. L. (2017). Adaptive variable universe
of discourse fuzzy control for a class of nonlinear systems with unknown dead zones.
International Journal of Adaptive Control and Signal Processing, 31(12), 1934–1951.
Xu, J., Qian, H., Ying, W., & Zhang, J. (2015). A deployment algorithm for mobile wireless sensor
networks based on the electrostatic filed theory. International Journal on Smart Sensing and
Intelligent Systems, 8(1), 516–537.
Zhang, Q., Yue, S. G., Yin, Q. J., & Zha, Y. B. (2013). Dynamic obstacle-avoiding path planning
for robots based on modified potential field method. Intelligent Computing Theories and
Technology, 9, 332–342.
Chapter 14
Area Coverage Algorithms for
Networked Multi-robot Systems

Lamia Iftekhar, H. M. Nafid Rahman, and Imran Rahman

Abstract The area coverage problem in robotics alludes to the challenge of getting
a network of mobile robots to efficiently cover or sense a spatial region of interest in
an optimal way. A significant number of coverage algorithms exists, especially from
works on sensor networks and related computer science fields, offering a wide range
of coverage solutions. However, many of them are not readily applicable to robotic
systems due to computational complexity of the theoretical algorithms as well as
the physical limitations of the robots in terms of sensing and processing. In this
chapter, we take a look at various area coverage algorithms that can be implemented
upon a system of networked robots and try to identify the ideal characteristics of an
area coverage algorithm whose target application is in mobile robotics. We propose
a set of simple flocking-based algorithms which inherently captures the identified
characteristics of a good coverage algorithm for a multi-robot system. The proposed
algorithm is simulated under various conditions to demonstrate its feasibility. This
chapter also gives the reader further direction towards open research problems on
this topic.

Keywords Distributed control · Swarm robots · Sensor networks · Coverage

optimization · Cooperative behavior

L. Iftekhar ()
Department of Electrical and Computer Engineering, North South University, Dhaka, Bangladesh
e-mail: [email protected]
H. M. Nafid Rahman
Ceridian HCM, Toronto, ON, Canada
I. Rahman
The Bradley Department of Electrical and Computer Engineering, Virginia Polytechnic Institute
and State University, Blacksburg, VA, USA
e-mail: [email protected]

© Springer Nature Singapore Pte Ltd. 2019 301

N. Derbel et al. (eds.), New Developments and Advances
in Robot Control, Studies in Systems, Decision and Control 175,
https://doi.org/10.1007/978-981-13-2212-9_14
302 L. Iftekhar et al.

14.1 Introduction

Multi-robot systems have become the preferred framework in solving many different
kinds of challenges in various fields of robotics. Instead of using a single, gigantic,
prohibitively expensive, computationally complex and power-hungry robot, the
same task can be completed with a group of simpler robots efficiently. On top of this,
with the large-scale advent of distributed control and collaborative protocols, multi-
robot systems have been yielding fascinating results, smartly utilizing comparatively
inexpensive robots with limitations in both sensing and communication, limited
battery life and low-power builds. One illustrative example can be found in the
problem of coverage, where an area has to ‘covered’ or kept under surveillance
all the time. Rather than having a lone patrolling robot (whose slightest malfunction
may jeopardize the whole mission of coverage), a multi-robot system can perform
the task in a more intelligent and robust manner, without compromising the overall
performance, even if a few of the robots here or there are impaired.
The task of Area Coverage started out as a problem in robotics. Douglas W
Gage in 1992 (Gage 1992) presented the problem of coverage in a systematic
manner for a multi-robot (which he termed as ‘many-robot’) system and introduced
new taxonomy for different aspects of the problem. Interestingly, computer and
communication scientists captured the idea and applied it to Wireless Sensor
Networks, making vast progress in devising algorithms for a network of sensors
(nodes) to spread out over an area of interest, extracting and exchanging information
efficiently. Recently, with the advancement in networked vehicular robotics as well
as mobile robotics in general, a growing number of researchers have been working
on getting robotic systems to perform coverage tasks with varied applications in
military surveillance (Mahamuni 2016), structural surveillance (Adaldo et al. 2017),
monitoring of forests and wildlife (Zheng et al. 2010), agriculture (Hameed 2014;
Janani et al. 2016) indoor and outdoor cleaning (Xin et al. 2017; Yuming et al.
2011), clearing mines (Habib 2017), and many more (e.g. Chen et al. 2014 and
Piciarelli et al. 2016). Although, robotics researchers build upon the algorithms
from the large body of literature on wireless sensor networks, various challenges
need to be addressed in new ways when nodes are replaced by mobile robots. In
such cases, algorithms have to be customized or devised to suit the complexities
associated with different kinds of robot dynamics, such as differential-drive robots,
four-wheeled robots, fixed wing planes, quad-rotor models etc.
This chapter serves as a quick primer on latest works on coverage algorithms,
especially designed for multi-robot systems (Sect. 14.2). It also applies modified
flocking algorithms to achieve the area coverage task. We demonstrate that the
proposed algorithms are simple compared to many existing ones. Yet these area
coverage algorithms incorporate advantages, such as being able to cover an unstruc-
tured area as well as not waste coverage resources on obstacles and non-spaces,
unnecessary. The description of the flocking-based area coverage algorithms is
detailed in Sect. 14.3, after which numerical simulation results are presented in
Sect. 14.4. The chapter closes with some guidance for researchers interested in
delving into this interesting aspect of robotics by offering open problems.
14 Area Coverage Algorithms for Networked Multi-robot Systems 303

14.2 Discussion on the Coverage Problem

and Existing Solutions

We take a tour of the recent advances in the area coverage problem, focusing on
the application of robotics, in this section. The coverage problem is an old one
stemming from military applications three decades ago, and have gone through
phases of different terminologies and popular solution strategies, but we focus on
the recent works, preferably devised for robotics applications (Fig. 14.1).

14.2.1 What Is Coverage?

The task of coverage by a multi-robot system is the efficient autonomous distribution

of the robots over an area or region of interest (ROI). Each robot’s sensing
capabilities extend over a limited spatial region around itself, and that region is
said to be covered by the robot. In wireless sensor network (WSN) terminology,
the sensors themselves are the discrete elements analogous to the robots, and are
denoted as nodes. The objective in coverage tasks would then be to find the ideal
placement of the nodes so that any event that might happen in the region of interest
falls in the sensing region of at least one node and thus does not go undetected. An
event could be an occurrence of interest which the sensors are attempting to detect,
such as an intrusion by a foreign object. In this chapter, we use sensors, nodes and
robots interchangeably (Fig. 14.2).
From literature, four types of coverage tasks can be extracted.
1. Area coverage or Blanket coverage: The robots are required to get distributed
over the whole region of interest and ensure as much coverage of the entire
area as possible. The area of interest is generally a convex area, bounded and
connected. (E.g. Cheng and Savkin 2013, Nasimov and Matveev 2014, Santoso
2011)

Fig. 14.1 An α-agent and its

neighbors. The triangles
represent α agents. Agent i
has four neighbors because
they fall within its sensing
and communication range.
Neighbors are assumed to
have a mutual sensing and
communication relationship.
Ni denotes the set of
neighbors of agent i. Agent 5
is not a neighbor of agent i
Olfati-Saber (2006)
304 L. Iftekhar et al.

Fig. 14.2 An α-lattice and a

quasi α-lattice. Note the
triangular formations.
Source: Olfati-Saber (2006)

2. Barrier coverage: The robots are required to get distributed along the perimeter
of an area of interest, or a slice of the perimeter, such as an entry point or exit
point. (E.g. Tao and Wu 2015)
3. Sweep coverage: The robots are required to move in groups and continuously
move around a region of interest, to maximize detection of any event. This is a
preferred coverage protocol if the number of robots available is small compared
to the size of the region of interest. (E.g. Banimelhem et al. 2017)
4. Target coverage: The robots are required to get distributed in a region of interest
and always keep one or several target points under surveillance. The targets could
be static or dynamic (in the latter case, the coverage algorithm would overlap
with a target tracking algorithm). (E.g. Yang et al. 2015).
In the earliest of coverage problems, the sensors/nodes could be considered
stationary and in that case, the placement of the sensors to achieve maximum
coverage is calculated beforehand and the sensors are manually placed. However,
in robotics, we are of course interested in dynamic nodes, that is, a group of mobile
robots are deployed in the region of interest and they are expected to get themselves
distributed in an automated fashion to achieve the desired coverage goals. This
mobility and self-organization capability of the robots are beneficial, not only for
establishing the necessary coverage of the region, but also to re-organize in case
there is a sudden gap in coverage due to malfunction of a robot (node failure) or
any other disturbance. Hence, we will limit our discussion to algorithms that can
show this kind of robustness, even if at a minimum level, and thus be of practical
importance to robotics applications. It is helpful to note that, for blanket, barrier and
static target coverage, a good control law to achieve coverage will eventually reach
steady state once the desired coverage objective has been achieved and the ideal
formation/distribution of the robots has been identified. For sweep and dynamic
target coverage, the control law would generally have to ensure perpetual motion of
the robot groups (Fig. 14.3).
If resource constraint is not an issue, then to guarantee coverage of a region of
interest, a multi-robot system can opt for multiple layers of coverage. A k-layer
14 Area Coverage Algorithms for Networked Multi-robot Systems 305

Fig. 14.3 Agent-based

representation of obstacles:
Hexagonal Wall.
Source: Olfati-Saber (2006)

coverage of a region ensures that each point in the region is covered by at least
k robots. The concept gains increased significance in barrier coverage problems
where 1-barrier has a perimeter guarded by one layer of robots, whereas k-barrier
is k layers of strings of robots – making it more difficult for an intruder to penetrate
the barrier.
Sensing capability of a robot is normally assumed to be omnidirectional
(isotropic sensing), where the sensor strength is maximal at its own position and
fades out radially. Although this is a good enough model for some real-life sensors
such as temperature sensor or circular laser scans, experimental robotics projects
generally utilize sensors such as cone-shaped lasers, monocular cameras or the
simple ultrasonic sensor, all of which are better modeled using directional sensor
regions, an example of which can be found in Tao and Wu (2015).
In the coverage problem dealt within wireless sensor network community,
two parameters of concern exist – adequate coverage and adequate connectivity.
Coverage is associated with the sensing capability of a node, while connectivity is
associated with the communication capability of the node. Connectivity denotes the
state of robot being in communication range of another robot. In WSN, connectivity
is of high importance, as generally information is passed from one node to the base
(or sink node) through its connected nodes. In fact, to ensure connectivity is not
compromised, many algorithms emphasize on having k-connectivity, which denotes
having the state of one node lying within the communication range of at least k other
nodes. In the coverage problems dealt within the robotics community, generally
connectivity can be given less importance at point if we assume information picked
by any robot in the system can be directly communicated to the base or the human-
in-the-loop. Hence we can focus on achieving adequate coverage only.
We focus on area or blanket coverage algorithms in this chapter as it is an
intriguing topic with a plethora of robotics applications, having much exciting work
already done, yet with many open problems still present.
306 L. Iftekhar et al.

14.2.2 Existing Algorithms for Area Coverage

Various control laws have been proposed which can aid a multi-robot system spread
over an area in an intelligent manner. We limit our discussion to distributed control
laws as they are more appropriate for practical robotics application without being
cost-prohibitive (in terms of energy, computation and consequently, money). There
are numerous highly cited survey papers on coverage from the wireless network
perspective (e.g. Ghosh and Das 2008, Mulligan and Ammari 2010, Fazli 2010,
Munishwar and Abu-Ghazaleh 2013), but we would like to provide an updated
collection of work, especially from the field of robotics.
The simplest method to deploy a group of robots over an area is to spread
them out in a random distribution. But this does not guarantee proper coverage
and performance is highly dependent on initial conditions. Some researchers have
taken the next logical step and used probabilistic distributions spatially to control
deployment of robots for area coverage (Chang et al. 2017; Fan et al. 2015; Prakash
and Patidar 2014). For example, authors in Prakash and Patidar (2014) have utilized
three distributions – Beta, Gamma and Fisher-Snedecor on large group of sensors,
demonstrating that coverage of a region of interest is easily achieved (the amount of
uncovered area gets very close to zero), the last distribution yielding the quickest
results. To calculate the amount of uncovered area, their performance metric, the
authors used two approaches, grid-based and disc-based. In grid-based approach,
area coverage is achieved when each single cell of the entire grid has at least one
sensor, whereas in disc-based approach, the coverage task is complete if only each
and every point is covered by a sensor, making the latter metric a more difficult one
to fulfill. A drawback to these probabilistic approaches would be the assumption that
the distribution is made over a square or rectangular box overlay above the region
of interest, and unless other explicit constraining laws are introduced, many of the
sensors or robots get automatically placed outside the region of interest, wasting
precious resources.
Partitioning a region of interest in Voronoi cells assigned to each robot and
updating the robot’s state using some distributed control law has been a popular
approach (see references in Mulligan and Ammari 2010). Authors in Yang et al.
(2014) took it a step further and divided the area of interest in Voronoi cells
dynamically and applied modified bacterial foraging optimization to move the
robots accordingly. An excellent example of Voronoi based application on a real
multi-robot system can be found in the works of Lei Ma such as in Han et al. (2017),
where multiplicatively-weighted Voronoi diagrams are used so that a group of
heterogenous robots, with varying physical design and capabilities can perform area
coverage utilizing each different robot’s highest potential. This work as well as Han
et al. (2017) and Zhang et al. (2016) implemented their area coverage algorithms on
two wheeled differential robots as well as four wheeled car-like robots, in contrast
to most works on multi-robot area coverage problems where the focus is mostly on
algorithmic level, and hence on theoretical robots with particle dynamics.
14 Area Coverage Algorithms for Networked Multi-robot Systems 307

Assuming a virtual geometric configuration akin to a lattice-like structure

covering the region of interest and figuring out a control strategy to get the robots
situated at the vertices is an underlying motivation for a huge body of grid-based
and computational geometry-based approaches for area coverage. Earlier works
(Cortes et al. 2005) devised protocols to make the robots or sensors form rectangular
grids, e.g. authors in Wang et al. (2007) utilized grid structure over an area and
divided sensors into clusters based on distance. Hexagonal and rhombus based
topologies exist too but the equilateral triangle topology is the optimal pattern
for minimum overlapping of sensing region and maximum connectivity with six
neighbors (Gupta et al. 2016). A major positive feature of our coverage algorithm
is that it automatically results in the robots forming triangular lattices without any
explicit command.
The control laws used to achieve the above coverage formations are often force-
based, capitalizing on attraction and replusion forces between neighboring robots.
Nearest neighbor techniques are used in Santoso (2011) in which the algorithm
was designed for airdrop deployment of robots which will perform a square-grid
area coverage upon release. Compared to some of the results in the previous works,
where a few robots in the group extends beyond the region of interest, in works such
as Cheng and Savkin (2010) and Santoso (2011), the robots themselves manage to
stay within the ROI, but the ones at the outer perimeter situate themselves at the
border, wasting sensing resources. Our algorithm to be described later, attempts to
keep not only the robots, but even their sensing regions within the region of interest,
to make optimal use of all resources. In the past few years, consensus algorithms
have been an increasingly popular choice to achieve robust and stable steady state
for area coverage problems in finite time, such as algorithms in Cheng and Savkin
(2010, 2013) and Nasimov and Matveev (2014), all of which achieve asymptotically
optimal triangular blanket coverage.
Desired characteristics of an ideal area coverage protocol From the above dis-
cussion, we identify our objective as to devise simple rules to coordinate the motion
of the robots that would result in intelligent and powerful control law to achieve
the desired coverage. We also identify some desired characteristics of an ideal area
coverage protocol for robotics applications: the final distribution should result in
triangular lattice formation, achieve steady state quickly, minimize overlapping of
sensing region (to cover more space using fewer robots and avoid redundancy and
wastage of computational and sensing resources) and be implementable of real-life
robot dynamics. Moreover, all robots should stay within the boundary of region of
interest. Two more criteria may be added: it is preferred that the robots can work in
a previously unknown environment (many of the prior algorithms only work when
the sensors know of the boundaries beforehand) and eventually be able to complete
the task of area coverage in the presence of obstacles efficiently.
Our flocking-based area coverage algorithm described in the remaining sections
of this chapter is related to the consensus-based protocols and employs nearest
neighbor rules, but using very simple laws. Compared to other consensus protocols,
the parent algorithm here by Olfati-Saber (2006) is highly robust, not suffering
308 L. Iftekhar et al.

from pitfalls such as defragmentation of an established structure over a period

of time, or collapse of two neighboring robots into one singular space. It is also
easily extendable to two-wheeled and four-wheeled robot dynamics (Olfati-Saber
and Iftekhar 2012), and so is the proposed algorithm here, making our algorithm a
perfect candidate to be implemented in real-life robots.

14.3 Flocking-Based Area Coverage Algorithms

In this section, a set of area coverage algorithms based on Olfati-Saber’s flocking

algorithms (Olfati-Saber 2006) are described. We start with the description of the
multi-robot systems, assuming particle dynamics, to establish the terminology. The
work in this chapter is easily extendable to two-wheeled robots with non-linear
dynamics (Olfati-Saber and Iftekhar 2012).

14.3.1 Preliminaries

The Area Coverage algorithm we developed is heavily based on Reza Olfati’s

(Olfati-Saber 2006) Flocking Algorithms. The concept and rules of these particular
flocking algorithms have never been applied to solve the problem of area coverage
and hence the application of flocking presented in our work is a novel one. In this
chapter, we present the background of flocking, and the notations and equations for
the existing Flocking Algorithm explained in Olfati-Saber (2006).
Flocking is a form of collective behavior of a large number interacting agents,
which includes a cooperative movement to stay in close proximity and maintain
a coveted formation while moving. It is a behavioral trait that can be observed
in multi-agent systems in nature, demonstrating a strongly adaptive capability to
achieve a common objective. Examples of these agents include birds, fish, ants,
bees, people, robots, etc. where each entity or members of flocks are referred as an
α-agent in this chapter (following the terminology of Olfati-Saber 2006).
In 1986, Craig Reynolds proposed three heuristic rules for flocking which
are Flock Centering, Obstacle Avoidance and Velocity Matching, also known as
cohesion, separation and alignment respectively. Numerous studies have attempted
to give these heuristic rules an algorithmic face, but most of them suffered
from drawbacks and could not capture the smooth flocking behavior shown in
nature. However, Olfati-Saber’s flocking algorithms in Olfati-Saber (2006), finally
provided a robust mathematical protocol to achieve the dynamic collective behavior
researchers have been aiming for, without the common pitfalls of a formed flock
falling apart over time (fragmentation) or the nodes inside the flock colliding with
each other (collapse). They could also perform well in the presence of obstacles and
successfully demonstrated splitting, rejoining and squeezing maneuvers.
14 Area Coverage Algorithms for Networked Multi-robot Systems 309

Flocking has also been applied on achieving autonomous driving (Iftekhar 2012),
in aerial vehicle formation (Crowther 2002) and target seeking (Lindhe et al. 2005).
However it has not been used explicitly for area coverage yet, to the best of our
knowledge. The unique lattice-formation capabilities of the highly scalable flocking
algorithms in Olfati-Saber (2006) (that neither experience fragmentation nor have
the nodes collapse upon each other) make them an excellent potential base to
develop area coverage algorithms, which we do in this chapter.
We now provide some background and notations on Olfati-Saber’s flocking
algorithm (Olfati-Saber 2006). As we have mentioned before, the agents performing
flocking behavior are denominated as α-agents.
A group of n α-agents moving in m-dimensional space are modeled by the
following equations:

q˙i = pi
(14.1)
ṗi = ui

where qi , pi , ui ∈ R m represents the position, velocity and control input respec-

tively of ith agent. The systematic configuration of all the α-agents in a system is
denoted by the vector equation q = col(q1 , ...qn ) ∈ R mn . An α-agent is assumed
to have a sensing and communication radius of r > 0 and if any other α-agent
enters that disc, they can interact with each other and are said to be neighbors. This
two-way neighbor relationship for all the agents in a system can be represented by
a graph where agents are the nodes V = {1, 2, ...n} and the interaction between
neighbors are represented by edges. The set of neighbors of α-agent i can thus be
denoted by:

Ni (q){j : (i, j ) ∈ ε(q)} (14.2)

2 2
where ε(q) = {(i, j ) ∈ V × V : 2qj − qi 2 ≤ r} is a set of boundaries and V =
{1, 2, ...n}. where . is the Euclidean norm. A conformation q is said to form a
quasi α−lattice, if each agent has at least two neighbors and the distances between
neighboring agents are approximately equal to a desired distance d i.e.
2 2
∃, d > 0 : −  ≤ 2qj − qi 2 − d ≤ , ∀j ∈ Ni (q) (14.3)

A quasi α-lattice is an α-lattice when  = 0. Moreover,

2 2 the attraction or repulsion
between the agents depends on the variation of 2qj − qi 2
Once a group of α-agents have created a quasi α-lattice, they can be considered
to be demonstrating flocking behavior.
Olfati-Saber also introduces two more agents. The first one is a virtual γ agent,
which incorporates the objective of the flock enabling them to move from one place
to another without fragmentation. The second one is also a virtual agent, termed β-
agent, which each α-agent creates on an obstacle, when the obstacle falls within the
α-agent’s sensing region, to maintain appropriate distance from it. Hence to achieve
310 L. Iftekhar et al.

flocking according to Olfati-Saber (2006), each α-agent is subjected to the following

control input:

β γ
ui = uαi + ui + ui (14.4)

where uαi refers to the (α, α) interaction term used to keep the agents in a lattice
β
form, and uj refers to the (α, β) interaction term used for obstacle avoidance, where
the virtual β-agent is the nearest point on the obstacle from the α-agent.
Two scalable flocking protocols for α-agents have been used in uαi , a gradient-
based one and a velocity consensus protocol.
2 2
uαi = φα (2qj − qi 2σ )nij + aij (q)(pj − pi ) (14.5)
j ∈Ni j ∈Ni

The first algorithm, uαi represents all three features of Reynolds, i.e. the gradient-
based term comprises flock centering and collision avoidance, and the consensus
term is equivalent to velocity matching.
 2 22
nij = (qj − qi )/ 1 +  2qj − qi 2 (14.6)

nij denotes an enclosed vector along the line connecting qi and qj , and

1
sσ = ( 1 + s2 − 1) (14.7)


denotes the σ -norm of vector s and  ∈ (0, 1). The group has a potential function
defined as V (q):

1 2 2
V (q) = ψα (2qj − qi 2σ ) (14.8)
2
i j =i

-s
where ψα = d φα (h)dh. The scalar function φα (s) is defined as follows:

φα (s) = ρh (s/rσ )φ(s − dσ ) (14.9)

1
φ(s) = [(a + b)σ1 (s + c) + (a − b)]
2
s
where σ1 (s) = √ . φ(s) depicts an uneven sigmoidal function having
1 + s2 √
parameters that satisfy the conditions 0 < a ≤ b and c = |a − b|/ 4ab to ensure
φ(s) = 0. ρh (s), on the other hand, is a scalar function that varies in between
the range of 0 and 1. The purpose of this function is to construct smooth potential
14 Area Coverage Algorithms for Networked Multi-robot Systems 311

functions with finite cut-offs and smooth adjacency matrices. In order to do that, the
following are used:
3
3 1 ( if s ∈ [0, h]
3  )
3 1 s−h
ρh (s) = 33 1 + cos π if s ∈ [h, 1]
3 2 1−h
3 0 otherwise

where h ∈ (0, 1).

Using the algorithm discussed above, one can achieve α-agents forming beautiful
lattice structures.

14.3.2 Area Coverage Algorithm in an Unknown Environment

A swarm of α-agents instructed to move over an area while ensuring as much

of the area as possible falls under the sensing region of at least one α-agent is
called area coverage. In this chapter we used the concept of Flocking from Olfati-
Saber (2006) to optimally cover a given area and minimize the elapsed time for the
coverage without human intervention. The Area Coverage problem addressed here
is particularly defined as follows: The multi-agent system finds stationary positions
to assume so that the agents can optimally cover the given area using their sensing
regions at given point of time. Once the agents reach a steady state/optimal position,
they assume a stationary position.
We have slightly modified and implemented Olfati-Saber’s (2006) flocking
algorithm by using two agents only: α and β and ignoring the γ -agent. The
algorithm allows α-agents to optimally cover the whole area, move in clusters and
avoid any obstacles/boundaries which are encountered. In the new algorithm, each
α-agent is subjected to the following control law:

β
ui = uαi + ui (14.10)

Here, the first term uαi is as explained in Sect. 14.3.1.

The second term of the algorithm is as follows:
2 2
φβ (2q̂i,k − qi 2σ )n̂i,k +
β
ui = bi,k (q)(p̂i,k − pi ) (14.11)
β β
k∈Ni k∈Ni

where bi,k represents a heterogeneous adjacency between an α-agent at qi and its

neighboring β-agent at q̂i,k , and is denoted by:
2 2
bi,k (q) = ρh (2q̂i,k − qi 2σ /dβ ) (14.12)
312 L. Iftekhar et al.

In addition, the φβ in Eq. (14.11), constitutes a repulsive action f unction, that

is defined as φβ (z) = ρh (z/dβ )(σ1 (z − dβ ) − 1), and the vector n̂i,k is given by
q̂i,k − qi
n̂i,k =  2 22 .
2
1 +  q̂i,k − qi 2
The obstacles in the territory are represented through virtual agents called β-
agents, that move on the boundary of the obstacles and correspond the nearest
distance between them and the α-agents. If the α-agents come in close contact with
the β-agents, repulsion takes place, and hence, any collision with the obstacles is
avoided.
Olfati-Saber (2006) mentions two different obstacles: hyperplane boundaries
(walls) and spherical boundaries. To develop the algorithm for area coverage, we
work with the m-dimensional space of an area, i.e. the hyperplane boundaries. We
apply the concept of individual hyperplane boundaries to create walls which form
an enclosed edge of the area to be covered by the α-agents. When any one of the
walls is within the sensing range of an α-agent, a virtual β-agent is formed on that
particular wall. This virtual β-agent then exerts a repulsive force on the α-agent,
ensuring that the α-agent remains within the boundaries of the coverage area and
ensures that there are no collisions between the α-agents and the walls. The position
and velocity of a β-agent on a wall is denoted by q̂i,k and p̂i,k respectively, generated
by an α-agent with state (qi , pi ) on a given obstacle, such as Ok . Subsequently, the
following applies:
• Consider an obstacle with a hyperplane boundary. Assume that is has a unit
normal ak , and passes through the point yk . Then position and velocity of the
β-agent is as follows:

q̂i,k = P qi + (I − P )yk
(14.13)
p̂i,k = Ppi

where P = I − ak akT is a projection matrix.

γ
The third part of Olfati-Saber’s flocking algorithm, ui , which is used for the
movement of agents from one point to another in a m-dimensional space, is omitted
as our primary goal is optimal area coverage. Instead of specifying predetermined
coordinates and directing the α-agents to move from point A to point B, we directed
our focus specifically on optimal coverage of an area. We applied the concept of a
multi-robot system getting deployed in the targeted area first, then apply flocking to
move around in cluster and lattice-like form, eventually covering the entire territory.

14.3.3 Area Coverage Algorithm with Obstacle Avoidance

Here, we show multi-robot system with obstacle avoidance capabilities. In the last
section, two different kinds of obstacles were mentioned, out of which we used the
concept of hyperplane obstacles to set the basics of our area coverage algorithm. For
14 Area Coverage Algorithms for Networked Multi-robot Systems 313

Fig. 14.4 Agent-based representation of obstacles: Approximate model of arbitrarily-shaped

obstacle using spheres

this section, our objective is to upgrade this algorithm to make it capable of letting
the robots in charge of the coverage task be able to avoid any sort of obstacles within
the region of interest, while they themselves are getting distributed over the region.
For that, we utilize the model of spherical obstacles – the second kind of obstacles
from Olfati-Saber (2006).
To achieve the above, the α-agents now also keep an eye out for spherical
obstacles by being able to project a new kind of β-agents which is defined below.
• Consider a spherical obstacle centered at yk and having a radius Rk . μ Then the
position and velocity of the β-agent are as follows:

q̂i,k = μqi + (1 − μ)yk

(14.14)
p̂i,k = μPpi

Here P = I − ak akT , μ = Rk / qi − yk  and ak = (qi − yk )/ qi − yk 

The strength of such spherical obstacles is that, multiple small versions of them
can be used to make an approximate model of many kinds of arbitrarily shaped
obstacles (Fig. 14.4).

14.4 Simulation Results

In this section, we present simulation results for Area Coverage by a multi-robot

system, where each robot is modeled as an α-agent, for three different environments:
square, pentagonal and hexagonal boundaries. The following parameters remain
314 L. Iftekhar et al.

Table 14.1 Total final Type of boundary Final covered area

coverage in three different
boundary structures Square 98.29%
Pentagon 99.03%
Hexagon 96.73%

fixed throughout the simulations that have been carried out: the minimum distance
to maintain between α-agents is d = 7, the sensing radius of α-agents is r = 1.2d,
the minimum distance to maintain between α and β agents is d = 8, the sensing
radius of β-agents is r = 1.5d , a = 5 and b = 4.5 for φ(s) and h = 0.15 for the
bump functions φα (s) and φβ (s). Initial conditions that are set in each case include:
i. The number of agents n
ii. The initial position of the agents q
iii. The initial velocity of the agents p
In the case of a square boundary, initial position coordinates are chosen at
random from the box [0, 21]2 and velocity coordinates are chosen randomly within
the box [0, 3]2 . For a pentagonal boundary the initial conditions are set such that
initial position coordinates are chosen at random within the box [0, n]2 and velocity
coordinates are chosen at random from the box [0, 5]2 . In the event of the hexagonal
boundaries the initial position coordinates are also chosen at random from the box
[0, n]2 and velocity coordinates are chosen at random within the box [0, 5]2 .
The following figure shows steady state snapshots from different region of
interests with illustrative boundaries:
Table 14.1 projects an approximate percentage area coverage for the three afore-
mentioned boundaries. Coverage is calculated manually using geometric analysis of
the two-dimensional space. It is seen that the area coverage of the agents in any given
area is at least 96%, thereby making it a highly efficient area coverage algorithm.
Figure 14.5 gives a series of snapshots show the simulation results for the
pentagonal boundaries.
To demonstrate the performance of our area coverage algorithm in the presence
of obstacles (as described in Sect. 14.3.3), we run another set of simulations with
different kinds of convex obstacles (Fig. 14.6). Here we present the snapshots of a
single circular obstacle in an area of interest with hexagonal boundary.1
It is worth mentioning that once steady state is reached the lattice formation
covering the ROI becomes fully still and no oscillation is observed no matter how
long the simulation is run, demonstrating stability of the system.

1 Stepsize for the simulation in Fig. 14.7 was different than the one used in obstacle-free
environment simulations.
14 Area Coverage Algorithms for Networked Multi-robot Systems 315

Fig. 14.5 Final coverage by multi-robot systems applying the proposed algorithm in Sect. 14.3.2
in different environments. The boundary structure is not known to the robots. The green disks are
the robots’ sensing regions. (a) Square ROI. (b) Pentagonal ROI. (c) Hexagonal ROI

14.5 Conclusion

In this chapter, we have explored briefly the task of area coverage by a group of
mobile robots and demonstrated the performance of some cooperative area coverage
algorithms. Our flocking-based algorithms are decentralized control laws which
enable the mobile robots to autonomously form a lattice-like coverage over a
region of interest. The region of interest could be of any arbitrary convex shape
and the robots do not need prior knowledge of the boundaries. Implementation of
the algorithm results in quick distribution of robots in triangular lattices, ensuring
maximum coverage and minimal sensing region overlap, as well as attempting to
limit extension of robot sensing region beyond the boundary of the area of interest,
thus offering more intelligent resource usage. Most significantly with very minimal
computational addition, the robots can also avoid any arbitrarily shaped obstacles
within the region of interest and distribute themselves in optimal configurations
in the presence of any number of obstacles, once again, not wasting any sensing
resource by being too close to the obstacles or other non-spaces.
We now provide some future directions and discuss some aspects off the area
coverage algorithms that still require more research on.
1. Heterogeneous sensing regions: We have only used a specific circular sensing
region with a fixed radius for all the α-agents in this chapter. Future work involves
implementing various sensing regions for α-agents of different shapes and sizes,
thereby giving more freedom to explore and increase the efficiency of the α-
agents during area coverage. This would enable a multi-robot group to utilize
different kinds of robots available with varying capabilities. For more practical
applications, algorithms that can handle robots directional sensing regions should
be focused upon, which would also reduce the issue of overlapping caused by
316 L. Iftekhar et al.

Fig. 14.6 A snapshot of area coverage by 30 robots in a pentagonal region. The robots have
no prior knowledge of the boundary. They are initially distributed randomly within the region
of interest. They spread out, detect boundary and distribute themselves quite uniformly over the
region of interest. (a) t = 0.0 (sec) (b) t = 4.03 (sec) (c) t = 8.91 (sec) (d) t = 11.12 (sec) (e) t = 14.08
(sec) (f) t = 18.02 (sec)
14 Area Coverage Algorithms for Networked Multi-robot Systems 317

Fig. 14.7 A snapshot of area coverage by 30 robots in a hexagonal region with a large obstacle.
The robots have no prior knowledge of the boundary or the obstacle. They are initially distributed
randomly within the region of interest. They spread out, detect boundary and obstacle and distribute
themselves quite uniformly over the region of interest. They maintain adequate distance from the
boundary and obstacle and thus do not waste sensing region over non-spaces or beyond region of
interest. (a) t = 0 (sec) (b) t = 35 (sec) (c) t = 70 (sec) (d) t = 105 (sec) (e) t = 140 (sec) (f) t = 179
(sec)
318 L. Iftekhar et al.

isotropic sensing. One recent attempt towards that direction can be found (Santos
et al. 2018) in where different sensing regions are modeled as different density
functions.
2. Adaptive configuration in dynamic environment: Our algorithm and most oth-
ers have a predefined desired distance between neighboring robots. Algorithms
which can give stronger and more definite ability for the robots to re-configure
and adapt to a smaller region of interest (in the event of the number of deployed
robot being higher than required) or slightly larger region (in case of sudden loss
of few robots). Devising a control law for multi-robot system in an environment
with moving obstacle and/or moving boundaries is still an open problem.
3. Implementation in 3D space and on complex robot dynamics: Most of the
existing coverage algorithms have been demonstrated in 2D space, barring some
exceptions. Ours is easily extendable to 3D space as its base algorithm has been
demonstrated to work well in three-dimensional regions (Olfati-Saber 2006).
Although most algorithms, including ours have been numerically simulated for
robots with particle dynamics, ours can also be easily extended to differential
drive robots (Olfati-Saber and Iftekhar 2012) and four-wheeled robots. Some
works do exist which apply other coverage algorithms to multi-robot systems
with complex robot dynamics, for example, the previously cited Zhang et al.
(2016) and Han et al. (2017) as well as the work in Varga et al. (2015) where
area coverage algorithm is implemented on fixed wing aerial vehicles, yet more
research needs to be done to demonstrate proof of stability for such systems.

References

Adaldo, A., Mansouri, S. S., Kanellakis, C., Dimarogonas, D. V., Johansson, K. H., & Niko-
lakopoulos, G. (2017). Cooperative coverage for surveillance of 3D structures. In 2017
IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), Vancouver, BC,
Canada (pp. 1838–1845).
Banimelhem, O., Naserllah, M., & Abu-Hantash, A. (2017). An efficient coverage in wireless
sensor networks using fuzzy logic-based control for the mobile node movement. In 2017
Advances in Wireless and Optical Communications (RTUWO), Riga, Latvia (pp. 239–244).
Chen, Y., Zhang, H., & Xu, M. (2014). The coverage problem in UAV network: A survey. In
Fifth International Conference on Computing, Communications and Networking Technologies
(ICCCNT), Hefei, China (pp. 1–5).
Chang, G. Y., Charng, C. W., Sheu, J. P., & Ruei-Yuan, L. (2017). Probabilistic k-weighted
coverage placement in wireless sensor networks. In 19th Asia-Pacific Network Operations &
Management Symposium (APNOMS), Seoul, Korea (pp. 382–385).
Cheng, T. M., & Savkin, A. V. (2010). Decentralized control of mobile sensor networks for
asymptotically optimal self-deployment in blanket coverage. In 49th IEEE Conference on
Decision and Control (CDC), Atlanta, Georgia, USA (pp. 1257–1262).
Cheng, T. M., & Savkin, A. V. (2013). Decentralized control of mobile sensor networks for
asymptotically optimal blanket coverage between two boundaries. IEEE Transactions on
Industrial Informatics, 9(1), 365–376.
Cortes, J., Martinez, S., & Bullo, F. (2005). Spatially-distributed coverage optimization and control
with limited-range interactions. ESAIM: Control, Optimisation and Calculus of Variations,
14 Area Coverage Algorithms for Networked Multi-robot Systems 319

11(4), 691–719.
Crowther, B. (2002). Rule-based guidance for flight vehicle flocking. Proceedings of the Institute
of Mechanical Engineers, Part G: Journal of Aerospace Engineering, 218(2), 111–124.
Fan, X., Yang, J., & Wang, C. (2015). An effective construction scheme of probabilistic k-barrier
coverage in WSN. In IEEE International Conference on Cyber Technology in Automation,
Control, and Intelligent Systems (CYBER), Shenyang, China (pp. 671–676).
Fazli, P. (2010). On multi-robot area coverage. In 9th International Conference on Autonomous
Agents and Multiagent Systems. AAMAS’10 (Vol. 1, pp. 1669–1670). International Foundation
for Autonomous Agents and Multiagent Systems, Toronto, Canada.
Gage, D. W. (1992). Command control for many-robot systems. Naval Command Control and
Ocean Surveillance Center RDT and E Div, San Diego, CA, Technical Report
Ghosh, A., & Das, S. K. (2008). Coverage and connectivity issues in wireless sensor networks: A
survey. Pervasive and Mobile Computing, 4(3), 303–334. [Online]. Available:
http://www.sciencedirect.com/science/article/pii/S1574119208000187.
Habib, M. K. (2017). Coordinated multi robotic system for demining activities. In The First
International Conference on Landmine: Detection, Clearance and Legislations (LDCL2017),
Faculty of Engineering, Lebanese University, Campus of Hadath, Beirut, Lebanon (pp. 1–5).
Hameed, I. (2014). Intelligent coverage path planning for agricultural robots and autonomous
machines on three-dimensional terrain. Journal of Intelligent and Robotic Systems, 74(3–4),
965–983.
Han, X., Ma, L., & Feng, K. (2017). Distributed coverage control of networked heterogeneous
robots. In 6th Data Driven Control and Learning Systems (DDCLS), Chongqing, China (pp.
286–291).
Gupta, P., Prakash, V., & Suman, P. (2016). Noticeable key points and issues of sensor deployment
for large area wireless sensor network: A survey. In 2016 International Conference System
Modeling Advancement in Research Trends (SMART), Moradabad, India (pp. 164–169).
Iftekhar, L., & Olfati-Saber, R. (2012). Cooperative autonomous driving for vehicular networks.
In 9th International Conference on Informatics in Control, Automation and Robotics ICINCO,
Rome, Italy (Vol. 2, pp. 345–352). SciTePress.
Janani, A.-R., Alboul, L., & Penders, J. (2016). Multi robot cooperative area coverage, case study:
Spraying. In Towards Autonomous Robotics Systems: 17th Annual Conference TAROS 2016,
Sheffield, UK (Proceedings of the lecture notes in computer science, Vol. 2716, pp. 165–176).
Springer.
Lindhe, M., Ogren, P., & Johansson K. H. (2005). Flocking with obstacle avoidance: A new dis-
tributed coordination algorithm based on Voronoi partitions. In IEEE International Conference
on Robotics and Automation, Barcelona, Spain (Vol. 2, pp. 1785–1790).
Mahamuni, C. V. (2016). A military surveillance system based on wireless sensor networks
with extended coverage life. In 2016 International Conference on Global Trends in Signal
Processing, Information Computing and Communication (ICGTSPICC), Jalgaon, India (pp.
375–381).
Mulligan, R., & Ammari, H. M. (2010). Coverage in wireless sensor networks: A survey. Network
Protocols & Algorithms, 2(2), 27–53. [Online]. Available: https://doi.org/10.5296/npa.v2i2.
276.
Munishwar, V. P., & Abu-Ghazaleh, N. B. (2013). Coverage algorithms for visual sensor networks.
ACM Transactions on Sensor Networks, 9(4), 1–45.
Nasimov, M., & Matveev, A. (2014). Suboptimal decentralized blanket coverage control of mobile
autonomous sensor networks. In 6th International Congress on Ultra Modern Telecommunica-
tions and Control Systems and Workshops (ICUMT), St. Petersburg, Russia (pp. 366–371).
Olfati-Saber, R. (2006). Flocking for multi-agent dynamic systems: Algorithms and theory. IEEE
Transactions On Automatic Control, 51(3), 401–420.
Olfati-Saber, R., & Iftekhar, L. (2012). Flocking for networks of mobile robots with nonlinear
dynamics. In 9th International Conference on Informatics in Control, Automation and Robotics
ICINCO, Rome, Italy (Vol. 2, pp. 353–359). SciTePress.
320 L. Iftekhar et al.

Piciarelli, C., Esterle, L., Khan, A., Rinner, B., & Foresti, G. L. (2016). Dynamic reconfiguration
in camera networks: A short survey. IEEE Transactions on Circuits and Systems for Video
Technology, 26(5), 965–977.
Prakash, S. K. L. V. S., & Patidar, N. (2014). Area based coverage for three deployment
distributions and their connectivity in wireless sensor networks. In 4th International Conference
on Communication Systems and Network Technologies, Bhopal, India (pp. 128–132).
Santos, M., Diaz-Mercado, Y., & Egerstedt, M. (2018). Coverage control for multirobot teams with
heterogeneous sensing capabilities. IEEE Robotics and Automation Letters, 3(2), 919–925.
Santoso, F. (2011). A decentralised self-dispatch algorithm for square-grid blanket coverage
intrusion detection systems in wireless sensor networks. In 2011 IEEE Vehicular Technology
Conference (VTC Fall), San Francisco, CA, USA (pp. 1–5).
Tao, D., & Wu, T. Y. (2015). A survey on barrier coverage problem in directional sensor networks.
IEEE Sensors Journal, 15(2), 876–885.
Varga, M., Basiri, M., Heitz, G., & Floreano, D. (2015). Distributed formation control of fixed
wing micro aerial vehicles for area coverage. In 2015 IEEE/RSJ International Conference on
Intelligent Robots and Systems (IROS), Hamburg, Germany (pp. 669–674).
Wang, Y. H., Yu, C. Y., & Fu, P. F. (2007). A coverage and connectivity method to cluster topology
in wireless sensor networks. In 21st International Conference on Advanced Information
Networking and Applications Workshops, AINAW’07, Niagara Falls, ON, Canada (Vol. 1, pp.
97–102).
Xin, B., Gao, G. Q., Ding, Y. L., Zhu, Y. G., & Fang, H. (2017). Distributed multi-robot motion
planning for cooperative multi-area coverage. In 13th IEEE International Conference on
Control & Automation (ICCA), Ohrid, Macedonia (pp. 361–366).
Yang, B., Ding, Y., & Hao, K. (2014). Target searching and trapping for swarm robots with
modified bacterial foraging optimization algorithm. In 11th World Congress on Intelligent
Control and Automation, Shenyang, China (pp. 1348–1353).
Yang, B., Ding, Y., & Hao, K. (2015). Area coverage searching for swarm robots using dynamic
voronoi-based method. In 34th Chinese Control Conference (CCC), Hangzhou, China (pp.
6090–6094).
Yuming, L., Ruchun, W., Zhenli, Z., & Junlin, Z. (2011). Dynamic coverage path planning and
obstacle avoidance for cleaning robot based on behavior. In International Conference on
Electric Information and Control Engineering, Wuhan, China (pp. 4952–4956).
Zhang, J., Zhou, P., & Ma, L. (2016). Coverage control of multiple heterogeneous mobile robots
with nonholonomic constraints. In 35th Chinese Control Conference (CCC), Chengdu, China
(pp. 6272–6277).
Zheng, X., Koenig, S., Kempe, D., & Jain, S. (2010). Multirobot forest coverage for weighted and
unweighted terrain. IEEE Transactions on Robotics, 26(6), 1018–1031.
Chapter 15
Backstepping-Based Nonlinear RISE
Feedback Control for an Underactuated
Quadrotor UAV Without Linear Velocity
Measurements

Jawhar Ghommam, Luis F. Luque-Vega, and Maarouf Saad

Abstract In this paper, a robust three dimensional output feedback control problem
is proposed for a 6-degrees-of-freedom model of a quadrotor unmanned aerial
vehicle (UAV) to track a bounded and sufficiently smooth reference trajectory in the
presence of slowly varying force disturbances. Due to the underactuation structure
of the UAV, a nonlinear output feedback controller based on the robust integral
of the sign error signal (RISE) mechanism is first designed for the translational
dynamics to ensure position reference tracking without velocity measurement. The
angular velocity is then regarded as intermediate control signal for the rotational
dynamics to fulfill the task of attitude angle reference tracking. The torque input is
designed taking full advantage of the smooth exact differentiator that circumvents
derivatives computation of virtual controls, the backstepping technique is then
judiciously modified to allow the use of the RISE control technique to compensate
for the external disturbances. The proposed controller yields semi-global asymptotic
stability tracking despite the added disturbances in the dynamics. Simulation results
are shown to demonstrate the proposed approach.

Keywords RISE-backstepping · Exact differentiator · Underactuated quadrotor ·

Output feedback

J. Ghommam ()
Department of Electrical and Computer Engineering, College of Engineering, Sultan Quaboos
University, Muscat, Oman
e-mail: [email protected]
L. F. Luque-Vega
CIIDETEC-UVM, Universidad del Valle de Mexico, Mexico City, Mexico
e-mail: [email protected]
M. Saad
Department of Electrical Engineering, École de Technologie Superieure, Montreal, QC, Canada
e-mail: [email protected]

© Springer Nature Singapore Pte Ltd. 2019 321

N. Derbel et al. (eds.), New Developments and Advances
in Robot Control, Studies in Systems, Decision and Control 175,
https://doi.org/10.1007/978-981-13-2212-9_15
322 J. Ghommam et al.

15.1 Introduction

The last few decades have witnessed quick trends in the field of unmanned
aerial vehicle (UAV), with increasingly affordable devices endowing these types
of vehicles with versatile capabilities to perform many types of tasks in multitude
applications. Most of the tasks envisioned to be within the reach of the UAVs,
involve search and rescue, security, civil engineering inspection and weather
observation (Ilcev and Sibiya 2015; Magsino et al. 2016). In particular, quadrotor
type UAVs have recently captured the attention of the control community since
their characteristics being highly nonlinear dynamics poses an extremely good
challenge to control system designers (Castillo et al. 2004; Cunha et al. 2009;
Huang et al. 2010; Kim and Shim 2003). Moreover, the quadrotor type UAVs are
steadily becoming instrumental tool for real life missions, ranging from surveillance
to entertainment.
Given the difficulties encountered with underactuation and strong nonlinear cou-
pling in the quadrotor dynamics, some investigations have explored various control
solutions for this type of aircraft vehicle. Earlier attempts to achieve acceptable
autonomous flight control have been developed based on linear control design and
they basically involve PID controllers (Garcia et al. 2012; Salih et al. 2010), LQR
controllers (Rinaldi et al. 2013) and H∞ controllers (Ortiz et al. 2016). Of special
significance, the input/output linearization approach has been used in Ciulkin et al.
(2015) and Garcia et al. (2012) to design controllers for the stabilization and
trajectory tracking of the quadrotor. Of particular note, the quadrotor aircraft often
operates in three dimensional space. This limits the control design proposed in
the aforementioned work which restricts the control of the quadrotor in a vertical
plane only. Just recently the authors in Do (2015) and Ghommam et al. (2015),
proposed new schemes that give ability to the designer of controlling the quadrotor
in six degrees of freedom. In particular, the authors in Kendoul (2009), Wang and
Jia (2014) and Van Den Eijnden (2017) proposed a cascade design procedure for
the position control of a rotary aircraft UAV. In Bertrand et al. (2011) the authors
proposed a hierarchical design procedure for an UAV vehicle based on the singular
perturbation theory. Similar approach on using the hierarchical design procedure
has been presented in Yao (2017) and Zou (2017).
In almost all the aforementioned works, the control designs require most often
instantaneous measurements of the linear velocity of the quadrotor UAV. As there
is no sensor which directly measures the linear velocity of the quadrotor in the
frame body, those controllers require some kind of observer relying on the global
positioning sensor. The global position tracking problem of VTOL UAVs without
linear-velocity measurement has been addressed in Abdessameud and Tayebi
(2010), where the authors exploited the natural cascaded structure of the UAV to
design their controller based on thrust and desired attitude extraction algorithm.
In Lee et al. (2007), the authors proposed a an adaptive output feedback scheme
based on an exponential observer to estimate the velocities for a tracking control of
quadrotor UAV using only linear and angular positions. In Madani and Benallegue
15 Backstepping-Based Nonlinear RISE Feedback Control for an. . . 323

(2007) and Benallegue et al. (2008), sliding-mode observers were used to estimate
the effect of external perturbations using measurement of positions and yaw angle.
Although the nonlinear controllers realized along with the observer design to esti-
mate the linear velocity of the quadrotor can achieve the desired performance of the
system, but the resulting performance in the presence of structural uncertainties and
external disturbances can only achieve ultimately boundedness and not asymptotic
stability tracking.
In this paper, a nonlinear backstepping controller is designed concurrently with a
filter that estimates the linear velocity of the quadrotor vehicle. The robust integral
of the sign of the error (RISE) control structure is introduced in the design to
compensate for uncertain disturbances affecting the translational dynamics. The
resulting desired translational commands are then utilised as attitude commands
in the backstepping approach, in that the virtual force command is backstepped
through an exact differentiator to the rotational dynamics. The backstepping tech-
nique is then judiciously modified to allow the use of the (RISE) control structure in
order to compensate for disturbances. A Lyapunov based stability analysis is used to
prove that the proposed controller guarantees semiglobal asymptotic tracking. The
novelty of this approach is that the trajectory tracking controller is designed based
on the imbrication of the RISE feedback technique and makes use of the position
measurement only for the stability of the translational dynamics. Information about
the derivatives of the force commands being used as virtual command to the
rotational dynamics are estimated through exact differentiator through the super
twisting algorithm (Levant 2005).
The remainder of this paper is as follows: Sect. 15.2 provides mathematical
preliminaries. Section 15.3 presents the problem statement of this paper. Sec-
tion 15.4 details the main control development, Sect. 15.5 performs some numerical
simulation to validate the proposed approach and finally, Sect. 15.6, concludes the
paper. The stability analysis is provided in the appendix of the paper.

15.2 Preliminaries

In the following, Rm×n denotes the m × n Euclidean space, |x| denotes the absolute
value of a scalar x, and x denotes the Euclidean norm of a vector x ∈ Rn .
The infinity norm of a vector x is denoted x∞ . Sgn(x) denotes the extension
of the signum function to a vector x = [x1 , x2 , . . . , xn ] , is such that Sgn(x) =
[sign(x1 ), sign(x2 ), . . . , sign(xn )] ∈ Rn .
In this section, useful definition and lemma are presented, which are used to the
design of the control law and also for the purpose of the stability analysis of the
three dimensional motion of the underactuated quadrotor system.
Lemma 1 (Levant 2005) Consider the first order sliding mode differentiator
designed as follows
324 J. Ghommam et al.

1
˙ 0 = ρ0 = −0 |0 − f (t)| 2 sign(0 − f (t)) + 1
˙ 1 = −1 sign(1 − ρ0 ) (15.1)

where 0 , 1 and ρ0 are the states of the first-order differentiator, 0 and 1 are the
design parameters of the differentiator (15.1) and f (t) is a known scalar function.
d
The output of the differentiator ρ0 can approximate the term dt f (t) to an arbitrary
˙
accuracy if |0 − f (t0 )| and |ρ0 − f (t0 )| are bounded.
Definition 1 (Filippov Solution Filippov 1964:) The vector function ϑ is said to
be a solution of ϑ̇ = Υ (ϑ) on the interval [t0 , t1 ] with discontinuity on the right hand
side such that Υ is Lebesgue measurable and essentially locally bounded function.
If ϑ is absolutely continuous on [t0 , t1 ]and for almost all t ∈ [t0 , t1 ], ϑ̇ ∈ Ψ [Υ ](ϑ)
4 4 4
where Ψ [Υ ](ϑ)  ε>0 μN =0 coΥ Ψ (ϑ, ε0 ) − N, t in which μN =0 denotes
the intersection overall sets N of Lebesgue measure zero, co denotes the convex
closure and Ψ (ϑ, ε) is an open ball of radius ε0 centred at ϑ.
Lemma 2 (Xian et al. 2004) Let the auxiliary function P (t) ∈ R be defined as
follows:
, n
t  
P (t) = − r  Md −K1 Sng(x(μ)) dμ+ K1 |xi (0)|−M
d (t0 )r(t0 ) (15.2)
0 i=1

where the subscript i = 1, . . . , n denotes the ith element of the vector x. If the gain
K1 is chosen such that the following hold:

K1 > Md ∞ + Ṁd ∞ (15.3)

.n
then P (t) ≥ 0 and r  Md − i=1 K1 |xi (0)| − r
 Ṁ
d ≤ 0.

15.3 Problem Statement

15.3.1 Motion Equations of a Quadrotor

The quadrotor helicopter studied in this paper is shown in Fig. 15.1. Due to
reactional torques, the quadrotor’s rotors are divided by pairs. One pair (1,3) turns
in the same clockwise direction, while the pair (2,4) turns in the counter-clockwise
direction. The flight mechanism of the quadrotor can be explained as follows:
The vertical motion of the vehicle is produced by generating a lift force through
collectively varying the rotor speeds with the same velocity, this would enable the
take-off/landing of the quadrotor. The yaw angle is controlled by the difference of
speed of each motors pair. While rotation around the y-axis (pitch) will cause the
quadrotor to move along the x-axis direction. Consider a body fixed reference frame
15 Backstepping-Based Nonlinear RISE Feedback Control for an. . . 325

Fig. 15.1 Quadrotor model xb

B
T1
yb
xd

zb
yd

zd
x

I
y
z

{B} attached to the vehicle’s center of mass which moves relative to a fixed inertial
frame {I }. The kinematic and dynamic equations of motion for the quadrotor in the
body-fixed reference frame can be described as follows (Ding et al. 2017):

ṗ = Rυ
mυ̇ = −mS(ω)υ + N1 (υ) + G + B1 u1 + d1 (15.4)
Θ̇ = T(Θ)ω
If ω̇ = −S(ω)If ω − N2 (ω) + B2 u2 + d2

where p = [x, y, z] represents the relative position of the center of mass of the
quadrotor with respect to an inertial frame and Θ = [θ, φ, ψ] represents the three
Euler angles (roll, pitch, yaw) that define the orientation of the quadrotor in space.
υ = [υx , υy , υz ] ∈ R3 and ω = [ωx , ωy , ωz ] ∈ R3 denote the linear and angular
velocities expressed in the body-fixed frame {B} respectively. N1 (.) and N2 (.) ∈
R3 denote the nonlinear aerodynamic damping interactions. G ∈ R3 is the gravity
vector defined as G = mgR ez , where m is the mass of the quadrotor, g ∈ R
denotes the gravitational acceleration and ez = [0, 0, 1] is the unit vector in the
inertial frame. The rotation matrix R(Θ) is given as below:
326 J. Ghommam et al.

R(Θ) =
⎛ ⎞
cos ψ cos θ − sin ψ cos φ + sin φ sin θ cos ψ sin ψ sin φ + sin θ cos ψ cos φ
⎝ sin ψ cos θ cos ψ cos φ + sin φ sin θ sin ψ − cos ψ sin φ + sin θ sin ψ cos φ⎠
− sin θ sin φ cos θ cos φ cos θ
(15.5)

The input u1 ∈ R provides lifting force in the z-direction, u2 ∈ R3 represents the

generalized moment torques on the roll, pitch and yaw directions. B1 = [0, 0, 1]
and B2 = I3 where I3 denotes a 3 × 3 identity matrix. S(.) ∈ SO3 denotes a
general form of skew symmetric matrix. d1 , d2 ∈ R3 represent the external forces
disturbance expressed in the inertial frame {I }. If ∈ R3×3 denotes a positive definite
inertia matrix relative to the inertial frame {I }. The matrix T(Θ) is defined as
follows
⎛ ⎞
1 sin φ tan θ cos φ tan θ
⎜0 − sin φ ⎟
T(Θ) = ⎜ ⎝
cos φ ⎟ (15.6)
sin φ cos φ ⎠
0
cos θ cos θ

The transformation matrix T(Θ) is nonsingular if the pitch angle satisfies θ =

(2k − 1) π2 .
To facilitate the control design, the following assumptions are made:
Assumption 1 The reference trajectory pd (t) = [xd (t), yd (t), zd (t)] is suffi-
ciently smooth, twice differentiable and satisfies ṗ d ∞ ≤ !1d , p̈ d ∞ ≤ !2d
where !1d and !2d are positive constants.
Assumption 2 To ensure that the angular transformation T(Θ) is nonsingular, we
assume that θ = (2k − 1) π2 and φ = (2k − 1) π2 , where k ∈ N and that T(Θ)∞
is bounded (i.e., there exists a positive constant ε1 such that T(Θ)∞ ≤ ε1 ).
Assumption 3 We assume that the quadrotor translational velocity signal υ(t) is
not measurable and thus cannot be available for state feedback, whereas the angular
velocity is measurable via Inertial system.
Assumption 4 The mass m and the UAV inertia matrix If are assumed to be
known. The disturbances d1 and d2 are time varying and their first two time
derivatives (i.e., di , ḋi , d̈i , i=1,2 ) exist and are bounded by known constants. (i.e,
there exists i1 , i2 and i3 such that di  ≤ i1 , ḋi  ≤ i2 and d̈i  ≤ i3 ).
Remark 1 The assumption on the unavailability of the quadrotor’s linear velocity
measurement stems from the fact that most commercially available quadrotor are
often equipped with inertial measurement unit (IMU) which is a set of sensors
consisting of accelerometer, magnetometer and gyroscope from which the attitude
can be measured with acceptable accuracy. However, there is no specific sensor that
is capable of measuring the body-fixed velocity υ. This justifies Assumption 3.
15 Backstepping-Based Nonlinear RISE Feedback Control for an. . . 327

15.3.2 Control Objective

The aim of the control design is to force the quadrotor AUV to track a desired
a three dimensional trajectory p d = [xd , yd , zd ] with a desired reference yaw
angle ψd (t). Specifically for any initial conditions p(t0 ), Θ(t0 ) and ω(t0 ) and under
Assumptions 1, 2, 3 and 4, design the thrust force u1 and the torque input u2 such
that

lim ep (t) = 0, lim ψ(t) − ψd (t) = 0 (15.7)

t→∞ t→∞

while keeping all other states of the quadrotor bounded for all initial conditions
p(t0 ), Θ(t0 ) and ω(t0 ), where ep (t) is the position tracking error expressed in the
body-fixed frame to be defined later.

15.4 Control Development

To address the positioning tracking problem, we start by defining the error variable,
henceforth called position tracking error ep , which is expressed in the body-fixed
frame and it is given as follows

ep = R (p − p d ) (15.8)

Note that the tracking error coordinates (15.8) is expressed with respect to a body-
fixed frame {B} rather than the inertial frame {I}, this is advantageous as it avoids
the dependence on the choice of the inertial reference frame. The position tracking
error rate ėp (t) is obtained by tacking the time derivative of (15.8) as follows:

ėp (t) = Ṙ (p − p d ) + R ṗ − R ṗd

= −S(ω)R (p − p d ) + υ − R ṗ d
1 1
= −S(ω)ep + (mυ − R ṗ d ) + ( − 1)R ṗd (15.9)
m m
The translational linear velocity tracking error can now be defined from Eq. (15.9)
as eυ ∈ R3 to be defined in the body-fixed fame as

eυ = mυ − R ṗd − δ (15.10)

where δ ∈ R3 is a scalar vector defines as δ = [0, 0, δ] with δ ∈ R is a design

constant. The time derivative of the velocity tracking error along the solutions of the
second equation of the quadrotor dynamics in (15.4) yields
328 J. Ghommam et al.

ėυ = −mS(ω)υ + N1 (υ) + G + B1 u1 + d1 + S(ω)R ṗd − R p̈d

= −S(ω)eυ − S(ω)δ + N1 (υ) + G − R p̈ d + B1 u1 + d1 (15.11)

The yaw angle tracking error as defined in (15.7) is given by eψ = ψ − ψd ∈ R, its

rate error is obtained by taking its time derivative as follows

ėψ = ψ̇ − ψ̇d = Tz (Θ)ω − ωzd (15.12)

where Tz (Θ) ∈ R3×1 is the third row of T(Θ) defined in (15.6) and ωzd = ψ̇d is
the desired yaw rate angle expressed in the body-fixed frame. Based on the tracking
errors (15.9), (15.11), (15.12) and the last equation of the quadrotor’s dynamics
(15.4), the control scheme construction is proposed based on the backstepping
approach and the nonlinear robust control technique of Cai et al. (2006). The
technique unfolds in two steps.
Step 1 Due to the unmeasured state υ(t) in (15.10), the control design cannot
contains terms involving the tracking error eυ . To overcome such obstruction, the
velocity tracking error is estimated through a second order filter designed as follows:
, t
rf = q − (k2 + 1) eυ (λ)dλ (15.13)
0
q̇ = −rf − (k2 + 1)(ep + rf + δ) + ep − ef (15.14)
ėf = −ef + rf (15.15)

where rf , ef ∈ R3 are filtered signals, q(t) ∈ R3 is an auxiliary variable useful for

the implementation of the filter and k2 is a positive constant to be determined later.
The initial conditions for the filter (15.13), (15.14) and (15.15) are as follows:

ef (t0 ) = 0 (15.16)
q(t0 ) = (k2 + 1)t0 eυ (t0 ) (15.17)
-t
Remark 2 Note from (15.13), that the second right hand term 0 eυ (λ)dλ is
completely known and can be evaluated directly by the measured signals p, Θ and
pd .
In order to further proceed with the development of the control input for the
translational dynamics, new error states are defined. Let the signal η ∈ R3 be defined
as follows:

η = eυ + ep + rf + δ (15.18)

Take the time derivative of (15.13) using, (15.14) and (15.18), we obtain
15 Backstepping-Based Nonlinear RISE Feedback Control for an. . . 329

ṙf = q̇ − (k2 + 1)eυ

= −rf − (k2 + 1)(ep + rf + δ) + ep − ef − (k2 + 1)(η − ep − rf − δ)
= −rf − (k2 + 1)η + ep − ef (15.19)

Take the time derivative of η and using (15.19), we obtain the following:

η̇ = ėυ + ėp + ṙf

= ėυ + ėp − (k2 + 1)η − rf + ep − ef
1 1 1
= −S(ω)η + S(ω)rf − S(ω)δ + eυ + δ + ( − 1)R ṗ d − R p̈d
m m m
+N1 (υ) + G + d1 + S(ω)δ + B1 u1 − (k2 + 1)η − rf + ep − ef (15.20)

Rearranging the terms in (15.20) yields

* +
 + Md + d1 + − S(δ)ω + B1 u1 − ep
η̇ = −(k2 + 1)η − S(ω)η + M (15.21)

 ∈ R3 and Md ∈ R3 are defined as

where the auxiliary functions M

 = S(ω)rf + S(ω)δ + 1 eυ + N1 (υ) − Md − (rf + ef − 2ep ) (15.22)

M
m
1 1
Md = ( − 1)R ṗd − R p̈ d + G + δd1 (15.23)
m m
The filtered position tracking error (15.21) can now be combined with the yaw
tracking error (15.12) to create a composite tracking error χ (t) ∈ R4 such that
χ = [η , eψ ] . The dynamics of the composite tracking error is therefore obtained
as follows:
( ) ( ) ( )( )
η̇ −(k2 + 1)η − S(ω)η + M  + Md − ep −S(δ) B1 ω
χ̇ := = +
ėψ −ωzd Tz (Θ) 0 u1
5 67 8 5 67 8
BC ζ
(15.24)
Note from (15.24), two components of the vector ζ := [ω , u1 ] can have influence
on the translational dynamics. The first component is u1 which is viewed as a
physical input to the system and can be directly specified in the control design,
while the second component is ω which can only be indirectly specified through the
design of the torque input u2 . The backstepping technique is therefore pursued to
control the translation dynamics via the angular velocity ω. To this end, we regard ω
as a desired angular velocity to force the filtered state χ to converge asymptotically
to zero. To this purpose, we define the backstepping variable error z as follows:

z = ω − Bz u1 (15.25)
330 J. Ghommam et al.

where u1 = [α  
ω , u1 ] ∈ R and the matrix Bz = [I3 , 03×1 ]. From (15.24) and
4

(15.25), it can be shown that:

( )
z
BC ζ = BC u1 + BC (15.26)
0

Then according to (15.26), the system dynamics (15.24) can be written as follows:
( ) ( ) ( )
η̇−(k2 + 1)η − S(ω)η + M  + Md − ep z
χ̇ := = + BC u1 + BC
ėψ −ωzd 0
(15.27)
From (15.27), the control input u1 can be designed using the robust mechanism
presented in Xian et al. (2004) as follows:
⎡ ⎤
0 − 1δ 0 0
⎢ ( )
⎢
1
0 0 0 ⎥ ⎥ −K1 Sgn(ep + ef ) + (k2 + 1)rf
u1 = ⎢ 1 δsin φ cos θ ⎥ (15.28)
⎣ − δ cos φ 0 0 cos φ ⎦ ωzd
0 0 1 0
5 67 8
BC−1

Notice that according to Assumption 2, the matrix BC is invertible and well defined,
this is due to the assumption that the pitch and the roll angles never reach singularity
± π2 . Clearly from (15.28), it is seen that the last term of this equation quantifies
an error term that can be controlled through the input to the rotational dynamics.
Substituting (15.28) into (15.27) to obtain the following closed-loop translational
system
( ) ( )
η̇  + Md − K1 Sgn(ep + ef ) − ep
(k2 + 1)(rf − η) − S(ω)η + M
=
ėψ 0
( )
z
+BC (15.29)
0

Remark 3 Note the importance of the scalar δ introduced in the filtering tracking
error in (15.14) as well as in (15.18), was to ensure that the matrix BC is nonsingular,
which based on Assumption 2 the matrix BC is invertible. The control input u1 is
therefore well defined.
Step 2 The control input u1 was designed to control the motion of the quadrotor
along the three dimensional axes. In order to control the quadrotor’s attitude, this
signal must be injected as a reference to be tracked by the rotational dynamics.
The design proceeds by modifying the standard backstepping procedure so that the
robustification mechanism of Cai et al. (2006) can be used in a simpler manner. To
do that, consider the following augmented system:
15 Backstepping-Based Nonlinear RISE Feedback Control for an. . . 331

1
ra = ż + z (15.30)
k3

where k3 is a positive design gain. Note that the augmented system (15.30) is
discontinuous and involves the time derivative of the control signal u1 . Since u1
contains signum function, its derivative is hard enough to figure out. To proceed with
the design an exact differentiator is needed to estimate the first two derivatives of
the control input u1 = [u1i ]
i=1,...,4 . The following finite time estimator is proposed:
9
/̇ ˙ 0i = ρ0i = 1i − 0i |0i − u1i |sign(0i − u1i )
u1  (15.31)
˙ 1i = −1i sign(1i − ρ0i ), i = 1, . . . , 4
 
˙ ˙
u1  ˙ 2i = ρ1i = 3i − 2i |2i − û1i |sign(2i − û1i )
/̈ (15.32)
˙ 3i = −3i sign(3i − ρ1i ), i = 1, . . . , 4

where 0i , 1i , 2i , 3i , ρ0i and ρ1i are the states of (15.31) and (15.32) respec-
tively. 0i , 1i , 2i and 3i are positive design constants. According to (15.32) and
Lemma 1, we have

/̈
u1 = ρ 1 + (15.33)

where ρ 1 = [ρ11 , ρ12 , . . . , ρ14 ] , is the estimation error of the first order sliding
mode differentiator (15.32). From Lemma 1, we have for each component of , the
following bound |#i | ≤ ςi , with ςi > 0, i = 1, 2, . . . , 4.
Remark 4 Note that if the filtered error ra of the state z converges asymptotically to
zero then the state tracking error z also does. The control design will be responsible
of ensuring boundedness of the filtered error as well as its asymptotic convergence
to zero.
Taking the time derivative of ra (t), multiplying by the inertia matrix If and using
the expression in (15.25), yields

If
If ṙa = z̈ + If ż
k3
If
= z̈ + k3 If (ra − z)
k3
1 
= Φ̇ + B2 u̇2 + D − If Bz/̈
u1 + k3 If (ra − z) (15.34)
k3

where Φ denotes the nonlinear part of the attitude dynamics in the last equation of
(15.4), such that Φ(t) = −S(ω)If ω − N2 (ω) and D is a disturbance compound
given as D = ḋ2 + Bz . Note that the differentiation of the backstepping variable
z involves the differentiation of the virtual control u1 which is a hectic operation as
332 J. Ghommam et al.

it contains signum function. It therefore has been replaced by its estimate using the
finite time estimator (15.32).
Based on (15.34), we design u̇2 as follows:

u̇2 = B−1
2 − Φ̇ + If Bz/̈
u1 − K2 Sng(z) − k3 (k3 If + I3 )ra (15.35)

where k3 > 0 and K2 is a positive gain control to be determined later. The first term
(15.35) is the nonlinear term to be canceled out form the attitude dynamics, the
second term is due to the time derivative of the virtual control law in the previous
step and third and the forth terms are used to compensate for the unknowns and
crossing terms during the Laypunov stability analysis. From (15.35), the actual
control law can be calculated according to the following:
  
u2 = B−1
2 − Φ(t) + Φ(t0 ) + If Bz /̇u1 (t) − /̇
u1 (t0 ) − (k3 If + I3 )(z(t) − z(t0 ))
, t
− k3 (k3 If + I3 )z(μ) − K2 Sng(z(μ))dμ (15.36)
t0

The control design is now completed, the main result of this paper can be
summarized in the following theorem:
Theorem 1 Consider the translational and rotational dynamics of the quadrotor in
closed loop system. the control input (15.28) and (15.36) ensure that all closed-loop
signals of the quadrotor system are bounded. In particular ep , ef , rf , eψ , η, z and
ra converge asymptotically to zero if the gain controllers K1 and K2 are chosen to
satisfy the following:

K1 > Md ∞ + Ṁd ∞ , K2 > D∞ + Ḋ∞ (15.37)

and the positive constant k2 is chosen to satisfy

" 2
ς max{1, m2 }
k2 > ϕ y(t0 ) (15.38)
4(ς + γ ) min{ 12 , m2 }

where an explicit expression for y(t0 ) is given as


y(t0 ) = ep (0)2 + eψ (0)2 + η2 + z(t0 ) + ra (t0 ) + P (t0 ) + Q(t0 )
(15.39)
In particular the velocity tracking error mυ − R ṗ d converges asymptotically to
zero.
Remark 5 The challenge of the proposed control for the aerial underactuated
quadrotor, is first to systematically incorporate the RISE feedback technique into
a backstepping design procedure which is made quite hard for most of underac-
15 Backstepping-Based Nonlinear RISE Feedback Control for an. . . 333

tuated systems. The second challenge is to achieve asymptotic convergence of the

closed-loop system in the presence of a disturbance. On the other hand, a major
disadvantage of the scheme is that the controller being designed has discontinuous
feedback due to the presence of signum functions. However, we emphasize that
although it is discontinuous control design, the proposed controller doesn’t suffer
from limitations revealed by other relevant control schemes such as the sliding mode
techniques characterized by their infinite bandwidth and chatter. An alternative way
to convert the proposed controller into a continuous feedback is to implement the
sigmoid function in place of the signum function. Nevertheless, incorporating the
sigmoid function in the implementation of the proposed controller would imply
different stability results (the reader is referred to Xian and Zhang 2017 for more
details).

15.5 Numerical Simulations

In this section, we illustrate the effectiveness of the proposed tracking controller

through a numerical simulation and assess the robustness property by adding
disturbances into the position and attitude loop. Simulations are carried out by
means of the “Matlab/Simulink” software program. For simplicity of the simulation,
the aerodynamics interaction terms N1 (.) and N2 (.) are neglected, hence the
physical parameters of the quadrotor UAV are taken as: m = 1.1 kg, d = 0.2125 m,
Irx = Iry = 0.013 kg/m2 , Irz = 0.024 kg/m2 , b = 1.41871 × 10−5 , c =
1.84 × 10−6 , and g = 9.81 m/s2 . The reference trajectories are taken pd =
[10 sin(0.1t), 10 cos(0.1t), 0.2t] m and the desired yaw angle ψd (t) = 0.1t rad.
The initial position and angle of the quadrotor at the center of mass are selected
as follows: p(t0 ) = [2, 1, 5] , Θ(t0 ) = [ π9 , 0, 0] and the constant vector design
δ is chosen as [0, 0, −1] . The control parameters for the filter and the controller
are chosen to be K1 = 30, k2 = 15, k3 = 5, K2 = 20, the gain parameters for the
differentiator are chosen as ε0i = ε2i = 10 and ε1i = ε3i = 10, for i = 1, . . . , 4. We
also consider that a normally distributed noise signal with zero mean and variance
of 0.1 is introduced into the motion equations of the quadrotor.
Figure 15.2 depicts the 3D trajectory tracking of the quadrotor with respect to the
desired trajectory, which clearly shows the accomplishment of the control objective
stated in (15.7). Figures 15.3 and 15.4 illustrates the asymptotic stability of the
trajectory tracking p − pd and the asymptotic convergence of the backstepping
error tracking variable z = [z1 , z2 , z3 ] respectively. These three figures illustrate
the validity of the robust position and attitude loops tracking results. Figure 15.5
illustrates the time variation of the resulting pitch and roll angles. Further, Fig. 15.6
illustrates the generated forces by the four motors of the quadrotor.
334 J. Ghommam et al.

Position in the earth referenced frame

40
Desired Trajectory
Qaudrotor Trajectory
35

30

25

20
z

15

10

-5
20 20
10 10
0 0
-10 -10
-20 -20
y x

Fig. 15.2 Figure of 3D trajectory tracking

15.6 Conclusion

This paper presented a nonlinear backstepping control design to achieve trajectory

tracking for quadrotor despite the presence of unknown external time varying
perturbation. The RISE technique has been employed within the backstepping
procedure in order to overcome the uncertainties in the system. During the position
tracking loop control design and due to the unavailability of the linear velocity
measurement a filter design is introduced to estimate the linear velocities in the
first step of the design. The integrator backstepping in combination with the RISE
feedback are then considered to overcome difficulties due to underactuation. It has
been demonstrated that, despite the presence of external unknown disturbances the
resulting closed-loop tracking error system is semi-globally asymptotically stable.
Future work will improve the proposed technique by considering parametric model
uncertainties affecting the system.
15 Backstepping-Based Nonlinear RISE Feedback Control for an. . . 335

1
x-x d (m)

-1
0 20 40 60 80 100 120 140 160 180 200
Time (sec)
2
y-y d (m)

-2
0 20 40 60 80 100 120 140 160 180 200
Time (sec)
10
z-z d (m)

-10
0 20 40 60 80 100 120 140 160 180 200
Time (sec)
5
(rad)

0
d
-

-5
0 20 40 60 80 100 120 140 160 180 200
Time (sec)

Fig. 15.3 Position and orientation tracking error variables

Appendix

For the clarity of the stability proof, we first write the closed loop-system by
substituting the control input (15.35) into the dynamics (15.34), yields
( ) ( )
η̇  + Md − K1 Sgn(ep + ef ) − ep
(k2 + 1)(rf − η) − S(ω)η + M
=
ėψ 0
+BC z (15.40)
1
If ṙa = D − K2 Sng(z) − k3 (k3 If + I3 )ra + If (ra − z) (15.41)
k3

where we have used the following relationship:

( ) ( ) ( )
z −S(δ)z −S(δ)
BC = = z
0 Tz (Θ)z Tz (Θ)
5 67 8
BC

For the closed-loop system (15.41) and (15.40), we consider the Lyapunov function
candidate function V which is Lipschitz positive definite function defined as:
336 J. Ghommam et al.

1.5
]

1
-1
z 1 [rads

0.5

-0.5
0 20 40 60 80 100 120 140 160 180 200
Time (sec)
0.5
]
-1

0
z 2 [rads

-0.5

-1
0 20 40 60 80 100 120 140 160 180 200
Time (sec)
6
]
-1

4
z 3 (m)[rads

-2
0 20 40 60 80 100 120 140 160 180 200
Time (sec)

Fig. 15.4 Backstepping variable time evolution z

V (y, t) = V1 + V2 + P + Q (15.42)

where we define the functions V1 and V2 as follows

m  1 1 1
V1 = e ep + e ef + r rf + χ  χ (15.43)
2 p 2 f 2 f 2
1 1
V2 = z If z + r If ra (15.44)
2 2 a
Let y ∈ R 20 be defined as
9 9
y = [X , P (t), Q(t)] (15.45)
X = [e     
p , ef , rf , η , eψ , z , ra ]

and the auxiliary function P ∈ R and Q ∈ R are defined as the Filippov solutions
(Filippov 1964) to the following differential equations respectively:

Ṗ = −η Md − K1 Sng(ep + ef ) (15.46)

r 
a
Q̇ = D − K2 Sgn(z) (15.47)
k3
15 Backstepping-Based Nonlinear RISE Feedback Control for an. . . 337

[rad] 2

1
Roll angle

-1
0 20 40 60 80 100 120 140 160 180 200
Time (sec)
[rad]

0.5

0
Pitch angle

-0.5
0 20 40 60 80 100 120 140 160 180 200
Time (sec)
[rad]

30

20
Yaw angle

10

0
0 20 40 60 80 100 120 140 160 180 200
Time (sec)

Fig. 15.5 Attitude angles

with initial conditions P (t0 ) and Q(t0 ) are given as follows

3
P (t0 ) = K1 |epi (0)| − e
p (0)Md (0) (15.48)
i=1

1
3
Q(t0 ) = K2 |zi (0)| − z(0) D(0) (15.49)
k3
i=1

with the subscript i = 1, . . . , n denotes the i-th element of the vectors ep (t) and
z(t). The gain control K1 and K2 are chosen according to the sufficient condition
(15.3) underlined in Lemma 2.
It can be verified that the Lyapunov function candidate defined in (15.42) can be
bounded as :

U1 (y) ≤ V (y, t) ≤ U2 (y) (15.50)

where U1 : R20 → R and U2 : R20 → R are positive definite functions defined

as U1 (y) = min{ 12 , m2 }y2 and U2 (y) = max{1, m2 }y2 . Under Filippov’s
framework, strong stability of the closed-loop system will be established in the
following.
338 J. Ghommam et al.

F1 F2
4 5

4.5
3.5
Force of M1

Force of M2
4
3
3.5

2.5
3

2 2.5
0 50 100 150 200 0 50 100 150 200
Time (sec) Time (sec)

F3 F4
4 4

3.5 3.5
Force of M3

Force of M4

3 3

2.5 2.5

2 2
0 50 100 150 200 0 50 100 150 200
Time (sec) Time (sec)

Fig. 15.6 Torques for Motors M1 , . . . , M4

The time derivative of (15.42) along the Filippov trajectories exists almost
everywhere (a.e), for almost all t ∈ [t0 , t], one can get:
:
˙ =
   1 12 1 1
ξ  Ψ [ė   
a.e
V̇ ∈ V p , ėf , ṙf , η̇ , ėψ , ż , ṙa , P Ṗ , Q Q̇, 1]
2
2 2
ξ ∈∂V (y,t)
(15.51)
where ∂V is the generalized gradient of V (Clarke 1990). Since V is continuously
differentiable with respect to y, then one has

˙ ⊂ ∇V  Ψ [ė , ė , ṙ , η̇ , ė , 1 P 12 Ṗ , 1 Q 12 Q̇, 1]


V (15.52)
p f f ψ
2 2
1 1
−2
where ∇V = Ψ [e     
p , ef , rf , η , eψ , z If , ra If , 2P , 2Q− 2 ] . Substituting
(15.9), (15.15), (15.24), (15.40) and (15.41) into (15.52) and using the calculus for
Ψ [.] from Paden and Sastry (1987), yields:

˙ ⊂ e (η − e − r − δ) + (1 − m)e R ṗ − e e + e r − r r

V p p f p d f f f f f f

−(k2 + 1)r η + r  
f ep − rf ef − (k2 + 1)η η + η M
15 Backstepping-Based Nonlinear RISE Feedback Control for an. . . 339


+η Md − K1 Ψ [Sng(ep + ef )] − η ep + (k2 + 1)η rf

+η BC z − Md − K1 Ψ [Sng(ep + ef )] − k3 z If z − r
a ra

1  1 
+ ra D − K2 Ψ [Sng(z)] − r D − K2 Ψ [Sng(z)] (15.53)
k3 k3 a

where Ψ [Sng(x)] = SGN(x) such that ∀x = [x1 , x2 , x3 ] , the set valued map
SNG(xi ) = 1 if xi > 0, [−1, 1] if xi = 0 and −1 if xi < 0 for i = 1, 2, 3. Using
the fact that the set in (15.53) reduces to a scalar equality since the right hand side
is continuous a.e, the following upper bound can be obtained on V̇ :

˙ ≤ 2 2
 2
−ep 2 − ef 2− rf 2− (k2 + 1)η2− k3 If z2−ra 2+2η M
a.e
V
2  2
2 2
+2 − e 
p δ + (m − 1)R ṗd 2 (15.54)

Denote by γ = δ + (m + 1)ṗd , using the Mean Value Theorem (De Queiroz
 can be upper bounded as M
et al. 1997), the function M  ≤ ϕ(X)X, then
using the young’s inequality yields:
  γ
 − e 
p δ + (m − 1)R ṗd  ≤ γ ep  ≤ X2 (15.55)
X

 ϕ(X)2
Mη ≤ + k2 η2 (15.56)
4k2
substituting (15.55) and (15.56) into (15.54), we obtain

˙ ≤

a.e ϕ(X)2
V −ep 2 − ef 2 − rf 2 − η2 − k3 If z2 − ra 2 +
4k2
γ
+ X2
X
a.e ϕ(X)2 γ
≤ −X2 + + X2 (15.57)
4k2 X

It follows through the choice of ς as any given positive constant such that X2 ≥
ς 2 , the expression in (15.57) can be upper bounded as:
 
˙ ≤

a.e γ ϕ(X)2
V − 1− − X2 (15.58)
ς 4k2

From (15.58), it can be concluded that for some positive constant C0 , we have:

˙ ≤

a.e ϕ(X)2
V −C0 X2 , for k2 > , and ς < γ (15.59)
4(1 − γς )
340 J. Ghommam et al.

It can be concluded from inequalities (15.50) and (15.59) that V (y, t) ∈ L∞ ,

therefore ep , ef , rf , η, z, ra , P and Q ∈ L∞ . From (15.13), (15.14) and (15.15), a
straightforward conclusion can be drawn to show that ėf , q, q̇ ∈ L∞ . From the fact
that the desired trajectory is sufficiently smooth then using (15.8) we can conclude
that p and ṗ ∈ L∞ . From (15.28) and (15.36) we know that u1 and u2 ∈ L∞ .
Given these boundedness statements, it is clear from that the time derivative of
U (X) = γ X2 is such that U̇ (X) ∈ L∞ , which also implies that U (X) is
uniformly continuous. Define the region SD as follows:
  " 2 ;
1 m −1 k2 (ς − γ )
SD = y ∈ D | U2 (y) < min{ , }ϕ 2 (15.60)
2 2 ς

The region of attraction in SD can be made arbitrary large by appropriately tuning

the control gain k2 . As a matter of fact, the region of attraction is calculated based
on the set given in (15.60) as follows:
"  " 
min{ 12 , m2 } −1 k2 (ς − γ )
y(t0 ) ≤ ϕ 2 (15.61)
max{1, m2 } ς

which can be rearranged as in (15.38). Based on the definition of the vector state
y in (15.45) and the initial conditions defined in (15.16) and (15.17) and (15.48)
and (15.49), an explicit expression for y(t0 ) can be derived as in (15.39).
From (15.59) it can be concluded that C0 X2 → 0 as time goes to
infinity, ∀y ∈ SD , it is then obvious to show form the definition of X that
ep , ef , rf , eψ , η, z, ra → 0 as t → ∞, ∀y ∈ SD . This further implies
that ev → [0, 0, −δ] as t → ∞, consequently the velocity tracking error
mυ − R ṗd → 0 as t → ∞. This completes the proof.

References

Abdessameud, A., & Tayebi, A. (2010). Global trajectory tracking control of VTOL-UAVs without
linear velocity measurements. Automatica, 46(6), 1053–1059.
Benallegue, A., Mokhtari, A., & Fridman, L. (2008). High-order sliding-mode observer for a
quadrotor UAV. International Journal of Robust and Nonlinear Control, 18(4–5), 427–440.
Bertrand, S., Guénard, N., Hamel, T., Piet-Lahanier, H., & Eck, L. (2011). A hierarchical controller
for miniature VTOL UAVs: Design and stability analysis using singular perturbation theory.
Control Engineering Practice, 19(10), 1099–1108.
Cai, Z., de Queiroz, M. S., & Dawson, D. M. (2006). Robust adaptive asymptotic tracking of
nonlinear systems with additive disturbance. IEEE Transactions on Automatic Control, 51(3),
524–529.
Castillo, P., Dzul, A., & Lozano, R. (2004). Real-time stabilization and tracking of a four-rotor
mini rotorcraft. IEEE Transactions on Control Systems Technology, 12(4), 510–516.
15 Backstepping-Based Nonlinear RISE Feedback Control for an. . . 341

Ciulkin, M., Pawluszewicz, E., Kaparin, V., & Kotta, U. (2015). Input-output linearization by
dynamic output feedback on homogeneous time scales. In 20th International Conference on
Methods and Models in Automation and Robotics (MMAR), Miȩdzyzdroje, Poland (pp. 477–
482).
Clarke, F. H. (1990). Optimization and nonsmooth analysis. Philadelphia: SIAM.
Cunha, R., Cabecinhas, D., & Silvestre, C. (2009). Nonlinear trajectory tracking control of
a quadrotor vehicle. In Proceedings of ECC 2009 – 10th European Control Conference,
Budapest, Hungary.
De Queiroz, M., Hu, J., Dawson, D., Burg, T., & Donepudi, S. (1997). Adaptive position/force
control of robot manipulators without velocity measurements: Theory and experimentation.
IEEE Transactions on Systems, Man, and Cybernetics, 27-B(5), 796–809.
Ding, X., Wang, X., Yu, Y., & Zha, C. (2017). Dynamics modeling and trajectory tracking control
of a quadrotor unmanned aerial vehicle. Journal of Dynamic Systems, Measurement, and
Control, 139, 1–11.
Do, K. D. (2015). Path-tracking control of stochastic quadrotor aircraft in three-dimensional space.
Journal of Dynamic Systems, Measurement, and Control, 137(10), 1–11.
Filippov, A. (1964). Differential equations with discontinuous right-hand side. American
Mathematical Society Translations, 42(2), 199–231.
Garcia, R., Rubio, F., & Ortega, M. (2012). Robust PID control of the quadrotor helicopter. IFAC
Proceedings Volumes, 45(3), 229–234.
Ghommam, J., Charland, G., & Saad, M. (2015). Three-dimensional constrained tracking control
via exact differentiation estimator of a quadrotor helicopter. Asian Journal of Control, 17(3),
1093–1103.
Huang, M., Xian, B., Diao, C., Yang, K., & Feng, Y. (2010). Adaptive tracking control of
underactuated quadrotor unmanned aerial vehicles via backstepping. In Proceedings of the
IEEE American control Conference (ACC), Baltimore, MD, USA (pp. 2076–2081).
Ilcev, D. S., & Sibiya, S. S. (2015). Weather observation via stratospheric platform stations. In
2015 IST-Africa Conference, Lilongwe, Malawi (pp. 1–12).
Kendoul, F. (2009). Nonlinear hierarchical flight controller for unmanned rotorcraft: Design,
stability, and experiments. Journal of Guidance Control and Dynamics, 32(6), 1954–1958.
Kim, H. J., & Shim, D. H. (2003). A flight control system for aerial robots: Algorithms and
experiments. Control Engineering Practice, 11(2), 1389–1400.
Lee, D. B., Burg, T. C., Xian, B., & Dawson, D. M. (2007). Output feedback tracking control of
an underactuated quad-rotor UAV. In Proceedings of the 2007 American Control Conference,
New York, USA (pp. 1775–1780).
Levant, A. (2005). Quasi-continuous high-order sliding-mode controllers. IEEE Transactions on
Automatic Control, 50(11), 1812–1816.
Madani, T., & Benallegue, A. (2007). Sliding mode observer and backstepping control for a
quadrotor unmanned aerial vehicles. In 2007 American Control Conference, New York, USA
(pp. 5887–5892).
Magsino, E. R., Chua, J. R. B., Chua, L. S., de Guzman, C. M., & Gepaya, J. V. L. (2016). A rapid
screening algorithm using a quadrotor for crack detection on bridges. In 2016 IEEE Region 10
Conference (TENCON), Singapore (pp. 1829–1833).
Ortiz, J. P., Minchala, L. I., & Reinoso, M. J. (2016). Nonlinear robust h-infinity pid controller for
the multivariable system quadrotor. IEEE Latin America Transactions, 14(3), 1176–1183.
Paden, B., & Sastry, S. (1987). A calculus for computing filippov’s differential inclusion with
application to the variable structure control of robot manipulators. IEEE Transactions on
Circuits System, 34(1), 73–82.
Rinaldi, F., Chiesa, S., & Quagliotti, F. (2013). Linear quadratic control for quadrotors UAVs
dynamics and formation flight. Journal of Intelligent and Robotic Systems, 70(1–4), 203–220.
Salih, A. L., Moghavvemi, M., Mohamed, H. A. F., & Gaeid, K. S. (2010). Modelling and PID
controller design for a quadrotor unmanned air vehicle. In 2010 IEEE International Conference
on Automation, Quality and Testing, Robotics (AQTR), Cluj-Napoca, Romania (Vol. 1, pp. 1–5).
342 J. Ghommam et al.

Van Den Eijnden, S. (2017). Cascade based tracking control of quadrotors. Master’s thesis,
Eindhoven University of Technology.
Wang, L., & Jia, H. (2014). The trajectory tracking problem of quadrotor UAV: Global stability
analysis and control design based on the cascade theory. Asian Journal of Control, 16(2),
574–588.
Xian, B., Dawson, D. M., de Queiroz, M. S., & Chen, J. (2004). A continuous asymptotic tracking
control strategy for uncertain nonlinear systems. IEEE Transactions on Automatic Control,
49(7), 1206–1211.
Xian, B., & Zhang, Y. (2017). A new smooth robust control design for uncertain nonlinear systems
with non-vanishing disturbances. International Journal of Control, 89(6), 1285–1302.
Yao, Z. (2017). Nonlinear robust adaptive hierarchical sliding mode control approach for
quadrotors. International Journal of Robust and Nonlinear Control, 27(6), 925–941.
Zou, Y. (2017). Nonlinear hierarchical control for quad-rotors with rotation matrix. International
Journal of Control, 90(7), 1308–1318.
Chapter 16
On the Tip–Path-Plane Flapping Angles
Estimation for Small–Scale Flybarless
Helicopters at Near–Hover

Mohammad K. Al–Sharman and Mamoun F. Abdel–Hafez

Abstract Observing the Tip-path-plane (TPP) flapping motion is essential for

characterizing the helicopter’s main rotor dynamics. The high-order harmonics
of the rotor blade flapping periodic motion are considered the major sources of
helicopter vibration. However, incorporating precise information of the flapping
dynamics with the overall helicopter dynamic model describes better the main rotor
motion. This enables designers to innovate blade designs that have less dynamic
vibration. Moreover, the flapping states are crucial for analysing the main rotor axial
force and moment components. However, obtaining measurements for the flapping
states is not directly possible. This mandates researchers in the field to exclude the
flapping dynamics, despite being essential, and use some algebraic expressions and
numerical approximations instead. This chapter addresses the problem of designing
a model-based estimation algorithm for the unobservable flapping angles while
a near-hover flight is being carried out. A Kalman state estimation algorithm is
designed to provide accurate flapping estimates for the Maxi Joker 3 flapping angles.
The presented estimator has succeeded in obtaining accurate longitudinal and lateral
flapping angles estimates with small root mean square error of estimation of 0.3770◦
and 0.2464◦ , respectively. Besides several simulation tests, a real outdoor near-hover
flight was performed to validate the proposed estimation method.

Keywords Small-scale helicopter · Flapping angles estimation · State

estimation · Near–hover dynamics

M. K. Al–Sharman ()
Department of Electrical and Computer Engineering, University of Waterloo, ON, Canada
e-mail: [email protected]
M. F. Abdel–Hafez
Mechanical Engineering Department, American University of Sharjah, Sharjah, UAE
e-mail: [email protected]

© Springer Nature Singapore Pte Ltd. 2019 343

N. Derbel et al. (eds.), New Developments and Advances
in Robot Control, Studies in Systems, Decision and Control 175,
https://doi.org/10.1007/978-981-13-2212-9_16
344 M. K. Al–Sharman and M. F. Abdel–Hafez

16.1 Introduction

Unmanned miniature aerial vehicles should be in a high degree of agility and

maneuverability (Al-Sharman et al. 2015, 2018a; Marantos et al. 2017). Agile
helicopter platforms have been deemed candidates in a range of civil and military
applications (Al-Sharman et al. 2018b; Alshoubaki et al. 2015). Nonetheless,
helicopters are highly nonlinear unstable dynamic systems; their dynamic model
suffers from a high number of states, subsystems cross coupling, model uncertainties
and noise associated with measurements. In addition, the unmeasured flapping
dynamic states makes the problem of helicopter’s states estimation and control more
sophisticated.
The blade flapping motion of the full-scaled helicopter’s main rotor has been
extensively studied (Chen 1987, 1990; Johnson 1981). A correlation between the
measured and the calculated values of the lateral flapping angle for the helicopter’s
rotor model was presented in Johnson (1981). A discrepancy was discovered
between the calculated flapping angle and the one obtained from static inflow
models. This discrepancy is because of the nonuniform distribution of the induced
velocity by the rotor blades vortices. However, this discrepancy was reduced by
combining the high-order aerodynamic models to the existed rotor model. In Chen
(1987), a thorough formulation of the flap lag equations of motion for an articulated
rotor was proposed. Recently, researches have addressed the problem of observing
the flap-lead-lag motion for full-scaled helicopter models (Allred et al. 2017; Zappa
et al. 2016). A real-time contactless instrumentation system was designed to observe
the TPP pitch flap lag motion for an articulated motor (Zappa et al. 2016). The
system consists of a set of laser and other vision sensors installed on the rotor head
of an Agusta Westland AW139 helicopter. While in Allred et al. (2017), a set of
gyros and MEMS accelerometers was evenly placed to monitor the flap, lead-lag
and pitch motion of the main rotor.
Small-scale helicopters have noticeably different flapping dynamics from that of
the full-scaled helicopters. Practically, full-scaled helicopters have flapping hinges;
nonetheless, these hinges do not exist in the small-scale ones. In addition, some
small-scale helicopters have a mechanical stabilizing flybar attached to the rotor
head to slow down the rapid dynamic response of the helicopter. Compared to
the full-scaled helicopters, the dynamic response of small-scale helicopters is
remarkably faster and more sensitive to disturbances (Zhu and Zuo 2017). However,
a robust control and effective sensor fusion design is required to navigate them
precisely (Al-Sharman 2015; Alshoubaki et al. 2015). Some research efforts have
focused on minimizing the state space model of the small-scale helicopter by
replacing the flapping dynamics with static algebraic approximation to reduce the
complexity of the control design (Gavrilets et al. 2002; Zhu and Zuo 2017). In
Gavrilets et al. (2002), a robust control design for an autonomous aerobatic flight
for a small-scale helicopter was introduced. Instead of implementing a flapping
angles state estimator, an approximation of the flapping angles was proposed for
them to maintain the robustness of the aerobatic flight controller. In Zhu and Zuo
16 On the Tip–Path-Plane Flapping Angles Estimation for Small–Scale. . . 345

(2017), the flapping dynamic states were approximated by algebraic expressions,

and a Bell-Hiller stabilizing bar of the small-scale helicopter was used to slow down
the model dynamics and provide some damping for the rotor dynamics. However,
in this study, a helicopter with no mechanical stabilizer (flybarless) is used to
perform an autonomous near-hover flight. While designing near-hover controllers,
the helicopter dynamic states, including the flapping states, are linearized at near
hover point. These linearized states are precisely estimated and controlled by a set
of linear estimators and controllers (Al-Sharman 2016; Liu et al. 2013).
It is discerned that the existing research effort in the field of helicopter’s
navigation and control removed the need to estimate the flapping states of small-
scale helicopters by excluding their dynamics or using static algebraic expressions.
However, to improve the accuracy of the helicopter’s maneuverability, an accurate
knowledge of each vehicle’s state is needed. Therefore, an accurate estimation of the
flapping angles states is required for a precise hovering performance of the miniature
flybarless helicopter. The flapping angles estimation problem is considered difficult
due to the fact that no sensor can be directly attached to the main rotor of the small-
scale helicopter to provide the estimator with flapping angles measurements. In this
case, the accuracy of the flapping angles estimation at near hover is only based on
the accuracy of the open loop near-hover dynamics of the helicopter. Therefore,
this chapter addresses the problem of small-scale flapping angles estimation and
proposes an implementation of a model-based estimator for these angles at near
hover conditions.
The present chapter is structured as follows. Section 16.2 describes the dynamic
model of the small-scale helicopter. In Sect. 16.3, the flapping angles state estima-
tion algorithm is presented. Section 16.4 illustrates the flapping angles estimation
results of the simulation tests. In Sect. 16.5, the experimental validation of the
proposed estimation method is illustrated. Finally, the proposed work is concluded
in Sect. 16.6.

16.2 Small-Scale Flybarless Helicopter Modeling

The dynamic model of the small-scale flybarless helicopter is a highly nonlinear

unstable dynamic model. This is due to the high number of states along with
its subsystems cross-coupling, model uncertainties and the noise associated with
measurements. In addition, the immeasurable main rotor flapping dynamic states
makes the problem of helicopter’s navigation and control more sophisticated.
Therefore, incorporating accurate helicopter’s state estimates of the unobservable
flapping angles into the near-hover control design would boost the robustness of
the near-hover flight performance. This study uses the nonlinear dynamic model of
the flybarless Maxi Joker 3 helicopter for designing a model-based flapping angles
estimators (Al-Sharman et al. 2014; Alshoubaki et al. 2015) (See Fig. 16.1).
346 M. K. Al–Sharman and M. F. Abdel–Hafez

Fig. 16.1 Maxi Joker 3

helicopter

16.2.1 Reference Frames

Two frames are commonly used to describe the helicopter motion and orientation:
the inertial earth frame (EF) and the body frame (BF). Both frames are used to
describe the position and attitude of the helicopter. The inertial earth frame is
positioned at a stationary point on earth, where the ground station is located;
whereas, the origin of the body frame is in the center of gravity of the vehicle.
The body frame is vital to describe the equations of motion and characterize the
aerodynamic forces and torques.

16.2.2 Rotation Matrix

The rotation matrix illustrated in Eq. (16.1) is used to transfer vectors from the body
frame to the earth frame and vice versa. The orthonormal rotation matrix Rbe is
used to change the representation of the helicopter forces and moments from frame
to another. The matrix is formulated by performing successive roll, pitch and yaw
rotations about the x, y, and z-axes, respectively.
⎡ ⎤
cos θ sin ψ cos ψ sin θ sin φ − sin ψ cos φ cos ψ sin θ cos φ + sin ψ sin φ
Rbe = ⎣ cos θ sin ψ sin ψ sin θ sin φ + cos ψ cos φ sin ψ sin θ cos φ − cos ψ sin φ ⎦
− sin θ cos θ sin φ cos θ cos φ
(16.1)

where we denote by c(.) = cos(.) and s(.) = sin(.).

16.2.3 Rigid Body Equations

The helicopter rigid body equations describe the position and the attitude in the
body frame of the helicopter. Equations (16.2), (16.3) and (16.4) represent the
rotational dynamics of the helicopter while Eqs. (16.5), (16.6) and (16.7) represent
16 On the Tip–Path-Plane Flapping Angles Estimation for Small–Scale. . . 347

the helicopter’s translational dynamics. The body rates around the Xb , Yb and
Zb axes are p, q and r as shown in Fig. 16.2, respectively. Equations (16.2),
(16.3), (16.4), (16.5), (16.6) and (16.7) represent rigid body equations of the
helicopter (Alshoubaki et al. 2015):

L Jz − Jy
ṗ = − qr (16.2)
Jx Jx
M Jx − Jz
q̇ = − pr (16.3)
Jy Jy
N Jy − Jx
ṙ = − pq (16.4)
Jz Jz
FXb
u̇ = rv − qw + (16.5)
m
FY b
v̇ = −ru + pw + (16.6)
m
Fzb
ẇ = qu − pv + (16.7)
m
where: p, q and r denote the angular rates around the XB , YB and ZB axes,
respectively; u, v and w denote the velocities in the body frame along the XB , YB
and ZB axes, respectively; JX , JY and JZ are the moments of inertia about the
XB , YB and ZB axes, respectively; m is the mass of the helicopter; FXb , FY b and
FZb define the force components in the body frame along the XB , YB and ZB axes,
respectively. The body moments about the XB , YB and ZB axes are denoted by
L, M and N .

16.2.4 Euler Angles Dynamics

Euler angles characterize the orientation of the helicopter with respect to earth
frame. These angles are formulated based on the 6 DOF vehicle model. Equations
(16.11), (16.12) and (16.13) define the Euler angles as:
 
φ̇ = p + q sin φ + r cos φ tan θ (16.8)
θ̇ = q cos θ − r sin φ (16.9)
1  
ψ̇ = p sin φ + r cos φ (16.10)
cos φ
348 M. K. Al–Sharman and M. F. Abdel–Hafez

16.2.5 Main Rotor Thrust Model

The thrust generated by the helicopter rotor is needed to sustain the vehicle in the
air. This thrust is governed by the collective command as given in Eq. (16.11) below
(Leishman 2000; Padfield 1996):
*  1 μ2  μ + λ + as
z i
TMR = θ0 + ar − ρ(ΩRmr )2 Ad (16.11)
3 2 2 2
where: μar represents the advanced ratio, which is the ratio between the rotor tip
speed and the helicopter translational velocity; λi is the rotor inflow; μz defines the
ratio of vertical velocity to the rotor tip speed; a and s denote the rotor lift curve
slope and the rotor solidity, respectively. Rmr is the radius of the main rotor, and Ad
is the rotor disk area.

16.2.6 Main Rotor Flapping Dynamics Model

The problem of modeling the flapping angles has been investigated in numerous
published works (Perdomo and Wei 2017; Taamallah 2011). In Taamallah (2011),
the blade pitch-lag-flap angles of a small-scale flybarless helicopter rotor were
expanded using the Fourier series as shown in Eq. (16.12). In Perdomo and Wei
(2017), a numerical solution was proposed for computing the high order coefficients
of equation (16.12) in order to estimate the high order harmonics of the blade
flapping periodic motion. Estimating these harmonics offers better understanding
of innovating blade designs that have lesser rotor dynamic vibration.

βbl (ϕbl ) = β0 + β1c cos ϕbl + β1s sin ϕbl (16.12)

where: ϕbl is defined by the blade’s azimuthal angular position, β0 describes the
TPP coning angle of the rotor, and β1c describes the TPP pitch angle with respect
to the X-axis of the body frame. Whereas, β1s describes the TPP roll angle with
respect to the Y -axis.
The flapping dynamic model of the small-scale helicopter is presented in
Gavrilets et al. (2001). As shown in Eq. (16.13), the flapping angles commands,
c and β c , are proportionate to the longitudinal and latitudinal input commands,
β1c 1s
respectively.
Through the system identification process (Gavrilets et al. 2001), the main rotor
time constants τf c and τf s are determined. At near-hover, τf c , τf s , Aδlon and Bδlat
have the values of 0.113 s, 0.101 s, 4.2 and 4.2, respectively.
⎛ ⎞
0 1( ⎡ ⎤ p
( ) ) Aδlon
β̇1c − τf c
1
0 β̇1c 0 −1 0 ⎜ ⎟
= +⎣ τf c ⎦⎜ q ⎟ (16.13)
β̇1s 0 − τf s
1
β̇1s −1 0 0
Bδlat ⎝ βc ⎠
τf s 1c
c
β1s
16 On the Tip–Path-Plane Flapping Angles Estimation for Small–Scale. . . 349

Fig. 16.2 Rotor lateral

T
flapping moment acting on
the fuselage (Gavrilets 2003)
Mk
kb

hmr

c.g.

16.2.7 Rotor Forces and Moments

The vector of the force acting on the main rotor of the helicopter consists of three
components along the axes X, Y and Z. Assuming that the generated main rotor
thrust is normal to the rotor disc, the main rotor forces are expressed as a function
of the flapping angles and the main rotor thrust. These forces are described in
Eqs. (16.14), (16.15) and (16.16) as follows:

fx,MR = −TMR sin β1c cos β1c (16.14)

fy,MR = TMR sin β1s cos β1s (16.15)
fz,MR = −TMR cos β1s cos β1c (16.16)

The rotor flapping motion produces the dominant moments of the main rotor.
Figure 16.2 displays the generated moment from the lateral TPP flapping angle.
A linear torsional spring with Kβ stiffness coefficient is used to express the restraint
in the blade attachment (Gavrilets 2003).
The rotor has three axial components of moments that are described in the
bodyframe. These components are given in Eqs. (16.17), (16.18) and (16.19) as
follows:

LMR = fy,MR hmr − fz,MR ymr + Kβ β1s − Lb τf s p (16.17)

MMR = fx,MR hmr − fz,MR ymr + Kβ β1c − Ma τf c q (16.18)
bcR 2 MR
NMR = −ρ(ΩR)2 CQ (16.19)
s
where: hmr is the distance in the z-axis from CG to the main rotor. lmr is the distance
in the x-axis from CG to the main rotor. ymr is the y-axis distance from CG to the
main rotor.
350 M. K. Al–Sharman and M. F. Abdel–Hafez

16.2.8 Helicopter Model and Linearization

Several studies address the linearization process of the helicopter’s dynamics at near
hover (Al-Sharman 2016; Hald et al. 2015). This is due to its importance and need
for designing linear near-hover controllers and estimators (Al-Sharman 2016). As a
linear estimator is being implemented in this study, the linearized helicopter open
loop dynamics are obtained at a near hovering. At near hovering point, the attitude
rates are equal to zero while the attitude angles are constant. The linearized system
is given in Eq. (16.20) as follows:

Ẋ = AX + BU + w (16.20)
Y = CX + DU + v

where X, U and Y represent the state space vector, the system’s input vector and
the output vector, respectively. The state-space matrices are defined in the A, B, C
and D matrices. w and v are the dynamics and the measurement Gaussian white
noise processes, respectively. As shown in Eq. (16.21), the helicopter’s state vector
is reduced to have the attitude, attitude rates and the main rotor flapping angles.
!T
X = φ θ ψ φ̇ θ̇ ψ̇ β1c β1s (16.21)

The input vector is given as:

⎡⎤
ulon
⎢ ulat ⎥
U =⎢ ⎥
⎣ ucol ⎦ (16.22)
uped

where: ulon is the longitudinal stick input ulat is the lateral stick input, the collective
lever input is denoted by ucol , and uped is the rudder pedal input. The dynamics
matrix, A, is computed to be:
⎡ ⎤
0 00100 0 0
⎢0 ⎥
⎢ 00010 0 0 ⎥
⎢0 ⎥
⎢ 00001 0 0 ⎥
⎢ Kβ + π +K86
m hmr ⎥
⎢0 −36.63 ⎥
⎢ 00000 Jx ⎥
A=⎢ Kβ + π +K86
m hmr ⎥ (16.23)
⎢0 ⎥
⎢ 00000 Jy 14.29 ⎥
⎢ ⎥
⎢0 00000 0 0 ⎥
⎢ ⎥
⎣0 00000 −Ta 0 ⎦
0 00000 0 −Tb
16 On the Tip–Path-Plane Flapping Angles Estimation for Small–Scale. . . 351

The input matrix B is shown in Eq. (16.24). The C matrix is a 6 × 8 unity diagonal
matrix while the D matrix is a 6 × 4 zero matrix. The Maxi Joker 3 parameters,
including the parameters of Eqs. (16.23) and (16.24), can be found in Al-Sharman
(2016).
⎛ ⎞
0 0 0 0
⎜ 0 ⎟
⎜ 0 0 0 ⎟
⎜ ⎟
⎜ 0 0 0 0 ⎟
⎜ Kt ltr ⎟
⎜ 0 0 0 ⎟
B=⎜
⎜ 0
Jx ⎟
(16.24)
⎜ 0 0 0 ⎟ ⎟
⎜ 0 Kt ltr ⎟
⎜ 0 0 Jz ⎟
⎜ ⎟
⎝ Ta 0 0 0 ⎠
0 Tb 0 0

16.3 Flapping Angles Estimation at Near–Hover

Precise state estimation technique is needed to ensure noise reduction and accurate
navigation (Abdel-Hafez and Speyer 2008; Abdel-Hafez 2010). This section pro-
poses the flapping angles state estimator design under near-hover conditions using
a Kalman filter. The Kalman state estimation cycle is described, and the simulation
results are presented. The results demonstrate accurate flapping states estimation
with small RMSEE values.

16.3.1 Kalman Filter Design

Kalman filter is deemed an optimal minimum mean square error state estimator.
Therefore, this study uses the Kalman state estimation algorithm to obtain opti-
mal estimates for the TPP flapping angles. This linear model-based estimation
algorithm uses the linearized near-hover dynamic model described in Sect. 16.2.
It relies on the accurate open loop dynamics of the system along with the
measurements of the onboard sensors to estimate the helicopter’s states (Al-
Sharman 2016). The Kalman state estimation cycle is described in the set of
Eq. (16.25), (16.26), (16.27), (16.28), (16.29), (16.30) and (16.31) below:
1. State and measurement prediction

x̂k+1|K = AK x̂k|K + BK Uk (16.25)

PK+1|k = AK PK|k A
K + Qk (16.26)
352 M. K. Al–Sharman and M. F. Abdel–Hafez

2. State update
vK+1 = zK+1 − HK+1 x̂k+1|K (16.27)

SK+1 = RK+1 + HK+1 PK+1|k HK+1 (16.28)
 −1
WK+1 = PK+1|k HK+1 SK+1 (16.29)
x̂k+1|K+1 = x̂k+1|K + WK+1 vK+1 (16.30)

PK+1|K+1 = PK+1|k − WK+1 SK+1 WK+1 (16.31)

By assuming the prior knowledge of the matrices AK , BK , HK , RK and QK

noise matrices, the Kalman filter begins its recursive state estimation cycle. Equa-
tion (16.28) represents the innovation sequence which provides the Kalman filter
with the additional information as new measurement becomes available. Equa-
tions (16.28), (16.29), (16.30) and (16.31) incorporate the measurement updates
into a priori estimate to enhance the posterior estimate and covariance.
The variances of the diagonal measurement noise covariance matrix and process
noise covariance matrix are precisely chosen to avoid divergence of the estimator.
As no measurements are available for the flapping angles, RK contains only the
attitude and the body rates variances.

16.4 Simulation Results

This section presents the simulation results of the proposed flapping angles state
estimation method. Since there are no measurements for the flapping angles, their
estimation is still challenging unless an accurate dynamic helicopter model is
identified. The IMU measurements of body angles and the body rates were obtained
by adding some measurement white noise sequences to the truth signals of the Maxi
Joker 3 helicopter dynamic states. A near-hover simulated test was conducted to
demonstrate the robustness of the proposed estimator. As shown in Figs. 16.3, 16.4
and 16.5, the Maxi Joker 3 is controlled to change its orientation by small degrees
to maintain the near-hover dynamics. It can be seen that the added noise has been
filtered out, and the attitude states are accurately estimated. The attitude estimation
problem for the flybarless is thoroughly investigated in Al-Sharman (2016).
Results illustrated in Figs. 16.6, 16.7 and 16.8 exhibit the TPP flapping angles
estimation of the Maxi Joker 3. Figure 16.6 shows the real and estimated flapping
angles. The significant coincidence between the actual and the estimated flapping
angles can be seen. Figure 16.7 shows the estimation result of the longitudinal
flapping angle. The estimated longitudinal flapping angle coincides well with the
truth angle. It is estimated with RMSEE of 0.3770◦ . The latitudinal angle is also
estimated accurately as shown in Fig. 16.8. The proposed estimator has succeeded
in estimating the latitudinal flapping angle with RMSEE of 0.2464◦ . The flapping
angles estimation errors are illustrated in Fig. 16.9. It is obvious that the estimation
algorithm for both angles has small value of errors.
16 On the Tip–Path-Plane Flapping Angles Estimation for Small–Scale. . . 353

15
Measured
Estimated
Actual
10
Roll angle (deg)

-5
0 1 2 3 4 5 6
Time (Sec)

Fig. 16.3 Roll angle response at near-hover

15
Measured
Estimated
10 Actual
Pitch angle (deg)

-5
0 1 2 3 4 5 6
Time (Sec)

Fig. 16.4 Pitch angle response at near-hover

16.5 Experimental Validation

The proposed estimation algorithm is experimentally validated by conducting an

actual near-hover flight. In real flight scenarios, the estimation error cannot be
accurately calculated because the truth values of the helicopter states are not
available. Nonetheless, in a near-hover experiment, the real values of the helicopter
states are expected to have approximately known values. A Mazari autopilot was
used to conduct the outdoor near-hover flight of the Maxi Joker 3. As shown in
Fig. 16.10, the Mazari autopilot uses the MPC555 32 bits microcontroller hardware
to sample the measurements of the MIDG sensor at 50 Hz. The specifications of
measurements of MIDG are illustrated in Table 16.1.
The attitude response of the Maxi Joker 3 helicopter during the actual near-hover
is illustrated in Fig. 16.11. The helicopter was controlled to perform short rolling
and pitching to maintain the near hover dynamics.
354 M. K. Al–Sharman and M. F. Abdel–Hafez

15
Measured
Estimated
Actual
Heading angle (deg)
10

-5
0 1 2 3 4 5 6
Time (Sec)

Fig. 16.5 Heading angle response at near-hover

Longitudinal angle (deg)

10
Actual
Estimated
0

-10
0 1 2 3 4 5 6
Time (Sec)
Latitudinal angle (deg)

10
Actual
Estimated
0

-10
0 1 2 3 4 5 6
Time (Sec)

Fig. 16.6 Flapping angles estimation

It is expected that the flapping angles are tilting with small degrees while near-
hover flight is being performed. Figures 16.12 and 16.13 display the performance of
the flapping angles estimation algorithm at near-hover. Although the flapping states
are unobservable, it can be seen that the estimated values coincide with the expected
values at near-hover.

16.6 Chapter Summary

Accurate information of the small-scale helicopter’s flapping states is needed for

characterizing the main rotor dynamics. However, these angles cannot be directly
measured by attaching sensors to the small rotor head to measure the main rotor
flapping motion. Therefore, an estimation algorithm has to be used to estimate
16 On the Tip–Path-Plane Flapping Angles Estimation for Small–Scale. . . 355

10
Actual
Estimated
Longitudinal angle (deg)

-5 0
-2

5.3 5.35
-10
5 5.2 5.4 5.6 5.8 6
Time (Sec)

Fig. 16.7 Flapping longitudinal angle estimation at near hover-flight

1.5
Actual
1 Estimated
Latitudinal angle (deg)

0.5

0.8
-0.5 0.6
0.4
-1 0.2
5.1 5.15
-1.5
5 5.2 5.4 5.6 5.8 6
Time (Sec)

Fig. 16.8 Flapping latitudinal angle estimation at near-hover flight

these states. The existing research efforts in the field of helicopter’s modeling and
control has focused on reducing the vehicle model to exclude the flapping dynamics
or replace them with algebraic expressions and numerical approximations. This
chapter presents a model-based estimation algorithm for the unobservable flapping
angles while a near-hover flight is being carried out. A Kalman state estimation
algorithm is designed to provide accurate flapping estimates for the Maxi Joker 3.
The proposed estimation technique was used to estimate the flapping angles while
simulated and real near-hover flights are carried out. The results prove the accurate
flapping estimation with small RMSEE values at near-hover conditions. Finally,
despite the essentiality of the flapping states in describing the rotor orientation,
356 M. K. Al–Sharman and M. F. Abdel–Hafez

Longitudinal angle
2
Longitudinal estimation error

-2
0 1 2 3 4 5 6
time [s]
Latitudinal angle

1
Latitudinal estimation error

-1
0 1 2 3 4 5 6
time [s]

Fig. 16.9 Flapping angles estimation errors at near-hover

Fig. 16.10 Inside view of the

Mazari autopilot box

Table 16.1 MIDG sensor Measurements Mean error (ME)

specifications
Angular rates
Range ±300◦ /s
Non-linearity 0.1% of FS
√
Noise density 0.05◦ /s/ Hz
3 dB Bandwidth 20 Hz
Acceleration
Range ±6 g
Non-linearity 0.3% of FS
√
Noise density 150 μg/ H z
3 dB bandwidth 20 Hz
Attitude accuracy
0.4◦ (1σ )
Position accuracy
2 m CEP, WAAS/EGNOS
16 On the Tip–Path-Plane Flapping Angles Estimation for Small–Scale. . . 357

Fig. 16.11 Maxi Joker 3 attitude response at near-hover

-1
Estimated
Longitudinal angle (degree)

-1.5

-2

-2.5

-3
4 4.2 4.4 4.6 4.8 5 5.2 5.4 5.6 5.8 6
Time (Sec)

Fig. 16.12 Longitudinal flapping estimation

estimating them in a stochastic framework has not been yet studied sufficiently for
small-scale platforms. As a future work of this research, intelligent state estimation
methods will be implemented to improve the accuracy of the flapping angles state
estimation. Another future research study could be in estimating the main and tail
rotors forces based on the estimated flapping angles. This would help in designing
a fault tolerant control design of the tail rotor at near-hover conditions.
358 M. K. Al–Sharman and M. F. Abdel–Hafez

1
Estimated
Latitudinal angle (degree)
0.5

-0.5

-1

-1.5
4 4.2 4.4 4.6 4.8 5 5.2 5.4 5.6 5.8 6
Time (Sec)

Fig. 16.13 Latitudinal flapping estimation

References

Abdel-Hafez, M. F., & Speyer, J. L. (2008). GPS measurement noise estimation in non ideal
environments. Navigation , 55(1), 55–66.
Abdel-Hafez, M. F. (2010). The autocovariance least-squares technique for GPS measurement
noise estimation. IEEE Transactions on Vehicular Technology, 59(2), 574–588.
Al-Sharman, M. (2015). Auto takeoff and precision landing using integrated GPS/INS/Optical flow
solution. Master thesis, American University of Sharjah, Sharjah.
Al-Sharman, M. (2016). Attitude estimation for a small-scale flybarless helicopter. In Multisensor
attitude estimation: Fundamental concepts and applications. Boca Raton: CRC Press.
Al-Sharman, M., Abdel-Hafez, M., & Al-Omari, M. (2014). State estimation for a small scale
flybar-less helicopter. In 2nd International Conference on System-Integrated Intelligence:
Challenges for Product and Production Engineering, Bremen, Germany.
Al-Sharman, M., Abdel-Hafez, M., & Al-Omari, M. (2015). Attitude and flapping angles
estimation for a small-scale flybarless helicopter using a Kalman filter. IEEE sensors Journal,
15(4), 2114–2122.
Al-Sharman, M. K., Emran, B. J. M., Jaradat, A., Najjaran, H., Al-Husari, R., & Zweiri, Y. (2018a).
Precision landing using an adaptive fuzzy multi-sensor data fusion architecture. Applied soft
computing, 69, 149–164.
Al-Sharman, M., Al-Jarrah, M., & Abdel-Hafez, M. (2018b). Auto takeoff and precision
terminal-phase landing using an experimental optical flow model for GPS/INS enhancement.
ASCE-ASME Journal of Risk and Uncertainty in Engineering Systems, Part B: Mechanical
Engineering, 5(1), pp. 011001-1–011001-17.
Allred, C. J., Churchill, D., & Buckner, G. D. (2017). Real-time estimation of helicopter rotor
blade kinematics through measurement of rotation induced acceleration. Mechanical Systems
and Signal Processing, 91, 183–197.
Alshoubaki, A., Al-Sharman, M., & Al-Jarrah, M. A. (2015). Flybarless helicopter autopilot
rapid prototyping using hardware in the loop simulation. In 10th International Symposium
on Mechatronics and its Applications (ISMA), Sharjah, UAE.
Chen, R. T. (1987). Flap-lag equations of motion of rigid, articulated rotor blade with three hinge
sequences. Washington, DC: NASA.
Chen, R. T. N. (1990). A survey of nonuniform inflow models for rotorcraft flight dynamics and
control applications. VERTICA, 2(14), 147–184.
16 On the Tip–Path-Plane Flapping Angles Estimation for Small–Scale. . . 359

Gavrilets, V. (2003). Autonomous aerobatic maneuvering of miniature. Ph.D. thesis, Massacheutes

Institute of Technology.
Gavrilets, V., Mettler, B., & Feron, E. (2001). Nonlinear model for a smallsize acrobatic
helicopter. In Proceedings of AIAA Guidance, Navigation, and Control Conference, Montreal,
QC, Canada, 2001, Paper AIAA-2001-4333.
Gavrilets, V., Martinos, I., Mettler, B., & Feron, E. (2002). Control logic for automated aerobatic
flight of miniature helicopter. In AIAA Guidance, Navigation and Control Conference,
Monterey, CA, USA 2002.
Hald, U., Hesselbaek, M., Holmgaard, J. T., Jensen, C. S., Jakobsen, S. L., & Siegumfeldt, M.
Autonomous helicopter-modelling and control. Aalborg: Aalborg University.
Johnson, W. (1981). Comparison of calculated and measured helicopter rotor lateral flapping
angles. Journal of the American Helicopter Society, 26(2), 46–50.
Leishman, J. G. (2000). Principles of helicopter aerodynamics. Cambridge: Cambridge University
Press.
Liu, H., Lu, G., & Zhong, Y. (2013). Robust LQR attitude control of a 3-DOF laboratory helicopter
for aggressive maneuvers. IEEE Transactions on Industrial Electronics, 60(10), 4627–4636.
Marantos, P., Bechlioulis, C. P., & Kyriakopoulos, K. J. (2017). Robust trajectory tracking control
for small-scale unmanned helicopters with model Uncertainties. IEEE Transactions on Control
Systems Technology, 25(6), 2010–2021.
Padfield, G. D. (1996). Helicopter flight dynamics: The theory and application of flying qualities
and simulation modeling (AIAA education series). AIAA, Washington, DC.
Perdomo, O., & Wei, F.-S. (2017). On the flapping motion of a helicopter blade. Applied
Mathematical Modelling, 46, 299–311.
Taamallah, S. (2011). Small-scale helicopter blade flap-lag equations of motion for a flybarless
pitch-lag-flap main rotor. In AIAA Modeling and Simulation Technologies Conference,
Portland.
Zappa, E., Trainelli, L., Liu, R., Rolando, A., Cordisco, P., Vigoni, E., & Redaelli, M. (2016). Real
time contactless sensor for helicopter blade angle measurement. In 2016 IEEE Metrology for
Aerospace (MetroAeroSpace), Florence, Italy.
Zhu, B., & Zuo, Z. (2017). Approximate analysis for main rotor flapping dynamics of a model-
scaled helicopter with Bell-Hiller stabilizing bar in hovering and vertical. Nonlinear Dynamics,
85(3), 1705–1717.