Application of Model-based Control in Autonomous Vehicle Chassis Control and

Motion Planning

by

Yijun Li

A dissertation submitted in partial fulfillment

of the requirements for the degree of
Doctor of Philosophy
(Mechanical Sciences and Engineering)
in the University of Michigan-Dearborn
2021

Doctoral Committee:

Professor Taehyun Shim, Chair

Assistant Professor Youngki Kim
Dexin Wang, Ford Motor Company
Professor Yi Zhang
 Yijun Li
[email protected]
ORCID iD: 0000-0002-8559-4690

© Yijun Li, 2021

Dedication

To my family

ii
 Acknowledgements

First and foremost, I would want to express my gratitude to Professor Taehyun Shim, who
served as my advisor. Working with Professor Shim over the last few years has been a
pleasure. His understanding and experience in the controls and automotive dynamics have
been tremendously beneficial and informative for my study. I’m grateful for the knowledge
he has shared with me, as well as the fruitful and gratifying collaborative relationship we
have had over the years. We had numerous intriguing talks throughout this research that
generated and shaped many of the concepts offered in this dissertation. Getting to know him
and learning from him has been a true pleasure.

I would like to thank Professor Yi Zhang, Assistant Professor Youngki Kim, and Dr.
Dexin Wang for serving on the dissertation committee. Professor Zhang provided several
helpful comments on the system modeling. Professor Kim’s Predictive Control of Dynamic
Systems course introduced me to the field of model predictive control. Dr. Wang guided me
through my first Ford-UM alliance research project, and provided the test bench for validating
my control algorithm. My gratitude also extends to those who worked with me in various
research projects. Specifically, Timothy Offerle offered many helpful suggestions on the first
alliance project. Lodewijk Wijffels and Joshua Guerra helped guide the second alliance
project. I also appreciate Dong-Hwan Shin, Seonghun Lee, and Sungho Jin from DIGST for
conducting experimental tests of the electromechanical braking system and providing helpful
suggestions on the control system development.

I would like to thank my colleagues, Chinmay Patil, Bhushan Naik, Aditya Sanjay
Kothari, Arun Joshi, Heeseong Kim, Syed Adil Ahmed, and Pavan Anandareddy in our
research group. I also wish to thank my friends, Wei Guo, Xiaopeng Guo, and Yiteng Tu at
the University of Michigan-Dearborn. I cannot imagine completing this long journey without
their continuous encouragement and support.

iii
 Table of Contents

Dedication ii

Acknowledgements iii

List of Figures viii

List of Tables xiii

List of Appendices xiv

Abstract xv

Chapter 1: Introduction 1
1.1 Motivation and Objectives . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.3 Dissertation Outline . . . . . . . . . . . . . . . . . . . . . . . . . . . 6

I Control and Estimation for Vehicle Chassis System 8

Chapter 2: Steering Feel Control for EPAS System 9

2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.2 EPAS System Modeling . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.2.1 Mechanical Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.2.2 Basic Control Strategy . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.3 Torque Overlay Control System . . . . . . . . . . . . . . . . . . . . . 20
2.3.1 Reference Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
2.3.2 Tracking Controller . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.4 Rack Force and Load-Dependent Friction Estimation . . . . . . . . . 26
2.4.1 Disturbance Observer based Rack Force Estimator . . . . . . . . . . 26
2.4.2 Load-Dependent Friction . . . . . . . . . . . . . . . . . . . . . . . . 28

iv
 2.5 Simulation and Experimental Results . . . . . . . . . . . . . . . . . 31
2.5.1 Rack Force Estimation Simulation . . . . . . . . . . . . . . . . . . . 31
2.5.2 Steering Feel Control Simulation . . . . . . . . . . . . . . . . . . . . 35
2.5.3 Hardware-in-the-Loop Test . . . . . . . . . . . . . . . . . . . . . . . 37
2.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42

Chapter 3: Clamping Force Control for Electromechanical Brake System 43

3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
3.2 EMB Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
3.2.1 Mechanical Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
3.2.2 Clamping Force Model . . . . . . . . . . . . . . . . . . . . . . . . . . 49
3.3 Clamping Force Control System Development . . . . . . . . . . . . . 55
3.3.1 Clamping Force Estimator . . . . . . . . . . . . . . . . . . . . . . . 55
3.3.2 Gap Distance Estimation . . . . . . . . . . . . . . . . . . . . . . . . 58
3.3.3 Clamping Force Tracking . . . . . . . . . . . . . . . . . . . . . . . . 62
3.3.4 Unified Control Structure . . . . . . . . . . . . . . . . . . . . . . . . 67
3.4 Simulation Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
3.4.1 Clamping Force Estimation . . . . . . . . . . . . . . . . . . . . . . . 70
3.4.2 Clamping Force Tracking Control . . . . . . . . . . . . . . . . . . . . 72
3.4.3 Robust Performance Evaluation . . . . . . . . . . . . . . . . . . . . 76
3.5 Conculsions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81

II Optimization-based Control for High-Level Motion Planning and Tra-

jectory Tracking 82

Chapter 4: NMPC based Trajectory Tracking Control for Vehicle with

4WS and 4WID 83
4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
4.2 Vehicle Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
4.2.1 Vehicle Model Development . . . . . . . . . . . . . . . . . . . . . . . 86
4.2.2 Tire Slip and Force . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
4.2.3 Load Transfer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
4.3 Nonlinear Model Predictive Control of Trajectory Tracking . . . . . 92
4.3.1 Formulation of the Optimal Control Problem . . . . . . . . . . . . . 92
4.3.2 Objective Function with Frequency Shaping . . . . . . . . . . . . . . 93
4.3.3 Fast NMPC based on RTI and Control Parameterization . . . . . . . 97

v
 4.3.3.1 Sparse QP Construction . . . . . . . . . . . . . . . . . . . . . . . . . 101
4.3.3.2 The First Condensing . . . . . . . . . . . . . . . . . . . . . . . . . . 103
4.3.3.3 Control Parameterization . . . . . . . . . . . . . . . . . . . . . . . . 105
4.3.3.4 The Second Condensing . . . . . . . . . . . . . . . . . . . . . . . . . 110
4.3.3.5 Solution Expansion . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
4.4 Simulation Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
4.4.1 High µ Braking . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112
4.4.2 Low µ Braking . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112
4.4.3 Split µ Braking . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
4.4.4 Sine with Dwell . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
4.4.5 High Speed Double Lane Change . . . . . . . . . . . . . . . . . . . . 120
4.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129

Chapter 5: Motion Planning for Autonomous Vehicle based on NMPC 130

5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130
5.2 Vehicle Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
5.2.1 Point-Mass Vehicle Model . . . . . . . . . . . . . . . . . . . . . . . . 131
5.2.2 Spatial Transformation . . . . . . . . . . . . . . . . . . . . . . . . . 133
5.2.3 Convert Cartesian Coordinates to Frenet Coordinates . . . . . . . . 133
5.3 Motion Planning Problem Definition . . . . . . . . . . . . . . . . . . 134
5.3.1 Objective Function . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134
5.3.2 Control Input Constraints . . . . . . . . . . . . . . . . . . . . . . . . 135
5.3.3 State Constraints . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135
5.3.4 Waypoint Constraints . . . . . . . . . . . . . . . . . . . . . . . . . . 136
5.4 Collision Avoidance Constraints based on Control Barrier Function . 137
5.4.1 Conventional Formulation of Collision Avoidance Constraint . . . . . 137
5.4.2 Control Barrier Function for Safety Critical System . . . . . . . . . . 139
5.4.3 HOCBF based Collision Avoidance Constraint for NMPC . . . . . . 142
5.4.4 HOCBF with Vehicle Speed Scheduling . . . . . . . . . . . . . . . . 144
5.5 Summary of NMPC Formulation and Numerical Method . . . . . . . 147
5.6 Simulation Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149
5.6.1 Obstacle Avoidance with Double Lane Change . . . . . . . . . . . . 150
5.6.2 Oval Circuit Tracking with Obstacles . . . . . . . . . . . . . . . . . 153
5.7 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 158

Chapter 6: Summary and Future Work 159

6.1 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159

vi
 6.1.1 EPAS Steering Feel Control . . . . . . . . . . . . . . . . . . . . . . . 159
6.1.2 EMB Clamping Force Control . . . . . . . . . . . . . . . . . . . . . 159
6.1.3 NMPC based Trajectory Tracking Control . . . . . . . . . . . . . . . 160
6.1.4 NMPC based Motion Planning . . . . . . . . . . . . . . . . . . . . . 160
6.2 Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161
6.2.1 EPAS Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161
6.2.2 EMB Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161
6.2.3 Trajectory Tracking Control . . . . . . . . . . . . . . . . . . . . . . 162
6.2.4 Motion Planning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 163

Appendices 164

References 170

vii
 List of Figures

Figure 1.1 Schematic of autonomous vehicle system . . . . . . . . . . . . . . . . 2

Figure 1.2 Dissertation layout . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7

Figure 2.1 Steering feel curve with and without power assistance . . . . . . . . 10
Figure 2.2 Schematic of C-EPAS model . . . . . . . . . . . . . . . . . . . . . . 13
Figure 2.3 Schematic of intermediate shaft . . . . . . . . . . . . . . . . . . . . 14
Figure 2.4 Schematic of worm gear . . . . . . . . . . . . . . . . . . . . . . . . 15
Figure 2.5 LuGre fiction characteristics . . . . . . . . . . . . . . . . . . . . . . 16
Figure 2.6 Friction level validation . . . . . . . . . . . . . . . . . . . . . . . . . 18
Figure 2.7 Block diagram of C-EPAS assist control strategy . . . . . . . . . . . 18
Figure 2.8 Boost curve for power assistance . . . . . . . . . . . . . . . . . . . . 19
Figure 2.9 Schematic of Feed-forward steering feel control strategy . . . . . . . 19
Figure 2.10 Overall schematic of C-EPAS steering feel control . . . . . . . . . . 20
Figure 2.11 Schematic of reference model with C-EPAS and basic control systems 21
Figure 2.12 Comparison of steering feel hysteresis with reference and plant EPAS
model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
Figure 2.13 Rack force estimation without friction estimation . . . . . . . . . . . 28
Figure 2.14 C-EPAS system friction vs. assist motor torque . . . . . . . . . . . . 28
Figure 2.15 Relationship between applied torque and angle . . . . . . . . . . . . 30
Figure 2.16 Schematic of rack force estimation with friction estimation . . . . . . 31
Figure 2.17 Total applied torque vs. steering angle . . . . . . . . . . . . . . . . 32
Figure 2.18 Iterative regression results . . . . . . . . . . . . . . . . . . . . . . . 32
Figure 2.19 Load-dependent friction vs. motor torque . . . . . . . . . . . . . . . 33
Figure 2.20 Friction estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
Figure 2.21 Rack force estimation . . . . . . . . . . . . . . . . . . . . . . . . . . 34
Figure 2.22 Vehicle trajectory and steering angle of DLC maneuver . . . . . . . 35
Figure 2.23 Friction and rack force estimation of DLC maneuver . . . . . . . . . 36
Figure 2.24 Comparison of steering feel hysteresis with and without torque overlay
control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36

viii
Figure 2.25 Comparison of tracking controller performance . . . . . . . . . . . . 37
Figure 2.26 Sine-Sweep steering feel frequency response . . . . . . . . . . . . . . 38
Figure 2.27 EPAS hardware-in-the-loop simulator . . . . . . . . . . . . . . . . . 38
Figure 2.28 Rack force estimation of HIL test . . . . . . . . . . . . . . . . . . . 39
Figure 2.29 HIL test results of torque overlay control . . . . . . . . . . . . . . . 39
Figure 2.30 Torque-to-rotate test results . . . . . . . . . . . . . . . . . . . . . . 40
Figure 2.31 Rack force estimation with increased pinion friction . . . . . . . . . 41
Figure 2.32 Steering feel of steering system with increased and nominal friction 41

Figure 3.1 EMB schematic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46

Figure 3.2 Schematic of ball screw . . . . . . . . . . . . . . . . . . . . . . . . . 47
Figure 3.3 Clamping force of EMB . . . . . . . . . . . . . . . . . . . . . . . . . 50
Figure 3.4 Clamping force measurement . . . . . . . . . . . . . . . . . . . . . . 50
Figure 3.5 Clamping force modeling . . . . . . . . . . . . . . . . . . . . . . . . 51
Figure 3.6 Comparison between clamping force model and experimental results 54
Figure 3.7 High-level schematic of clamping force control system . . . . . . . . 55
Figure 3.8 Motor angle input for clamping force estimation in applying . . . . . 57
Figure 3.9 Closed loop speed control system . . . . . . . . . . . . . . . . . . . . 59
Figure 3.10 Simulation results of IMP based speed control . . . . . . . . . . . . . 60
Figure 3.11 Simulation results of contact detection . . . . . . . . . . . . . . . . . 62
Figure 3.12 Block diagram of the DOB control of clamping force . . . . . . . . . 63
Figure 3.13 Step response of linear and nonlinear EMB models . . . . . . . . . . 65
Figure 3.14 Magnitude response of Gp (z −1 ) and ∆ (z −1 ) . . . . . . . . . . . . . . 66
Figure 3.15 Q-filter bandwidth selections . . . . . . . . . . . . . . . . . . . . . . 66
Figure 3.16 Frequency response of Kr (z −1 ) and Kr (z −1 ) Kf f (z −1 ) Gcl (z −1 ) . . . 68
Figure 3.17 High-level schematic of unified control system . . . . . . . . . . . . . 68
Figure 3.18 Augmented clamping force . . . . . . . . . . . . . . . . . . . . . . . 69
Figure 3.19 Identification of applying clamping force . . . . . . . . . . . . . . . . 70
Figure 3.20 Identification of releasing clamping force . . . . . . . . . . . . . . . . 71
Figure 3.21 Simulation of stair-step reference with large amplitude . . . . . . . . 73
Figure 3.22 Simulation of stair-step reference with small amplitude . . . . . . . . 74
Figure 3.23 Clamping force tracking of sinusoidal reference with 50% modulation 75
Figure 3.24 Clamping force tracking of sinusoidal reference with 10% modulation 76
Figure 3.25 Comparison with different controller . . . . . . . . . . . . . . . . . . 77
Figure 3.26 Effect of time delay . . . . . . . . . . . . . . . . . . . . . . . . . . . 78

ix
Figure 3.27 Simulation with model parameters variation . . . . . . . . . . . . . . 79
Figure 3.28 Simulation with noisy measurement . . . . . . . . . . . . . . . . . . 80

Figure 4.1 Schematic of NMPC trajectory tracking control system . . . . . . . 85

Figure 4.2 Schematic of chassis model . . . . . . . . . . . . . . . . . . . . . . . 86
Figure 4.3 Wheel local coordinate system and slip angle . . . . . . . . . . . . . 88
Figure 4.4 Wheel longitudinal slip schematic . . . . . . . . . . . . . . . . . . . 89
Figure 4.5 Tire force form Magic Formula Model with F z = 5718N . . . . . . . 90
Figure 4.6 Tire force form under combined slip condition with F z = 5718N . . 90
Figure 4.7 Comparison between actuator embedding and frequency shaping methods 97
Figure 4.8 Fast NMPC algorithm timeline . . . . . . . . . . . . . . . . . . . . . 101
Figure 4.9 Linear interpolating parameterization with np = 4 . . . . . . . . . . 108
Figure 4.10 High µ (µ = 1) braking with target deceleration of 1g . . . . . . . . 113
Figure 4.11 Low µ (µ=0.4) braking with target deceleration of 1g . . . . . . . . 115
Figure 4.12 Low µ (µ=0.4) braking with no frequency shaping and increased
control input weighting . . . . . . . . . . . . . . . . . . . . . . . . . 116
Figure 4.13 Control input weighting . . . . . . . . . . . . . . . . . . . . . . . . . 116
Figure 4.14 Split µ (left: µ = 1; right: µ = 0.4) braking with target deceleration
of 0.4g . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118
Figure 4.15 Split µ (left: µ = 1; right: µ = 0.4) braking with target deceleration
of 0.3g . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
Figure 4.16 Front wheel steering angle of sine with dwell simulation . . . . . . . 120
Figure 4.17 Open-loop trajectory of sine with dwell simulation. The green solid
line denotes vehicle C.G. trajectory, and the black polygon denotes
vehicle’s shape. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120
Figure 4.18 Closed-loop control of sine with dwell (µ = 1) . . . . . . . . . . . . . 121
Figure 4.19 Control inputs profiles of sine with dwell. The solid line denotes the
response of 4WS and the dashed line denotes the response of FWS. 122
Figure 4.20 DLC on asphalt at 140kph (µ=1). Top plot: the vehicle with 4WS is
outlined by black polygon, and the one with FWS is outlined by red
polygon. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
Figure 4.21 Vehicle state responses of DLC on asphalt (µ=1) at 140kph . . . . . 123
Figure 4.22 Control input demands of DLC on asphalt (µ=1) at 140kph. The
solid line denotes 4WS, and the dashed line denotes FWS. . . . . . . 124

x
Figure 4.23 DLC on snowy surface (µ=0.4) at 70kph. Top plot: the vehicle with
4WS is outlined by black polygon, and the one with FWS is outlined
by red polygon. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
Figure 4.24 Vehicle state responses of DLC on snowy surface (µ=0.4) at 70kph . 125
Figure 4.25 Control input demands of DLC on snowy surface (µ=0.4) at 70kph.
The solid line denotes 4WS, and the dashed line denotes FWS. . . . 126
Figure 4.26 DLC on icy surface (µ=0.17) at 50kph. Top plot: the vehicle with
4WS is outlined by black polygon, and the one with FWS is outlined
by red polygon. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126
Figure 4.27 Vehicle state responses of DLC on icy surface (µ=0.17) at 50kph . . 127
Figure 4.28 Control input demands of DLC on icy surface (µ=0.17) at 50kph. The
solid line denotes 4WS, and the dashed line denotes FWS. . . . . . . 127
Figure 4.29 Computation times of DLC simulations . . . . . . . . . . . . . . . . 128

Figure 5.1 Schematic of point mass model . . . . . . . . . . . . . . . . . . . . . 132

Figure 5.2 Collision avoidance by road boundary tightening . . . . . . . . . . . 137
Figure 5.3 Collision avoidance using Euclidean norm . . . . . . . . . . . . . . . 138
Figure 5.4 Comparison between space based and time based NMPC formulation 145
Figure 5.5 Effect of λ on collision avoidance . . . . . . . . . . . . . . . . . . . . 146
Figure 5.6 Direct collocation discretization . . . . . . . . . . . . . . . . . . . . 148
Figure 5.7 Motion planning and tracking control system . . . . . . . . . . . . . 150
Figure 5.8 Road configuration for DLC. The blue solid line denotes lane con-
dounary. The blue dashed line denotes lane center; red dashed line
denotes obstacle. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150
Figure 5.9 Collision avoidance at 70km/h (µ = 1). The red dotted ellipse denotes
CBF constraint. The blue dotted ellipse denotes distance constraint. 151
Figure 5.10 Collision avoidance at 85km/h (µ = 1). The magenta rectangle denotes
the shape of controlled vehicle. . . . . . . . . . . . . . . . . . . . . 151
Figure 5.11 Collision avoidance at various vehicle speeds (µ = 1) . . . . . . . . . 152
Figure 5.12 Collision avoidance with lateral acceleration constraints (µ = 1) . . 153
Figure 5.13 Collision avoidance with lateral acceleration constraints and additional
speed constraints (µ = 1) . . . . . . . . . . . . . . . . . . . . . . . . 154

xi
Figure 5.14 Motion planning with waypoint constraints (µ = 1) . . . . . . . . . 154
Figure 5.15 Oval circuit tracking with heavy weighting on lateral deviation and
heading angle deviation (µ = 1) . . . . . . . . . . . . . . . . . . . . 156
Figure 5.16 Oval circuit tracking with large and small weightings on speed (µ = 1) 157

Figure B.1 Brake torque profile . . . . . . . . . . . . . . . . . . . . . . . . . . . 167

Figure B.2 Slip ratio response . . . . . . . . . . . . . . . . . . . . . . . . . . . 167

xii
 List of Tables

Table 3.1 The sign of measurements and their derivatives before/after the contact 58
Table 3.2 Asymptotic properties of transfer functions with DOB control . . . . 64
Table 3.3 Uncertain parameters of plant model . . . . . . . . . . . . . . . . . . 65
Table 3.4 Control performance of sinusoidal reference with 50% modulation . . 75
Table 3.5 Control performance of sinusoidal reference with 10% modulation . . 75

Table 4.1 Tracking performance comparison . . . . . . . . . . . . . . . . . . . 129

Table A.1 Vehicle parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . 164

Table A.2 Tire parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 164
Table A.3 Parameters for the double lane change reference path . . . . . . . . . 165
Table A.4 NMPC tracking controller parameters . . . . . . . . . . . . . . . . . 165
Table A.5 NMPC motion planning controller parameters . . . . . . . . . . . . . 165

xiii
 List of Appendices

Appendix A List of Model and Controller Parameters . . . . . . . . . . . . . . . 164

Appendix B Determining Rate Constraint of Wheel Slip . . . . . . . . . . . . . 166
Appendix C Moving Obstacle Collision Avoidance Constraint . . . . . . . . . . 169

xiv
 Abstract

Automatic control systems are found ubiquitously in today’s automobiles. This dissertation
focuses on the development of control systems for autonomous vehicle applications using
model-based control design approaches. We start with the control system development for
chassis actuators. Firstly, we develop a closed loop torque overlay control system to improve
the steering feel of an electric power-assisted steering (EPAS) system. This system has a
reference model, a rack force estimator, and a tracking controller. A target steering feel is
generated from a reference model with an rack force estimate to reflect the actual vehicle
operating conditions. The performance of the proposed control system is evaluated through
simulation and a hardware-in-the-loop test. Secondly, we present an EMB clamping force
control system is developed that addresses several major challenges in practical implementation.
A nonlinear EMB model including a novel clamping force model is introduced. A clamping
force estimation and contact detection algorithm is proposed that requires only the existing
measurements. Furthermore, a unified architecture is proposed to realize a smooth transition
between gap closing and clamping force tracking. The performance of the control system is
evaluated based on simulation.

In the second part, we developed a hierarchical nonlinear model predictive control

(NMPC) framework for autonomous vehicle motion planning and control. At low level,
a trajectory tracking NMPC is used to track the reference trajectory given by the high-
level motion planner. A frequency shaped objective function is used to incorporate lower
level actuator dynamics. Real-time implementation is realized by a fast NMPC algorithm
based on RTI with condensing and control parameterization. Simulation results show
that the control system yields good performance under various road conditions when the
vehicle is operating at its handling limit. A high-level motion planner is formulated in the
NMPC framework incorporating high order control barrier function (HOCBF) based collision
avoidance constraint. The motion planner is able to generate dynamically feasible trajectories
with respect to various constraints. Simulations of the motion planner and trajectory tracking
controller demonstrate the effectiveness of the proposed control framework.

xv
 Chapter 1

Introduction

1.1 Motivation and Objectives

Automatic control has been playing an important role in automotive innovation since the
1990s. Tremendous efforts have been put into control systems development in the fields of
powertrain, driveline, chassis, and recently advanced driver assistance systems, in order to
meet the enhanced requirements for lower fuel consumption, lower emissions, and improved
safety and comfort demands.

In the area of vehicle chassis control, attention is mainly focused on the development of
active safety systems. Various active safety control systems, such as anti-lock braking system
(ABS) [1], electronic stability control (ESC) [2], and active front steering (AFS) [3] have been
commercially deployed on production vehicles. These standalone control algorithms aim to
improve the vehicle’s safety and stability by using a particular type of actuator, e.g. brake or
steering system. As more and more actuators are available to control and the requirements
become increasingly demanding, the amount of work on control systems development and
calibration quickly becomes overwhelming. This motivated the development of the integrated
vehicle dynamics control (IVDC) architecture in the 2000s [4]. In the IVDC framework, all
the vehicle subsystems are coordinated coherently by control algorithms to achieve multiple
control functions, such as yaw motion control, lateral motion control, and longitudinal motion
control. Control allocation is often used to distribute required chassis force and moment over
different actuators.

In recent years, conventional active safety control has evolved into a new generation of
control systems, namely the advanced driver assistance (ADAS) systems [5]. The introduction
of advanced sensing and computation systems enables new active safety functionalities such as
adaptive cruise control (ACC), automatic emergency braking (AEB), and collision imminent
steering (CIS).

1
 Figure 1.1: Schematic of autonomous vehicle system

The ultimate goal of the active safety system is fully autonomous driving. Autonomous
driving not only improves driving safety, but also traffic density and fuel efficiency. Various
research institutes and companies have invested a large amount of effort in the research and
development of autonomous vehicles. Fig. 1.1 shows a typical schematic of an autonomous
vehicle system. The whole system consists of sensors, perception systems, decision making
systems, and vehicle control systems [6–8]:

• The system has various types of sensors for measuring the states of ego vehicle and
surrounding objects. Some of the most commonly used sensors for autonomous vehicle
include GPS, IMU sensors, cameras, radars, and lidars.

• The perception system processes the measurements from various sensors to construct
the position of ego vehicle as well as environmental information around the vehicle,
such as lane markers, traffic signs and signals, obstacles, etc.

• The decision making system is responsible for planning a route between the vehicle’s
starting location and destination. The route can be a geometrical path or a set of way
points. Based on the planned route and environmental information, the behavioral
planning algorithm determines which action the vehicle should take, such as lane keeping,
lane changing, overtaking, etc. Finally, reference paths or waypoints are generated for
the lower level control system.

• The motion control system is responsible for guiding the vehicle to follow the high
level reference path while ensuring driving safety. A hierarchical architecture is usually

2
 adopted to separate the tasks into local motion planning and trajectory tracking. The
local motion planning generates a vehicle’s motion trajectory over a horizon based
on the reference path, vehicle states, and surrounding objects. Then the trajectory
tracking controller computes the required actuation like steering angle, braking/driving
force, etc, such that the vehicle follows closely to the reference trajectory.

• The actuation system consists of different low-level actuators such as steering, braking,
powertrain, and suspension systems. It takes the commands from the motion control
system and applies appropriate control input to each individual system to track the
reference command.

In this dissertation, we aim to develop control systems for the motion control layer and
the actuation layer. In particular, model-based control design methodology is used for
control algorithm development. Traditionally, in the automotive industry, model-free control
techniques are often applied. These methods are based on the time-consuming iterative
calibrations of lookup tables, PID controller gains, and so on. This tuning methods might
work well for standalone control algorithms like ABS or ESC, but quickly become intractable
when dealing with complex chassis control such as the motion control system in autonomous
driving applications.

On the contrary, model-based control design is a more systematic approach that relays on
knowledge of system dynamics for controller development. Controllers are synthesized based
on the control-oriented model and several weightings and tuning parameters. Simulations are
used to evaluate the control performance, and the controller can be retuned conveniently by
changing the corresponding weights and parameters based on the simulation results.

This dissertation presents our work on how model-based design approaches can be used
for control system development of autonomous vehicle. To be specific, we try to stress how
several important requirements can be addressed when designing model-based control systems
for autonomous vehicle:

• Control structure that is easy to tune: An increasing number of algorithms are added
to the existing base vehicle control systems to realize additional functionalities. The
designed control algorithm should allow fast and convenient tuning to achieve target
performance. For example, the control algorithm for the electric power-assisted steering
(EPAS) system should enable flexible adjustment for steering feel such that the car

3
 maker can distinguish its products of different classes, e.g. family, luxury, sport, etc.
The motion planning algorithm should be able to generate vehicle motions according to
personal driving style, e.g. comfort, aggressive, etc. In this dissertation, we will see how
reference model based torque overlay control is used for convenient adjustment of target
steering feel on the basis of existing power assist functions. In electromechanical brake
(EMB) control, we show how clamping force tracking bandwidth and gap closing speed
can be adjusted by a couple of intuitive parameters. And for motion planning and
control for autonomous driving, control performance can be tuned by relative weights
of control objectives versus control inputs.

• Control robustness with respect to disturbances and model uncertainties: Vehicle control
systems frequently experience various external disturbances and uncertainties, e.g.
mechanical friction, measurement noise and delay, during operations. The controller’s
performance should be robust against these factors. In the development of EPAS
steering feel control, we use sliding mode controller to realize robust reference tracking
against external disturbance. In the development of EMB clamping force control, we
will see a disturbance-observer based controller is used to find the balance between fast
tracking with disturbance rejection and robust stability against model mismatch.

• Estimation of system parameters and states using limit sensor measurements: Due to
hardware and cost limitations, some model parameters and states necessary for control
system are not available by sensor measurement. Estimation techniques must be used
to reconstruct these information. In the first part of the dissertation, we will see how
pragmatic estimation algorithms are developed to effectively estimate some essential
model parameters and states such as EPAS rack force, EMB clamping force and gap
distance, for the control systems.

• Interfacing between high-level and low-level control systems: Controllers in different

layers of the vehicle control system are often designed separately with different specifi-
cations and assumptions. For example, the actuator controllers in the actuation layer
usually operate at a much faster rate than those in the motion control layer. And the
motion control system might be developed based on simplified actuation models with
some assumptions. Therefore, controllers in different layers should work cohesively
in order to achieve good overall control performance. In this work, we show how the
trajectory tracking controller takes into account of actuator dynamics using frequency
shaping technique, how temporal based trajectory tracking nonlinear model predictive
control (NMPC) and spatial based motion planning NMPC work together, and how

4
 the motion planner takes care of commands from high-layer decision making systems
by formulating them into various path constraints.

• Real-time feasible controller design: Optimization-based controllers like the one proposed
in the second part of this dissertation are usually computationally intensive, since
optimization must be solved online in real-time. We show how specific numerical
methods can be applied with control objectives and system characteristics in mind,
such that real-time feasible control is realized.

1.2 Contributions

The major contributions of this dissertation are summarized as follows:

EPAS Steering Control

We propose a torque overlay control structure that achieves a flexible and intuitive steering feel
adjustment without tedious calibrations on feed-forward gains and lookup tables. Specifically,
a high-fidelity EPAS model reflecting realistic dynamics of the system is developed. In
particular, load-dependent fiction is included in the model and its effect on steering feel
is investigated. We propose a novel algorithm for load-dependent fiction estimation. By
incorporating it with a Kalman filter, accurate rack force estimation is achieved. The rack
force estimate is then used for the reference model to generate a realistic target steering feel.
Hardware-in-the-loop simulations are conducted which demonstrate the effectiveness of the
proposed estimation and control algorithm.

EMB Clamping Force Control

A novel clamping force model validated by testing results, which is able to accurately
capture the nonlinearity and hysteresis effect. A straightforward process is also presented
for estimating the clamping force. We develop a simple and effective gap distance estimator
based on the Kalman smoother and a robust clamping force tracking controller based on
the disturbance observer. In order to seamlessly integrate the gap closing and force tracking
without a controller switching logic, a unified clamping force control scheme is designed.
Extensive simulations are conducted to verify the effectiveness and robustness of the proposed
control system.

5
NMPC based Trajectory Tracking Controller

A NMPC based trajectory tracking controller for vehicles equipped with four-wheel-steering
(4WS) and four-wheel-independent-driving (4WID) are developed. In particular, a frequency
shaped objective function is used for NMPC formulation to accommodate the dynamics
of low-level actuators. A fast NMPC solving algorithm based on real-time-iteration (RTI)
scheme with control parameterization is proposed to achieve a fast sample time with real-time
feasibility. The performance of the proposed controller on vehicle stabilization and trajectory
tracking is demonstrated by various simulations.

NMPC based Motion Planner for Autonomous Vehicle

A NMPC based motion planner for autonomous vehicle is developed. The motion planning
problem is formulated in the spatial domain to enable simple constraint formulations. The
NMPC simultaneously plans lateral and longitudinal trajectories of the vehicle that satisfy
various constraints on control inputs, vehicle states, waypoints, and collision avoidance. We
proposed a novel collision avoidance formulation in the NMPC framework using high-order
control barrier functions (HOCBF). The HOCBF based collision avoidance constraint provides
formal certification of safety in the sense of set invariance, which is able to ensure safety even
with short prediction horizon. In order to allow flexible constraints on waypoints with fast
computation, NMPC problem is solved based on the direct collocation method.

1.3 Dissertation Outline

Fig. 1.2 shows the layout of this dissertation. The dissertation consists of 6 chapters, which
are organized as follows. Referring to Fig. 1.2, the main 4 chapters (Chapter 2 to Chapter
5) are divided into 2 parts. Part I includes Chapters 2 and 3, focusing on the control and
estimation development for low-level vehicle chassis system. In Chapter 2, the development
of EPAS steering feel control system is discussed. Chapter 3 discusses the clamping force
control system for EMB. Part II consists of Chapters 4 and 5, containing the development of
motion control system for autonomous vehicles. Chapter 4 describes the NMPC trajectory
tracking control design. And Chapter 5 discusses the details of motion planning NMPC
development. Finally, Chapter 6 summarizes the outcomes of the preceding chapters and
concludes the study. Possible future work is discussed as well.

6
Figure 1.2: Dissertation layout

7
 Part I

Control and Estimation for Vehicle

Chassis System

8
 Chapter 2

Steering Feel Control for EPAS System

The steering system is a key component of an automobile, affecting the vehicle handling
and stability and the driver’s comfort. It is also a key actuator for delivering the advanced
driver assistance or autonomous driving functions. However, the conventional power assist
control involves extensive tuning effort to meet the desired steering feel requirements. This
chapter aims to develop a steering feel control system for EPAS based on torque overlay. The
proposed control algorithm is able to realize the target steering feel given by a reference model.
Different steering feels can be achieved by simple tuning of the reference model parameters
without modifying the power assist functions.

EPAS system is also an essential part for autonomous vehicle. When the driver’s input is
absent as in autonomous vehicle, the assist motor of EPAS system is responsible for generating
required steering angle to control the vehicle’s motion. We will see from the development
process that the EPAS steering feel control shares many common challenges with autonomous
vehicle position control in terms of system modeling, disturbance estimation, and control.
And the outcomes of this chapter can be straightforwardly applied to autonomous vehicle
application.

2.1 Introduction

Electric power-assisted steering (EPAS) has become widely used in the automotive industry
to replace the conventional hydraulic power assist steering (HPAS) systems because of its
efficiency, modularity, and flexibility [9]. The EPAS system uses an electric motor rather than
a hydraulic pump to provide assist torque such that the driver’s steering effort is reduced.
Moreover, EPAS provides more freedom than the HPAS system to achieve advanced driver
assist functions, such as lane-keeping assist, lane-change assist, automatic parking, etc.

Despite the rapid development of advanced steering control systems, a fundamental

9
 Figure 2.1: Steering feel curve with and without power assistance

issue has continued to trouble developers and engineers since the appearance of the EPAS
system—the steering feel. Steering feel is defined as the haptic feedback from the steering
wheel to the driver’s hand. It originates from the tires’ self-aligning moment while steering
and conveys real-time information on the vehicle’s dynamic status to the human driver. Thus,
a good steering feel delivered by the EPAS system is critical for the driver’s maneuvering of
the vehicle [10]. However, good steering feel is not easy to achieve.

Steering feel is commonly evaluated by on-center handling test, where the steering angle
is small and at low frequency and the vehicle is at high speed [11]. This characterizes typical
highway situation, where lane-keeping and lane-changing are the primary maneuvers. As a
important index of the on-center handling quality, steering feel represents the torque feedback
to the driver with regard to the steering angle. The steering feel of Column-EPAS (C-EPAS)
can be depicted by hysteresis curves of steering torque versus steering wheel angle. Figure
2.1 shows a example of steering feel curve of C-EPAS. Roughly speaking, the objective of
steering feel control algorithm is to effectively shape this curve according to vehicle’s driving
conditions such that target steering feel is achieved.

The conventional methods for steering feel control are based on feed-forward ap-
proaches.The feed-forward approaches usually adopt different look-up tables and filters
to calculate the assist torque. For example, Ziman He [9] developed three steering control
modes as well as various compensation controls based on different boost curves. In [12],
the assist torque is calculated by nonlinear functions so that the yaw rate oscillation is

10
well damped and steering wheel returnability is improved. Yamazaki et al. [13] designed a
steering torque compensation strategy to cancel the influence of vehicle dynamics on steering
torque. One advantage of feed-forward control is that it is an analogue of the conventional
HPAS control method, and thus its implementation is straightforward. Two main issues,
however, limit the application of the feed-forward control methods. On the one hand, however,
the development of feed-forward control typically involves the tuning of a blend of look-up
tables and compensators in a range of vehicle speeds in order to achieve the target steering
feel. Therefore, the development procedure is extremely time-consuming, often requiring the
iterative tuning of the controller and target steering feel. On the other hand, feed-forward
control lacks robustness against model uncertainties and disturbances. Since the assist torque
is determined in feed-forward fashion, the resulting steering feel is subject to change in the
presence of parameter variations, e.g., inertia and road friction, and external disturbance,
e.g., road disturbance force and friction.

Distinct from the feed-forward methods, the feedback approaches control the steering
feel in a closed-loop scheme such that the response of the EPAS system tracks the target
steering feel. For instance, Tao [14] applied admittance control to an EPS system such that
the steering system has the desired admittance. Alaa Marouf et al. [15] designed a controller
for an EPS system based on the reference model to make the motor angle track the ideal one.
Carlos et al. [16] presented a position-model reference control strategy for an EPS system.
M. Moradkhani [17] et al. employed an H∞ loop-shaping control method in the EPS system
steering torque control. The feedback methods have two main advantages. Firstly, the control
system is more robust in the presence of model uncertainties and disturbances. Therefore, the
steering feel is kept the same regardless of the vehicle parameters and operation conditions.
Another advantage is that the design of the feedback controller is independent from the target
steering feel. Thus, tremendous tuning effort is saved.

Based on the preceding discussion, the feedback method is clearly more promising for
EPAS steering feel control. However, there are still open questions regarding the design of
feedback steering feel control. The steering feel of EPAS with feedback control is realized
by tracking a reference signal. Therefore, the choice of reference signal directly affects the
steering feel. The reference signal can be generated statically or dynamically. The static
reference signal is usually calculated from boost maps calibrated based on experiments [18],
while the dynamic reference is derived from some reference model that depicts the target
steering feel characteristics [14]. Nevertheless, neither method is able to produce ideal steering
feel under all operation conditions. On the one hand, the boost map represents only a steering

11
feel that is predetermined and cannot adapt to different operation conditions. On the other
hand, the artificial steering feel from a dynamic model is unrealistic to the driver due to the
discrepancy between the model and the real plant.

To address the aforementioned problem of feedback steering feel control, the authors
propose a torque overlay steering feel control algorithm based on a feedback control method.
In the proposed system, a reference model is used to generate a reference steering feel where
it uses a real-time rack force estimated from a rack force estimator and a tracking controller
is employed to drive the real system variable to the reference one by overlay torque.

The remainder of this chapter is organized as follows: firstly, the Column-Electric Power
Assist System (C-EPAS) plant model is presented in Section 2. Section 3 discusses the torque
overlay control system. In Section 4, the rack force estimation algorithm is elaborated, which
is needed in both reference model and tracking controller. The performance evaluation of the
control system is presented with simulation and hardware-in-the-loop testing in Section 5.
Finally, the conclusion is given in Section 6.

2.2 EPAS System Modeling

In this section, a detailed mechanical model of a C-EPAS system is built that includes
important dynamics affecting the steering feel, such as the worm gear, intermediate shaft,
pinion gear, and rack housing. This model is used as a plant model for validation of the
estimation and control algorithm in the later sections. The C-EPAS system detects driver
input torque from the torsion bar, and based on that the ECU sends a torque command
to a DC motor to generate assistance toque. The assistance torque along with the driver’s
torque is converted to a rack force via the intermediate shaft and the rack and pinion gear to
overcome the reactive force from the tie rods.

2.2.1 Mechanical Model

For modeling the C-EPAS, the steering wheel, motor armature, worm gear, intermediate
shaft, pinion shaft, and rack are considered. The schematic of the C-EPAS model is shown in
Fig. 2.2.

Equations (1) and (2) describe the dynamics of the steering wheel and the assist motor,
respectively.  
Tsw − Btb θ̇sw − θ̇g − Ktb (θsw − θg ) = Jsw θ̈sw , (2.1)

12
 Figure 2.2: Schematic of C-EPAS model

 
Tm − Btb θ̇m − θ̇w − Km (θm − θw ) = Jm θ̈m , (2.2)

where Tsw , θsw , and Jsw are steering wheel torque, angle, and inertia, respectively. Tm , θm ,
and Jm are motor rotor torque, angle, and inertia, respectively. Ktb and Btb are torsion
bar stiffness and damping, respectively. Km and Bm are motor rotor stiffness and damping,
respectively. θg and θw are gear and worm angle, respectively.

The dynamics of worm and worm gear are depicted by

 
Bm θ̇m − θ̇w + Km (θm − θw ) − Tbf,w + Fwx rw = Jw θ̈w , (2.3)

   
Btb θ̇sw − θ̇g + Ktb (θsw − θg ) − Bu1 θ̇g − θ̇u1,i + Ku1 (θg − θu1,i ) − Tbf,g + Fgy rg = Jg θ̈g ,
(2.4)
where Tbf,w and Tbf,g are motor shaft and gear shaft bearing friction, respectively. Fwx and
Fgy are worm and gear force, respectively. rw and rg are worm and gear radius, respectively.
Jw and Jg are worm and gear inertia, respectively.

Figure 2.3 illustrates the schematic of the intermediate shaft and u-joint. The equations
for the intermediate shaft are
Tu1,o − Tint = Ju1 θ̈u1,o , (2.5)

Tint − Tu2,i = Ju2 θ̈u2,i , (2.6)

13
 Figure 2.3: Schematic of intermediate shaft

h   i
Tu1,o = R−1 (θu1,i ) Bu1 θ̇g − θ̇u1,i + Ku1 (θg − θu1,i ) , (2.7)
h   i
Tu1,i = R (θu2,i ) Bu2 θ̇u2,o − θ̇p + Ku1 (θu2,o − θp ) , (2.8)
 
Tint = Bint θ̇u1,o − θ̇u2,i + Kint (θu1,o − θu2,i ) , (2.9)

where Tu1/2,i/o and θu1/2,i/o are upper/lower universal joint input/output torque and angle,
respectively. Tint is intermediate shaft torque. Ju1 and Ju2 are upper and lower intermediate
shaft inertia. Ku1 and Bu1 are stiffness and damping of worm gear shaft, respectively. Kint
and Bint are stiffness and damping of intermediate shaft, respectively.

The transmission ratio of the universal joint R is calculated byThe transmission ratio of
the universal joint R is calculated by

θ̇o cos β
= R (θi ) = , (2.10)
θ̇i 1 − sin2 β cos2 θi

where β is universal joint bending angle.

The following two equations represent the dynamics of pinion and rack, respectively.
 
Bu2 θ̇u2,o − θ̇p + Ku1 (θu2,o − θp ) − Tbf,p + Fpy rp = Jp θ̈p , (2.11)

Fpy − Ftl − Ff,r = mr ÿr , (2.12)

14
 Figure 2.4: Schematic of worm gear

where θp is pinion shaft angle. Ku2 and Bu2 are stiffness and damping of pinion shaft,
respectively. Tbf,p is pinion shaft bearing friction. Fpy is pinion gear force. rp is pinion gear
radius. Jp is pinion shaft inertia. Fry is rack gear force. FR is rack force from tie rod. Ff,r is
rack friction. mr is rack mass. yr is rack displacement.

As the key component of a C-EPAS system, modeling of the worm gear is of great
importance to obtaining a good C-EPAS model. In general applications, an intentional
backlash is usually provided in the worm gear mechanism to avoid teeth jamming and reduce
unnecessary frictional losses. However, considering the frequent change of transmission
direction of the worm gear in C-EPAS, a preload mechanism is adopted such that the rattling
noise caused by backlash can be reduced during gear inverse rotation. Such a structure
introduces additional mesh friction into the system whose value increases with increasing
load. In order to capture such a feature in a C-EPAS system, the worm gear contact force is
represented by a wedge-like model including worm gear compliance [19, 20]. The schematic
of the worm gear model is shown in Fig. 2.4.

The normal contact force between worm and worm gear in Eqns. (2.3) and (2.4) can be
calculated as a linear spring and dampers

N = kwg δ + cwg δ̇, (2.13)

where kwg and cwg are worm gear contact stiffness and damping, respectively. The deflection

15
 Figure 2.5: LuGre fiction characteristics

of gears can be calculated as

δ = θw rw sin γw − θg rg cos γw + δ0 , (2.14)

where δw is worm gear lead angle. δ0 is the initial deflection caused by worm gear preload F0 ,

F0
δ0 = . (2.15)
2kwg sin βw

The mesh friction of the worm gear is calculated as

Ff = µN, (2.16)

where the friction coefficient µ is modeled as a LuGre model [21]

µ = σ0 z + σ1 ż + σ2 v, (2.17)

|v|
ż = v − σ0 z, (2.18)
g (v)
2
g (v) = µc + (µba − µc ) e−( /vs ) , (2.19)
v

where σ0 , σ1 are bristle stiffness and damping, respectively. σ2 is viscous damping coefficient.
µc and µba are Coulomb and static friction, respectively. vs is Stribeck velocity. Fig. 2.5
shows the static characteristics of the LuGre friction model.

16
 Based on Eqns. (2.13)–(2.19), the contact force of the worm gear can be calculated as

Fwx = −N cos βw sin γw − Ff cos γw , (2.20)

Fgy = N cos βw cos γw − Ff sin γw , (2.21)

where βw is the pressure angle of the worm gear.

The contact force of the rack and pinion gear is also modeled in the same fashion. It
should be noted that the worm gear and the rack and pinion gear are modeled under the
assumption of single-tooth contact, while the worm gear normally has higher contact, meaning
that there are times when more than one tooth is in contact. The variation in the number of
teeth in contact might introduce position-dependent friction in the gears. It has been observed
that in transmission systems there exists position-dependent friction, which has oscillation
characteristics [22] For a C-EPAS system, the position-dependent friction originates from the
gear stiffness variation of the worm gear and the rack and pinion gear transmission.

The position-dependent friction of the C-EPAS system is modeled as a sine wave, shown
in below, whose amplitude and frequency is dependent on the external load and the gear
ratio.
τf (θ) = A sin (ωθ + φ) . (2.22)

In addition to the mesh friction of the worm gear and the rack and pinion, the friction of a
C-EPAS system also originates from the bearings of the worm gear, pinion shaft, and rack
housing. These frictions are also modeled by a LuGre model assumed to be independent of
load and position.

The overall friction level is examined by simulation test. Figure 2.6 shows the comparison
between simulation and test results. Figure 2.6(a) is the comparison result when the pinion
and rack are detached from the steering column, and Fig. 2.6(b) is the result from the
complete C-EPAS system. The simulation is carried out under constant steering wheel angle
input at different rates without load. It can be seen that the friction level of our model
matches pretty well to the test results.

2.2.2 Basic Control Strategy

The conventional power assistance functions have been embedded in the C-EPAS by the
supplier to realize basic steering feel control. Figure 2.7 shows a block diagram of the basic
steering control strategy used in this study. There are three basic control systems: power

17
(a) Steering column only (b) Whole steering system

Figure 2.6: Friction level validation

Figure 2.7: Block diagram of C-EPAS assist control strategy

18
 Figure 2.8: Boost curve for power assistance

Figure 2.9: Schematic of Feed-forward steering feel control strategy

assistance, active damping, and active return control [12]. The role of power assistance is to
reduce the steering effort, the active damping is to improve the steering stability, and the
active return function helps the steering wheel return to center when a driver releases the
steering wheel. The power assistance torque is calculated according to the boost curve shown
in [23], and active return control is released by a PI controller. Figure 2.8 shows a example
of boost curve for power assistance.

The basic power assistance is implemented as feed-forward torque injected to the C-EPAS
system, as shown in Fig. 2.9. In Fig. 2.9, the motor torque is added to the EPAS system
in feed-forward fashion calculated based on lookup tables. Although this method is easy to
implement, it requires tedious parameters tuning on the parameters of feed-forward functions
and lookup table. Another drawback of such a method is the lack of robustness against
parameter variation, disturbance, etc.

19
 Figure 2.10: Overall schematic of C-EPAS steering feel control

2.3 Torque Overlay Control System

In this section, we introduce the EPAS steering feel control algorithm based on torque overlay.
The torque overlay control is built upon the existing conventional power assistance algorithms.
Figure 2.10 shows a schematic of the C-EPAS control system. This system consists of a
reference model, tracking controller, and rack force estimator. The estimated real-time rack
force from the rack force estimator is used to the reference model to set a reference pinion
angle and angle rate. Then, the actual pinion angle is regulated by the tracking controller to
follow the reference value. The following sections discuss in detail for individual component.

Although steering feel control is our focus in this chapter, we want to stress that the
control structure in Fig. 2.10 can be applied to steering position control for autonomous
vehicle application by simply disable the reference model, and replace the source of reference
signal δref to the steering angle command from high-level motion control system, e.g. the
NMPC controller in Chapter 4. Moreover, the rack force estimation and friction estimation
developed later in Section 3 can be used as feed-forward compensations to improve the control
performance.

2.3.1 Reference Model

The reference model provides a target steering feel that the tracking controller must follow
with overlay torque. The target steering feel can be generated by various approaches such as
lookup table [18], spring and damper [24], and dynamic model [14].

20
 Figure 2.11: Schematic of reference model with C-EPAS and basic control systems

In our study, a joint reference model is proposed. This joint model combines basic
assistance functions with a reduced-order EPAS mechanical model. The mechanical model of
EPAS provides an intuitive basis for tuning the steering feel with physically-related parameters
such as friction and damping coefficients, and the basic assistance module takes advantage
of the conventional feed-forward steering feel control method so that the resulting reference
steering feel is familiar to the driver. Figure 2.11 shows the schematic of the reference model.

The dynamics of the reference model is given as

Tt − Tf,ref − TR,ref + Ta = Iref δ̈ref . (2.23)

In the above reference model, the steering system is considered as one stiff bar with assuming
that the inertias and frictions at different location are lumped together on the column
compared to the high order EPAS model. The inputs of the reference model are torsion bar
torque Tt , friction torqueTf,ref , rack load torque TR,ref , and assist torque Ta . Using these
inputs, the reference steering column angle δref is computed from Eq. (2.23). With torsion
bar sensor, the torsion bar torque can be measured and the assist torque is also accessible
from the basic assist control system. Since there is no sensor to measure directly the rack
reaction torque in production vehicles, a rack force emulation based on a rack force model
and vehicle states has been widely used [25]. Although this approach is straightforward, the
major downside is that realistic rack force is difficult to obtain due to modeling error. In
addition, the emulated rack force is unable to capture the change of rack force caused by

21
Figure 2.12: Comparison of steering feel hysteresis with reference and plant EPAS model

real-time driving condition changes, e.g., road friction variation. Therefore, the driver will
perceive a misleading steering feel since the actual rack force is masked by the rack force
model.

In the proposed control system, the actual rack force is estimated from the rack force
estimator unlike other emulation methods. In order to provide a realistic steering feel to
drivers, it is necessary that the estimated rack force should be closely matched with the
actual rack force. The rack force estimator will be discussed in Section 3.

As discussed earlier, the friction level in the steering system significantly affects the
steering feel. Since the target steering feel is based on the reference model, a reduced friction
level is used to the reference model compared to the plant model, i.e., Tf,ref is chosen to be
smaller than the overall friction of the plant, such that the target steering feel will have a
small hysteresis feel. Figure 2.12 compares the steering feel hysteresis between the reference
model and the plant model. As seen clearly from Fig. 2.12, it is possible to set different
target steering feels with different friction levels in the steering system. With nominal friction
in the reference model, the target steering feel is similar to that of the plant model. The
steering feel of the reference model with reduced friction has a narrower hysteresis compared
to the plant model, producing a lighter steering feel to the driver [23].

22
2.3.2 Tracking Controller

The role of the tracking controller is to regulate the C-EPAS to track the reference by applying
overlay torque. It is important for a tracking controller to have good robust performance
under disturbance, road friction variation, steering system friction variation, etc.

The reduced-order C-EPAS model for controller design is written as

Tt + Tm iwg + FR /ip − fe = Ie δ̈c . (2.24)

The reduced-order model lumps the inertia of the worm gear, worm and motor shaft,
intermediate shaft, pinion shaft, and rack into a single equivalent one, and the frictions in
several locations in the high-order model are also represented by an equivalent frictional
torque.

The equivalent inertia and friction in Eq. (2.24) are given as

Ie = Is + (Iw + Im )i2wg + Iuj + Ilj + Ip + mr /i2p , (2.25)

fe = fgb + (fwb + µw Nwp )iwg + µg Nwp + fpb + µr Npp , (2.26)

where µw = µsinwg rw
βw
, µg = µwgsin
rw tan γw
βw
µrp sin γp rp
, µp = sin βrp cos γr
, µr = µR − µrp tan γr
sin βrp
are, respectively,
the equivalent friction coefficients of worm, worm gear, pinion and rack on the column, due
to the preload. µwg and µrp are the mesh friction coefficient of worm gear and pinion gear,
respectively. Nwp and Npp are the preload of worm gear and pinion gear, respectively.

The sliding mode control (SMC) technique is widely used to control systems with large
uncertainty. For the tracking controller, a sliding mode controller is developed for this study.
The sliding variables used for our controller are given as

s = λ1 e + λ2 ė, (2.27)

e = δc − δref , (2.28)

where λ1 and λ2 are positive gains of tracking error and error rate, respectively, and δref is
reference pinion angle.

The equivalent control is obtained by letting ṡ(t) = 0, which is the necessary condition

23
for the tracking error to remain on the sliding surface. Subtracting (2.23) from (2.24) gives

Iref (δ̈c − δ̈ref ) + I∆ δ̈c = To − (TR − TbR ) − (fe − fref ), (2.29)

where I∆ is the inertia difference between reference model and plant, Ta is the overlay torque,
and reference rack force is the estimated rack force TbR from the Kalman filter.

Taking the derivative of (2.27) and substituting (2.29) becomes

 
1
ṡ = λ1 ė + λ2 (To − ∆TR − ∆f − I∆ δ̈c ) , (2.30)
Iref

where ∆TR = TR − TbR , and ∆f = fe − fref .

The equivalent control is derived by equaling (2.30) to 0,

λ1 Iref
To,eq = ∆TR + ∆f + I∆ δ̈c − ė. (2.31)
λ2

In Eq. (2.31), ∆TR and I∆ δ̈c are unknown, and ∆f can only be approximated based on the
reference and the nominal plant model, while its actual value will vary over time due to wear.
Therefore, the sliding mode cannot be obtained by equivalent control alone. The applicable
equivalent control is
λ1 Iref
To,eq = ∆fb − ė, (2.32)
λ2
where ∆fb is the nominal friction difference.

In order to compensate for the deviation of the representative point due to parameter
variations and disturbances, a switching controller must be designed to drive the output
trajectory to move toward and reach the sliding surface. For this purpose, the Lyapunov
function can be chosen as
1
V (t) = s2 (t), (2.33)
2
with V (0) = 0 andV (t) > 0 for s(t) ̸= 0. A sufficient condition to guarantee that the
trajectory of the error will approach the sliding phase is to select the control strategy, also
known as the reaching condition:

V̇ (t) = sṡ < −η|s|. (2.34)

24
 To satisfy the reaching condition, the equivalent control To,eq given in (2.32) is augmented
by a switching control term, To,N . The sliding mode controller is designed as

To = To,eq + To,N . (2.35)

If Eqns. (2.30), (2.32), and (2.35) are inserted into the reaching condition (2.34), it is obtained
that
λ2 h i
s To,N − ∆TR − (∆f − ∆fb) − I∆ δ̈c < −η|s|. (2.36)
Iref
The switching control To,N in (2.36) is chosen as

To,N = −Ksgn(s), (2.37)

where K is the a positive constant and means the upper bound of uncertainty, and sgn(·)
denotes signum function, defined as

1 if s > 0



sgn(s) = 0 if s = 0 . (2.38)


−1 if s < 0


Substituting (2.37) back to (2.36), one has

Iref
K > |∆TR | + |∆f − ∆fb| + |I∆ δ̈c | + η. (2.39)
λ2

This implies that V̇ (t) is a negative definite function as long as K is greater than the sum of
the rack force estimation error, the friction torque variation, and the inertia force difference.
Therefore, the sliding surface s is attractive for this system. The convergence rate can be
adjusted by η.

To alleviate the chattering issue of the sliding mode control, a saturation function has
been used via introduction of a thin boundary layer around the sliding surface. The switching
control To,N is modified as

s −Ksgn(s) if |s| ≥ Φ
K
b o,N = −Ksat( ) = , (2.40)
Φ −K s otherwise
Φ

where Φ is a positive constant that defines the thickness of the boundary layer affecting the

25
steady-state accuracy and robustness.

2.4 Rack Force and Load-Dependent Friction Estimation

The rack force estimator is needed to provide real-time rack force such that the target steering
feel reflects the real driving conditions. It is shown in [26] that the friction of the steering
system affects the performance of the rack force estimate significantly. A friction estimation
algorithm was proposed based on the relationship between steering torque and angle [27].
The main concept of the friction estimation algorithm is summarized here to make the paper
self-contained.

2.4.1 Disturbance Observer based Rack Force Estimator

Based on the low-order model, the rack force can be estimated by a disturbance observer.
The state space presentation of the C-EPAS system is given in the following:
" # " # " # " #
δ̇c δc Tt fe
=A + Bu + Bw , (2.41)
δ̈c δ̇c Tm FR
" # " #
δc δc
=C , (2.42)
δ̇c δ̇c
" # " # " # " #
0 1 0 0 0 0 1 0
where A = , Bu = , Bw = , and C = .
0 0 1/Ie iwg /Ie −1/Ie 1/ip Ie 0 1

There are two unknown inputs in Eq. (2.41), fe and FR . Augmenting the state vector
with both fe and FR by assuming that they are slowly changing, i.e., f˙e = ḞR = 0, we have
the extended state space equations
   
δ̇c δc " #
   
 δ̈c 
  = AE   + BR Tt ,
 δ̇c 
 f˙  f  (2.43)
 e  e Tm
ḞR FR

 
" # δ c
 
δc  δ̇c 
= CE 
f ,
 (2.44)
δ̇c  e
FR

26
 " # " #
A Bw Bu h i
where AE = , BE = , and CE = C 0 .
0 0 0

The observability of system (2.43) and (2.44) is checked by the rank of matrix Ob =
h iT
CE CE AE CE AE CE AE . It can be shown that the rank of Ob = 3, which is smaller
2 3

than the system order, indicating that the extended system is unobservable. It is thus impos-
sible to estimate all the system states no matter what observer is used. The unobservability
of the system can also be interpreted in the physical aspect: since the steering column is
modeled as a single rigid shaft, the rack force FR and steering friction fe have the same effect
on the dynamics of the column. Without including additional sensors or dynamics, they are
indistinguishable from each other.

In order to estimate FR based on steering system dynamics, fe must be known a priori

or determined by another mechanism. For the moment, we just assume that fe is a known
input to the system. The state space presentation of the C-EPAS system is discretized as

xk+1 = F xk + Guk , (2.45)

yk = Hxk , (2.46)
h iT h iT h iT
where x = δc δ̇c FR , u = Tt Tm fe , and y = δc δ̇c . The matrices F , G, and
H are derived using the zero-order-hold method.

The system represented by Eqns. (2.45) and (2.46) is fully observable so that the rack
force FR can be estimated using any linear observer [28], e.g., a Luenberger observer. In this
work, we adopt a Kalman filter to optimally estimate the rack force while attenuating the
noise effect. The Kalman filter is given by

Pk− = F Pk−1
+
F T + Q, (2.47)

Kk = Pk− H T (HPk− H T + R)T , (2.48)

x− +
k = F xk−1 + Guk−1 , (2.49)

x+ − −
k = xk + Kk (yk − Hxk ), (2.50)

Pk+ = (I − Kk H)Pk− (I − Kk H)T + Kk RKkT , (2.51)

where Q and R are the covariance matrices of system noise and measurement noise, respec-
tively.

27
 Figure 2.13: Rack force estimation without friction estimation

Figure 2.14: C-EPAS system friction vs. assist motor torque

The effect of steering system friction on steering feel can be seen from Fig. 2.13. It
shows the rack force estimation result when the steering system friction in the estimator is
approximated by Coulomb friction while the real friction in the plant is not. The estimation
shows great error near the peak of the rack force due to the discrepancy in the frictions.

2.4.2 Load-Dependent Friction

Figure 2.14 shows the overall friction of an C-EPAS system with respect to assist-motor
torque based on low-frequency sinusoidal steering test data from an C-EPAS test bench [29]
As shown in Fig. 2.14, the friction torque is quite different from the commonly used Coulomb
friction, as it varies almost linearly with motor torque during steering (A and C) and return
maneuvers (B and D).

28
 The model of load-dependent friction fei is expressed as

f
0 |Tm | < T0,i
fei = . (2.52)
f + λ (|T | − T ) otherwise
0 f,i m 0,i

where i = s, r represents the steering and return maneuvers, respectively. f0 is the load-
independent friction resulting from preload, λf,i is the load dependency coefficient, and T0,i
is the motor torque where load dependency starts to appear.

The algorithm is based on the relationship between the steering angle and torque that
are measured by torsion bar sensors. In the case of low-frequency steering input, e.g. 0.2Hz
sinusoidal input, the inertia force in Eq. (2.24) is negligible. Thus, the torque balance
equation of EPAS can be expressed as

Tt + Tm iwg = Tapp = −FR /ip + fe , (2.53)

where Tapp denotes the total applied torque from the measured driver and the assist motor.

When the vehicle is operated in the linear range at a constant forward velocity with
low-frequency input, the rack forces of steering and return at the same steering angle are
similar, i.e., FRs (δcr ) ∼ FRr (δcr ) when δcs = δcr . Therefore, by subtracting the steering and return
states, we have
s
|Tapp (δcs ) − Tapp
r
(δcr )| = |fes | + |fer |. (2.54)

Equation (2.54) indicates that the difference between the total applied torque of the steering
and return states at the same steering angle is equal to the sum of frictions of both states
at that angle. Figure 2.15 depicts this relationship, where the input is the low-frequency
sinusoidal steering angle. Figure 2.15 shows the hysteresis loop of the total applied torque
versus the steering angle from simulation. According to the relationship in Eq. (2.54), the
width of hysteresis indicates the sum of friction of the steering and return states. The torque-
versus-angle relationship in Fig. 2.15 can be divided into two sections in each steering and
return state. The hysteresis in the small steering angle remains constant, which corresponds
to the load-independent portion of the friction. As the steering angle keeps increasing, the
load-dependent effect appears after the change point δb . As a result, the hysteresis becomes
greater. In addition, the relationship between Tapp and δc can be approximated as the two
affine functions A1 and A2 in each section, respectively.

29
 Figure 2.15: Relationship between applied torque and angle

The relationship between the hysteresis width and friction torque is given as

s
Tapp (δcs ) − Tapp
r
(δcr ) = 2f0 + fL,s + fL,r , (2.55)

where fL,s and fL,r are the load-dependent portion of the friction in the steering and return
states, respectively. To estimate f0 , fL,s , and fL,r , the measurements of Tt , Tm , δc , and Tapp
are filtered and stored (denoted by superscript *). Then, the affine functions A1 and A2 are
identified using least square regressions by changing the break point δb∗ from δc,min
∗
to δc,max
∗

iteratively. After all iterations are completed, the optimal parameters (λf , f0 , and δb∗ ) with
the smallest regression error are chosen.

Next, the load-dependent portion of friction fL,s and fL,r can be determined by the
relationship between Tapp
∗
and δc∗ using optimal parameters identified in the previous step:

∗
fL,s ∗
= |Tapp − (ρf δc∗ + f0 sign(δ̇c∗ ))| if δc∗ δ̇c∗ > 0, (2.56)

∗
fL,r ∗
= |Tapp − (ρf δc∗ + f0 sign(δ̇c∗ ))| if δc∗ δ̇c∗ < 0, (2.57)

where ρf and f0 are the identified slope and offset of A1 .

Finally, the coefficients λf,i and T0,i can be obtained by the least squares method, using
the equation
∗
fL,i = λf,i (|Tm∗ | − T0,i ), i = s, r. (2.58)

30
 Figure 2.16: Schematic of rack force estimation with friction estimation

For the proposed estimation algorithm to have good performance, the vehicle should be
running in a low-frequency on-center steering maneuver at constant speed, such that the rack
forces in the steering and return states can be canceled. Figure 2.16 shows the schematic of
rack force estimation with friction estimation.

2.5 Simulation and Experimental Results

The proposed rack force estimation and steering feel control algorithm are evaluated using
simulations and experimental testings.

2.5.1 Rack Force Estimation Simulation

In this case, the simulated vehicle is running on a high µ (µ = 1) road surface at 75km/h while
the steering excitation of a 0.2Hz sinusoid with amplitude of 30◦ is applied. The response
of total applied torque vs. steering angle during 10s simulations is plotted in Fig. 2.17. As
discussed in the previous section, the load-dependency characteristic of friction is reflected in
the hysteresis width of the curve. Based on the raw measurements, the data points indicated
by circular marks are obtained, which will be used for friction estimation in the following.

Figure 2.17 also shows the lines of two linear regressions, which represent the load-
independent and load-dependent portions, respectively. The slopes change of the two portions
is clearly seen. The change point of the load-dependent and independent parts is found by
the iteration regression algorithm. Figure 2.18 shows the regression errors of the iteration
procedure, where the lowest point corresponds to the change point.

The load-dependent friction can then be estimated by the difference between the regression
line and the measurement data in Fig. 2.17 . The load-dependent friction is plotted with

31
Figure 2.17: Total applied torque vs. steering angle

Figure 2.18: Iterative regression results

32
 Figure 2.19: Load-dependent friction vs. motor torque

respect to the motor torque in Fig. 2.19. It can be seen clearly in Fig. 2.19 that the
relationship between friction and motor torque is almost linear. The different dependency
on motor load in the return and steering states is also captured. By linear regression, the
load-dependent friction model, which is later used for the rack force estimation, is identified
by two straight lines.

After the load-dependent friction is identified, the estimated friction parameters are
applied to the friction model used for rack force estimation. Root mean square error (RMSE):
v
u n F − Fb 2
uP  
t 1 R R
RM SE = , (2.59)
n

and normalized root mean square error (NRMSE) over the range of the measurement data:

RM SE
N RM SE = (2.60)
FR,max − FR,min

are introduced as quantitative error analysis.

Figure 2.20 and 2.21 show the friction and rack force estimation results, respectively. As
can be seen in Fig. 2.20, the estimated friction successfully captures the load-dependency
characteristic existing in the plant model. As a result, the rack force estimate with load-
dependent friction (RMSE=84N and NRMSE=2.0%) shown in Fig. 2.21 is much improved
compared to that with load-independent friction only (RMSE=511N and NRMSE=11.9%).

33
 Figure 2.20: Friction estimation

Figure 2.21: Rack force estimation

34
 (a) Vehicle trajectory of double lane change

(b) Steering angle input of double lane change

Figure 2.22: Vehicle trajectory and steering angle of DLC maneuver

The steering inputs are standard sinusoidal waves in the previous cases. In this section,
the vehicle is controlled by the Carsim driver model, such that the rack force estimation
is evaluated in more realistic scenarios. The vehicle performs a double lane change (DLC)
maneuver at 120km/h with µ=0.85. The load-dependent friction is assumed to be identified
using sinusoidal input. Figure 2.22 present the vehicle trajectory and steering wheel angle
during the simulation.

The friction and rack force estimation results are shown in Fig. 2.23, respectively. As
is shown, a good friction estimate is achieved, although estimation is slightly worse in the
return state compared to the steering state. A good rack force estimation is also obtained
with small RMSE=104N and NRMSE=2.6% in the presence of the saturation of rack force.

2.5.2 Steering Feel Control Simulation

The proposed steering feel control algorithm is evaluated through simulation in Carsim
and Matlab/Simulink environments. At first, the steering feel under low-frequency input is
simulated. For low-frequency steering feel simulation, a 0.2Hz, 30◦ sinusoidal steering wheel
angle input is applied while vehicle runs at 75km/h. The steering feel control results are
shown in Fig. 2.24. We can see that the steering feel without torque overlay control has a
larger hysteresis compared to the reference. On the other hand, with activation of torque
overlay control, the steering feel hysteresis matches the reference one.

35
 (a) Friction estimation of double lane change

(b) Rack force estimation of double lane change

Figure 2.23: Friction and rack force estimation of DLC maneuver

Figure 2.24: Comparison of steering feel hysteresis with and without torque overlay control

36
 Figure 2.25: Comparison of tracking controller performance

Figure 2.25 compares tracking performance of the tracking controller. It can be observed
that it tracks smoothly with the reference signal without any chattering and its tracking
error is less than 2%. The overlay torque used is also smooth, with the maximum value
below 0.2Nm. These simulation results show the effectiveness of the torque overlay control
algorithm on steering feel control.

The sine-sweep steering feel frequency response is shown in Fig. 2.26. Sine-sweep steering
wheel angle inputs (0.2 to 3Hz) with 15◦ amplitude are applied at a 120km/h vehicle speed.
As shown in Fig. 2.26, the steering feel with torque overlay control is well matched to the
reference one.

2.5.3 Hardware-in-the-Loop Test

In order to validate the proposed algorithm, Hardware-in-the-loop (HiL) simulations were

performed. Real measurement data from the torsion bar and motor sensors have been
gathered on an EPAS HiL simulator. Moreover, a rack force sensor is installed on the HiL
system in order to compare the estimated value and the real torque applied on the rack. The
hardware configuration of the HiL system is shown in Fig. 2.27. The data are recorded under
a 0.2Hz sinusoidal steering angle input with 30◦ amplitude. The vehicle speed is 75km/h.

Figure 2.28 shows the rack force estimation results. It is shown in Fig. 2.28 that the rack
force estimate is good. The amplitude of the estimate error is less than 200N. Figure 2.29
shows the responses of torque overlay control system from the HIL test with a 0.2Hz sinusoidal
steering input. The HIL system consists of C-EPAS hardware, dSpace simulator, rack load

37
Figure 2.26: Sine-Sweep steering feel frequency response

Figure 2.27: EPAS hardware-in-the-loop simulator

38
 Figure 2.28: Rack force estimation of HIL test

Figure 2.29: HIL test results of torque overlay control

39
 Figure 2.30: Torque-to-rotate test results

actuators, and steering robot. As can be seen in Fig. 2.29, the steering feel hysteresis response
is improved (a smaller steering feel hysteresis) with the proposed torque overlay control.

To evaluate the robustness of the control system’s performance, the actual friction on
the pinion shaft is adjusted to change the steering system friction level. Figure 2.30 shows
the steering torque vs. steering angle from the torque-to-rotate test. In Fig. 2.30, the one
with the higher torque-to-rotate corresponds to the system with increased pinion friction. It
can be seen that the friction of the steering system is increased by about 1Nm compared to
the nominal system.

Figure 2.31 shows the rack force estimation results in the case of increased pinion friction
under low-frequency sinusoidal input. We can see from Fig. 2.31 that the estimation of rack
force maintains a satisfactory performance. The error of the estimation is generally below
300N. The steering feel of the system with increased friction is compared with the nominal
system in Fig. 2.32. We first compare the steering feels of increased and nominal friction
without torque overlay control. The steering feel of the system with increased friction has a
bigger hysteresis than that of the nominal system. However, the steering feel with torque
overlay control is almost the same in both cases, which means that the torque overlay control
system performs consistently well in the presence of friction variation.

40
 Figure 2.31: Rack force estimation with increased pinion friction

Figure 2.32: Steering feel of steering system with increased and nominal friction

41
2.6 Conclusions

In this chapter, we proposes a steering feel improvement methodology utilizing overlay torque
and feedback control. The simulation and test results demonstrate that (1) the reference
model can provide a more realistic target steering feel due to the inclusion of real-time rack
force, (2) rack force can be successfully estimated by the Kalman filter integrated with friction
estimation, and (3) steering feel can be improved by tracking the target steering feel despite
variations in vehicle parameters and operation conditions.

42
 Chapter 3

Clamping Force Control for Electromechanical Brake System

Along with the steering system, the brake system is another important chassis component that
significantly affects vehicle’s performance and safety. In fact, many of the widely implemented
active safety systems, for example ABS and ESC, are built on the basis of brake system. In
particular, with growing interest of vehicle electrification and autonomous vehicle systems,
electromechanical brake (EMB) system becomes important and has widely gained attention.

In this chapter, we design a clamping force control system for EMB. EMB generates the
clamping force between the brake pad and disk through an electric motor and mechanical
transmission. Control of the clamping force has significant impact on the brake performance.
The EMB control shares many similarities to the EPAS control in previous chapter, for
example, mechanical friction being a major disturbance and estimation needed as feedback to
the controller. But it also has its uniqueness in terms of system modeling, control algorithm
and architecture. We propose a unified clamping force control framework that incooperates
estimation, contact detection, gap closing and tracking control.

3.1 Introduction

Inspired by the fly-by-wire system in the aerospace industry, drive-by-wire technology has
been getting increasing attention from the automotive industry in recent years. Regarding
the brake-by-wire (BBW) technology, the first implementation is the electro-hydraulic brake
(EHB) system. It retains the hydraulic linkages of the conventional brake system while it
realizes by-wire functionality. The electro-mechanical brake (EMB) system is introduced as
the second generation of BBW, which further improves performance. Compared to the EHB,
it replaces the hydraulic components with an electric motor and mechanical transmissions
that connect the actuator and brake pad.

The EMB system offers various advantages, such as the reduction of volume and weight

43
and ease of maintenance. More importantly, the fast response of the electric motor and the
computational power of the electronic control unit (ECU) facilitate the implementation of
advanced driver assistance systems (ADAS) such as automated emergency braking (AEB)
and electronic stability control (ESC) systems.

One major research topic of the EMB system is the clamping force control. The brake
torque is related to the clamping force through the friction between the brake pad and
disk. Thus, the quality of the clamping force control directly affects the brake performance.
The challenges of EMB clamping force control include the friction of the bearing and gear
transmission, as well as the nonlinear stiffness and hysteresis of the caliper and brake pad.
Furthermore, the difficulty of measuring real-time clamping force as feedback also must be
taken into consideration during controller design and implementation.

Many control architectures have been proposed for clamping force control. A popular
control method is the cascade PID control [30–32], which consists of three feedback controllers
positioned sequentially from the inner to the outer loop for regulating the current, speed,
and force, respectively. This control structure is conceptually simple and intuitive. However,
the effort for tuning the controllers is extensive, since three loops are coupled together and
must be tuned iteratively to achieve the best overall performance. To cope with the nonlinear
friction in the EMB system, a feedforward compensation of Coulomb friction is added to the
cascade control system [33]. It is shown that the tracking error is reduced, and a finer force
modulation is achieved. Similarly, an inverse model controller is developed in [34], where
various types of friction and nonlinearities of the individual components are modeled and
compensated for using an inverse model. Although the feedforward control can linearize
the EMB plant to improve the effectiveness of the feedback controller, its performance is
very sensitive to modeling error. Even slight model uncertainties will yield a significant
performance degradation or even instability.

To improve the robustness of the control system, more advanced control techniques have
been introduced. An H∞ controller for clamping force tracking is designed in [35], where
the parametric uncertainty and unmodeled dynamics of EMB are considered as structured
uncertainty. The simulation shows that the H∞ controller is less sensitive to the actuator
perturbations yet too conservative compared to the conventional PI controller, resulting in
a sluggish response. In order to obtain the ideal tracking performance in the presence of
uncertain parameters, an adaptive sliding mode controller is developed in [36], in which the
parameter of a lumped friction model is identified online and incorporated into a sliding

44
model controller such that the tracking performance is robust against varying friction. While
it presents a promising result, only friction variation is considered, and the adaptation
mechanism might fail when there are multiple fast-varying parameters in the plant. Other
control algorithms such as model predictive control [37, 38] and time-optimal control [39] are
also studied. Although these methods give optimal controllers in theory, they rely heavily on
accurate modeling of the plant and are not quite suitable for fast low-level control applications
due to the more demanding computational effort.

Aside from the tracking control, there are also practical problems to be considered. First
of all, the clamping force is usually not measured due to the difficulty of sensor installation
and cost. In [32], a stiffness model that describes the relationship between motor angle and
clamping force is proposed for clamping force estimation. A similar approach has also been
adopted in [36] because of its simplicity. However, the hysteresis of the clamping force is not
considered here, which could result in a large error if the hysteresis is significant in the real
system. In [40], the clamping force is divided into linear and nonlinear part, respectively.
Then a Kalman filter is designed to estimate the clamping force as well as other states. A
data-driven approach for clamping force estimation is proposed in [41] that relies only on
the motor current and voltage. However, this method requires a large amount of data from
various experiments. It does not consider clamping force hysteresis, either. The clamping
force models in [42] and [43] consider both the nonlinearity and the hysteresis, which requires
identification of extra parameters using experimental data. Another practical problem is
contact detection and gap management. An air gap must be maintained between the brake
pad and disk to avoid damage caused by unwanted contact. Furthermore, the air gap distance
must be kept constant for the best control performance despite the wear of the brake pad.
Several algorithms have been developed for this problem. For example, the gradient of motor
torque is used in [43] to detect the contact point. However, the accuracy of contact detection
will be affected by measurement noise.

In this chapter, a clamping force control system is developed that consists of the clamping
force estimator, gap distance estimator, and tracking controller. The rest of the chapter is
organized as follows. The mechanical and clamping force models of the EMB are introduced in
Section 3.2. Clamping force estimation, gap distance estimation, and a tracking controller are
developed respectively in Section 3.3. In Section 3.4, closed-loop simulations of the proposed
clamping force control system are performed, and the control performance is discussed.
Finally, conclusions are drawn in Section 3.5.

45
 Figure 3.1: EMB schematic

3.2 EMB Modeling

EMB modeling is introduced in this section. Figure 3.1 shows the schematic of the considered
EMB system. The rotational motion of the motor is converted into a linear motion of the
head through the reduction gear and the screw thread gear. The linear motion of the head
generates a clamping force between the pad and the disk, which provides the braking torque.

3.2.1 Mechanical Model

The permanent-magnet DC motor is modeled as an ideal DC model using the following two
equations:
Va = Ra Ia + Kemf θ̇m + La I˙a , (3.1)

Jm θ̈m = Kmotor Ia − Bm θ̇m − τf − TL , (3.2)

where Va is the armature voltage, Ra is the armature resistance, La is the armature inductance,
Kemf is the back-emf coefficient, Jm is the motor rotor inertia, Kmotor is the torque constant,
Bm is the rotor damping coefficient, θ̇m is the rotor angular speed, Ia is the armature current,
τf is the friction torque, and TL is the load torque.

In Eq. (3.2), the load torque TL is expressed as the torsional torque resulting from the
compliance and damping of the motor shaft and gear train
 
TL = km (θm − ig θs ) + cm θ̇m − ig θ̇s , (3.3)

46
 Figure 3.2: Schematic of ball screw

where km , and cm are respectively the stiffness and damping of the motor shaft and gear
train, ig is the reduction gear ratio, and θs is the rotational angle of the ball screw.

The reduction gear is modeled by the ideal reduction ratio, including transmission
efficiency, which is written as
Ts = ig ηg TL , (3.4)

where ηg is the gear efficiency and Ts is the drive torque on the ball screw.

The rotary motion of the ball screw is converted to the linear translation of the ball screw
nut. The unwrapped planar model is used for modeling the ball screw, with the assumptions
that the screw ball is rigid and the motion of screw and nut are constrained along the xs and
yn axes, respectively. Figure 3.2 illustrates the schematic of the ball screw model.

The dynamics of the ball screw and nut are expressed as following two equations,
respectively
meq ẍs = Fs − fb − Ns sin α − fs cos α, (3.5)

(mn + mp ) ÿn = Ns cos α − fs sin α − Fp − fn , (3.6)

where meq is the equivalent translational mass of the screw, α is the screw lead angle, mn
is the nut mass, mp is the lumped mass of the brake pad and head, and xs and yn are the
translation displacement of the ball screw and the nut along their axis, respectively. The
external forces of the ball screw are the driving force from reduction gear Fs , tapered bearing
friction fb , screw and nut contact force Ns , and screw ball rolling friction fs . The external
forces applied on the screw nut are the clamping force Fp and the housing friction fn .

47
 The torque and force of the ball screw can be converted by the following equations

Ts
Fs = , (3.7)
rs

Tf b
fb = , (3.8)
rs
where rs is the pitch radius of the ball screw and Tf b is the bearing friction torque calculated
by the LuGre friction model.

The screw ball rolling friction fs can be calculated as

f s = µ s Ns , (3.9)

where µs is the friction coefficient modeled by the LuGre friction model.

In Eq. (3.5), the equivalent translational mass meq is the lumped mass, including the
planetary gear, the spur gear, and the screw shaft, which are expressed by the following
equations
Jeq = Jpg i2g + Jsg + Js , (3.10)
Jeq
meq = , (3.11)
rs2
where Jpg is the planetary gear inertia seen from the motor side, Jsg is the spur gear inertia,
and Js is the ball screw inertia.

In the unwrapped model, the rotational angle θs of the screw shaft is expressed by the
equivalent translational displacement of xs using the equation below

xs = θs rs . (3.12)

The relationship of the screw shaft and nut displacement is expressed as

xs tan α = yn . (3.13)

Using Eqns. (3.6)–(3.13), we can combine (3.6) and (3.13) by eliminating the contact

48
force Ns , which gives
TL − Tf b − rs GFp
θ̈s = , (3.14)
Jeq + rs2 G (mn + mp ) tan α
ÿn = θs rs tan α, (3.15)

where G = tan α+µs

1−µs tan α
.

The dynamic LuGre friction model is adopted in our model, which is written as

Tf = σ0 z + σ1 ż + σ2 θ̇, (3.16)

θ̇
ż = θ̇ − σ0   z, (3.17)
g θ̇
  2
g θ̇ = Tc + (Tba − Tc ) e−( /θ̇s ) , (3.18)
θ̇

where σ0 and σ1 are the bristle stiffness and damping, respectively, σ2 is the viscous friction
coefficient, Tc and Tba are the Coulomb and static friction torque, respectively, θ̇s is the
Stribeck velocity; and θ̇ is the angular speed.

3.2.2 Clamping Force Model

A correct understanding of the clamping force characteristics is necessary before undertaking

the modeling. Experimental testing of the EMB hardware was conducted by our research
colleagues at DGIST to obtain the real clamping force. For the detailed experimental setup,
one can refer to [44].

Figure 3.3 shows the clamping force experimental test results. We can see that there
is a dead zone where clamping force remains zero before it is generated. This dead zone
corresponds to the initial air gap between the brake pad and disk, which is handled by the
gap management system. The clamping force is generated as soon as the brake pad contacts
the disk. Two main aspects of the clamping force are clearly shown, the nonlinearity and the
hysteresis, which are determined by the properties of the brake caliper and brake pad [45].

It has been shown that the nonlinearity can be approximated nicely by polynomials as
functions of motor angle [31, 46]. Figure 3.4 shows various clamping force from measurements.
From Fig. 3.4(a), we can see the polynomial model can be used to fit the test data pretty well.
However, the hysteresis cannot be presented by the memoryless polynomial model. There
are a few algorithms that try to incorporate the hysteresis into the clamping force model

49
 (a) Motor input profile (b) Clamping force

Figure 3.3: Clamping force of EMB

(a) Test data compared with polynomial stiffness (b) Clamping force of various reverse points
model

Figure 3.4: Clamping force measurement

50
 (a) Releasing clamping force (b) Applying clamping force

Figure 3.5: Clamping force modeling

[47], yet they usually involve complicated models and extensive experiments, which are not
applicable for real-time estimation.

In order to capture the hysteresis while preserving the simplicity of the polynomial model,
an alternative solution is proposed here. We first study the transition from the applying to
the releasing operation. Figure 3.4(b) shows the test results of the clamping force starting
from the same position and then reversing at various points. From Fig. 3.4(b), we see that
all the curves initially follow the same trajectory during the applying. From the overlapped
applying curves, the clamping force trajectories begin to differentiate during the releasing
operation according to their reverse points. One observation is that the releasing clamping
forces are almost parallel to each other despite differences in the reverse points. This indicates
that the clamping force with different reverse points might be derived from one reference
curve.

Based on the above observation, the approach to obtaining the releasing clamping force
operation is illustrated by Fig. 3.5(a). The blue curve is called the reference clamping force,
which represents the envelope of the EMB operation region. It is modeled by the polynomial
functions. The red curve is an arbitrary trajectory of the clamping force. Taking the contact
h iT h iT
point as the origin, two vectors are constructed, vrev = Frev θrev and vref = Fref θref ,
which are the vector from the origin to the reverse point and its extension on the reference
curve, respectively. The ratio between vrev and vref is

∥vrev ∥
rr = , (3.19)
∥vref ∥

51
where ∥·∥ denotes the 2-norm.

Using rr , the unknown clamping force at point v2 can be calculated by scaling the vector
v1 on the reference curve " # " #
Fbr (θ2 ) Pref (θ1 )
= rr , (3.20)
θ2 θ1

where θ1 and θ2 are respectively the motor angle corresponding to v1 and v2 , Fbr is the releasing
clamping force of interest, and Pr is the polynomial model of the reference releasing clamping
force.

The applying clamping force is modeled similarly, as shown in Fig. 3.5(b). The idea is
to model the unknown clamping force by scaling and rotating the reference curve. Equations
(21) and (22) give the scaling factor and rotation angle

∥vrev − vL ∥
ra = , (3.21)
∥vref − vL ∥

⟨vrev − vL , vref − vL ⟩
cos γ = , (3.22)
∥vrev − vL ∥ ∥vref − vL ∥
where vrev is the current reverse point, vL is the last reverse point, vref is the applying
clamping force on the reference curve at the same motor angle as that of vrev , and ⟨·⟩ denotes
the inner product.

The unknown applying clamping force is mapped by the reference clamping force between
vref and vL using follwing equation
" # " #" #
Fba (θa∗ ) − FL cos γ − sin γ Pa (θa ) − FL
= ra , (3.23)
θa∗ − θL sin γ cos γ θa − θL

where θa and θa∗ are respectively the motor angle of the reference and the modeled applying
clamping force, Fba is the applying clamping force of interest, and Pa is the polynomial model
of the reference applying clamping force.

The proposed clamping force model produces the hysteresis by simple linear transforma-
tion without introducing additional dynamics and parameters, which is suitable for real-time
implementation for clamping force control.

To validate the proposed clamping force model, experimental tests have been conducted.

52
Motor angle ramp inputs with different rate are performed to achieve different clamping
force profiles. Then the clamping force measurements are compared with the output of the
estimation to validate its performance. The results are shown in Fig. 3.6, where the input
rates are 16, 8, and 4 rev/s and the maximum clamping forces are 24, 18, 12, and 5kN.

We can see that the clamping force of the simulation agrees closely with the experimental
data under various input conditions. This demonstrates the proposed clamping force model
can be used to produce accurate clamping force.

53
 (a) Motor angular rate: 16 rev/s

(b) Motor angular rate: 8 rev/s

(c) Motor angular rate: 4 rev/s

Figure 3.6: Comparison between clamping force model and experimental results
54
3.3 Clamping Force Control System Development

The objective of the clamping force control system is to realize the clamping force requests
that might be from either the driver or the high-level control systems. Despite its simple
and generic SISO control structure, the clamping force control is subject to several practical
challenges. First, the clamping force is not measured, due to both the cost and the technical
constraints of installing a load cell on the brake pad. Second, the gap management algorithm
should be properly integrated to achieve a seamless transition between gap closing and
force tracking control. Last but not least, the controller needs to be robust against model
nonlinearities and disturbances such as varying pad stiffness and friction.

A high-level schematic of the proposed clamping force control system in Fig. 3.7. As
shown in Fig. 3.7, clamping force Fcl is estimated online using the available motor angular
position and speed measurements and a novel clamping force model. Motor torque command
Tm is generated from a disturbance-observer-based tracking controller with robustness to model
uncertainty and disturbance. The reference clamping force Fcl∗ is given by the gap management
block which unifies the gap closing and force tracking without controller switching.

3.3.1 Clamping Force Estimator

There are two unknown inputs in the system, i.e., the clamping force Fcl and the lumped
friction torque on the motor shaft Tf . One may attempt to use the common unknown input
disturbance observer (UIDO) to estimate these unknown disturbances. However, the real
challenge here is the lack of observability of this system. The idea of UIDO is to augment the
unknown disturbance into the system states by constructing an auxiliary dynamic system
based on which the disturbances are generated. The augmented plant is described as
   
θ̇m " # θm " #
A B d Cd   Bu
θ̈m  = θ̇m  + Tm , (3.24)
 
0 Ad 0
ż z

Figure 3.7: High-level schematic of clamping force control system

55
  
h i θm
y = C 0 θ̇m  , (3.25)
 

z
" # " # " # " #
0 1 0 0 0 1 0
where A = , Bu = ,A= , and A = . J is the lumped inertia
0 0 1 −itot/J −1/J 0 1
on the motor shaft, itot is the total transmission ratio of reduction gear and ball screw. Tm is
the motor torque, and z is the states of disturbances.

Then an observer can be designed to estimate the augmented states if the pair
(" # )
A B d Cd h i
, C 0 (3.26)
0 Ad

is observable.

A frequent choice is to model the disturbances as piecewise constant inputs, which

gives Ad = 0 and Cd = I. However, it turns out that the pair in (3.26) is not observable
in this way. An intuitive explanation is that Fcl and Tf perturb the state equation in the
same way through the same channel. Therefore, it is impossible to differentiate them by
the observer. A solution is to replace the fictitious dynamics of the step-wise friction model
with a physical-based model, e.g., the LuGre model as in [48]. However, such a method will
require the friction parameters to be precisely known. This is a rather strong and impractical
assumption, considering that the friction is supposed to be an “unknown disturbance” in the
first place.

In our study, an alternative approach for clamping force estimation is used. The algorithm
is based on the clamping force model introduced in Section 3.2.2, where the clamping force
is derived from two reference curves, respectively for applying and releasing. The clamping
force estimation is then equivalent to finding these two reference curves.

To obtain the coefficients of the polynomial model of the reference curves, we use an idea
similar to that first introduced in [49]. The torque balance of the EMB system is given by

Tm,A = Fcl,A itot + J θ̈m,A + Tf,A , (3.27)

Tm,R = Fcl,R itot + J θ̈m,R + Tf,R , (3.28)

where the subscripts A and R denote the applying and releasing operation, respectively.

56
 Figure 3.8: Motor angle input for clamping force estimation in applying

Adding Eqns. (3.27) and (3.28) together and rearranging the equation, we get

1 h i
Fcl,A + Fcl,R = Tm,A + Tm,R − J θ̈m,A − J θ̈m,R − Tf,A − Tf,R . (3.29)
itot

Equation (3.29) can be used for clamping force estimation given several assumptions. Firstly,
it is assumed that the absolute value of Tf is determined by θm only, that is, |Tf | = f (θm ).
Secondly, we assume that the hysteresis is small if the motor angle amplitude is small.

Given the preceding assumptions, when θm,A = θm,R and θ̇m,A θ̇m,R < 0, the following
relationships hold
Tf,A = −Tf,R , (3.30)

Fcl,A = Fcl,,R . (3.31)

Thus Eq. (3.29) can be simplified to

1 h i
Fcl = Tm,A + Tm,R − J θ̈m,A − J θ̈m,R , (3.32)
2itot

where Tm is directly measured and θ̈m can be obtained by the numerical derivative of
measurement θ̇m .

In Eq. (3.32), clamping force is obtained by the readily available motor torque and
angular speed. Then this estimation is used for polynomial reference curves identification
for the clamping force model. For implementation, the algorithm requires the motor to pass
the same angular position in opposite directions within a small amplitude. A specific input
profile is designed to facilitate the procedure, as shown in Fig. 3.8. It is shown in Fig. 3.8
that the motor angle consists of two components, a slow ramp input with large amplitude
superimposed by a high frequency sinusoid with small amplitude. The amplitude of the ramp
input decides the range of motor angles that the stiffness curves cover and produces a large
hysteresis. The sine wave enforces the motor passing the same angle in opposite directions

57
 Table 3.1: The sign of measurements and their derivatives before/after the contact
     
Mode sgn (Tm ) sgn Ṫm sgn θ̇m sgn θ̈m
Speed control 1/1 0/1 1/1 0/-1

in each period, and the small amplitude ensures that the local hysteresis is not excited too
much. For the releasing curve, the sinusoid will be added to the descending branch instead.

The shape of the reference curve will change due to pad wear and temperature [49].
Heuristic approaches can be used to compensate for their effect. Because the wear condition
changes slowly during the lifespan of the brake system, the characteristic curve can be
reidentified regularly in the maintenance process. It is shown in [49] that the reference curves
at different temperatures can be differentiated by a scaling factor. Therefore, reference curves
at different temperature can be adapted from a basic curve using a scaling factor as a function
of temperature [50].

3.3.2 Gap Distance Estimation

When the EMB is at rest, a certain air gap between the brake pad and disk must be maintained
to avoid unintended braking and overheating. However, the air gap distance will gradually
vary from its initial setup due to mechanical wear, degrading the performance of clamping
force estimation and control. Therefore, contact detection and air gap distance estimation is
indispensable for EMB clamping force control.

Due to the lack of an automatic wear adjustment mechanism [51] and contact pin [52] in
our EMB system, the contact point must be estimated through the existing measurements.
Depending on how the motor is controlled, measurements will change differently upon contact.
In previous studies, various signals such as Ia [47] and dTm
dθm
[49] have been used as indicators
for contact detection. Table 3.1 gives the sign of the measurements and their derivatives
before and after the instant of contact under motor speed control.
 
Under motor speed control, sgn Ṫm = 0 before contact, as only constant Tm is needed
 
to compensate for the bearing friction. Also sgn θ̈m = 0 because the motor speed is
regulated to the constant set point. As soon as contact occurs, motor
 speed
 will immediately
decrease due to the generated clamping force. Furthermore, sgn Ṫm will also change to
positive, because the controller will request more control effort trying to keep the speed to
the set point. Therefore, among the available measurements, both Ṫm and θ̈m are suitable for

58
 Figure 3.9: Closed loop speed control system

contact detection, as their directions change immediately upon contact.

In this study, θ̈m is chosen for contact detection. The condition is expressed as

θ̈m > S, (3.33)

where S is a small positive number to avoid false detection due to noise.

Choosing θ̈m as the indicator of contact, two tasks are left. First is to design a speed
controller such that the motor speed reaches the set point before the contact. The other is to
derive θ̈m from the noisy measurement of θ̇m .

The motor speed control model is written as

J ω̇m = Tm − Bωm + d, (3.34)

where ωm is the motor angular speed, B is the damping coefficient approximating the viscous
friction, and d is the external disturbance, including friction and clamping force.

The closed loop system is shown in Fig. 3.9. In Fig. 3.9, Ps (z) is the discretized plant
by the zero-order-hold method, and Ks (z) is the controller to be designed. This is a classic
motor speed control problem. However, for our application, it is not a good idea to make the
controller yield the best tracking and disturbance rejection, because it will mask the change
of θ̈m and raise the difficulty in detecting the contact point. Rather, we want the controller
to work just fine before the contact and then degrade after contact such that θ̈m changes as
much as possible upon the contact.

Based on the preceding reasoning, the well-known internal model principle (IMP) [53] is
used to design the controller. The key difference in control before and after contact is the

59
 Figure 3.10: Simulation results of IMP based speed control

type of disturbance. The disturbance pre-contact is mainly bearing friction, which resembles
a step disturbance, while the post-contact disturbance largely comes from the clamping force,
which appears like a nonlinear spring. The IMP tells us that in order to reject the input
disturbance asymptotically, the feedback controller must incorporate the “structure” of the
disturbance into it. For a step disturbance, the transfer function of controller thus must have
the term 1 − z −1 in this denominator.

The controller and plant are written as

Nk (z −1 )
Ks z −1 = (3.35)


,
Dk (z −1 ) Dd (z −1 )

z −1 NP (z −1 )
−1
(3.36)


Ps z = .
DP (z −1 )
According to the IMP and the control objective, we choose Dd (z −1 ) = 1 − z −1 to reject the
constant friction disturbance only. Then Nk (z −1 ) and Dk (z −1 ) are determined by solving
the Diophantine equation

z −1 NP z −1 Nk z −1 + DP z −1 1 − z −1 Dk z −1 = η z −1 , (3.37)

where η (z −1 ) = 1 + η1 z −1 + η1 z −1 + . . . + ηq z −q is the desired closed-loop characteristic

polynomial.

Figure 3.10 shows the simulation of the IMP controller. The reference speed input is a
step with amplitude of 10 rev/s passed through a 1st order low-pass filter. We can see from

60
Fig. 3.10 that the settling time of motor speed is about 0.3s. There is no steady-state error
after the transient. At 1.1s, when the brake pad touches the disk, the motor speed diverges
from the set point immediately and the motor torque also starts to increase. This is expected
because the IMP-based controller is able to handle only a constant disturbance.

Now the task is to calculate the motor angular acceleration from the speed measurement.
A Rauch–Tung–Striebel (RTS) smoother is used here to obtain the acceleration from a noisy
speed measurement. The RTS smoother is a two-way Kalman filter that finds the optimal
estimate of the state at each time step k while using the measurements up to and including
the final time N [54].

The discrete-time estimation model for the motor speed and acceleration is constructed
as
...
" # " #" # " 2#
Ts
θ̇m,k+1 1 Ts θ̇m,k
= + 2 θ m,k , (3.38)
θ̈m,k+1 0 1 θ̈m,k Ts
" #
h i θ̇
m,k
θ̇m,k = 1 0 + vk , (3.39)
θ̈m,k
... ...
where Ts is the sample time and θ m,k ∼ (0, Q) and v k ∼ (0, R) are respectively the process
and measurement noise, assuming both Gaussian, zero-mean, uncorrelated, and white.

The state of the system is first estimated by a standard forward Kalman filter

−
Pf,k +
= F Pf,k−1 F T + GQGT , (3.40)

−1
−
Kf,k = Pf,k −1 T
H T HPf,k H +R , (3.41)

b−
x b+
f,k = F x f,k−1 , (3.42)

b+ b− b− (3.43)


x f,k = x f,k + Kf,k yk − H x f,k ,

+
Pf,k −
= (I − Kf,k H) Pf,k (I − Kf,k H)T + Kf,k RKf,k
T
. (3.44)

Then the following RTS smoother equations are initialized and executed backwards for
k = N − 1, . . . , 1, 0
x
bN = xb+
f,N , (3.45)
+
PN = Pf,N , (3.46)

−1
−
If,k+1 −
= Pf,k−1 , (3.47)

61
 Figure 3.11: Simulation results of contact detection

+
Kk = Pf,k −
F T If,k+1 , (3.48)
+ −
− Pk+1 KkT , (3.49)


Pk = Pf,k − Kk Pf,k+1

b+ b− (3.50)


x
bk = x bk+1 − x
f,k + Kk x f,k+1 ,

where matrices F , G, and H are the state, input, and output matrix in (3.38) and (3.39),
respectively. Q and R are the covariance of process and measurement noise, respectively.
Pf is the state covariance matrix. Kf is the Kalman filter gain. x
bf is the state estimate. y
is the measurement. The superscript + and − represent a posteriori and a priori estimate,
respectively. Details on the RTS smoother can be found in [54].

Figure 3.11 shows the simulation results of contact detection. The real air gap between
the brake pad and disk is 9.5 motor revolutions. As shown in Fig. 3.11, the motor acceleration
estimation from the RTS smoother is very close to the actual acceleration. The threshold S
is set as small as 11rev/s2 and the resulting gap distance estimation is 9.51rev. Figure 3.11
also shows the acceleration estimation using finite difference passed by a 1st order low-pass
filter. The obtained signal not only suffers from significant noise, but also shows noticeable
phase lag. In order to rule out the fault detection, the threshold must be raised to 25rev/s2 ,
which gives an overestimated air gap of 9.72rev.

In summary, the proposed distance detection and distance estimation algorithm is fast
and accurate. In practice, the whole procedure should be performed on a regular basis when
the vehicle starts up. The estimated gap distance can be used for adjusting the initial motor
position to maintain the preset value.

3.3.3 Clamping Force Tracking

The clamping force is controlled to track the reference by the motor. Generally, the EMB
system should have a fast and robust response with respect to disturbances and model
uncertainties. The mathematical model for the EMB clamping force control can formulated

62
 Figure 3.12: Block diagram of the DOB control of clamping force

as
J θ̈m = Tm − itot Kcl θm − Tf , (3.51)

where Kcl is a time-varying nonlinear spring element to capture the nonlinearity and hysteresis
of the actual clamping force.

A disturbance-observer-based (DOB) control is used in our study. This method has been
proven to be very effective in high-precision position control such as hard disk drives [55]
and positioning tables [56]. The control schematic is shown in Fig.3.12. Each block in Fig.
3.12 represents a discrete-time transfer function. On the prototype EMB system, the CAN
sample time is 2ms, which is used for clamping force regulation, while the DC motor current
controller is operated at a much faster 200µs sample time. Therefore, the motor controller is
neglected in the following study.

In Fig. 3.12, Gp (z −1 ) is the plant model with unknown parameters and dynamics.
n (z ) is the inverse of the nominal model factorized out of the delay term z
G−1 . Q (z −1 )
−1 −m

is a low-pass filter, whose role will be discussed shortly. r (k), Tf b (k), d (k), and n (k) are the
reference, feedback control input, disturbance torque, and measurement noise, respectively.
By a straightforward algebra, the transfer function from Tf b (k), d (k), and n (k) to the output
Fcl (k) are respectively expressed as

Gp (z −1 ) Gn (z −1 )
GFcl Tf b = , (3.52)
Gn (z −1 ) + (Gp (z −1 ) − Gn (z −1 ) z −m ) Q (z −1 )

Gp (z −1 ) Gn (z −1 ) (1 − Q (z −1 ) z −m )
GFcl d = , (3.53)
Gn (z −1 ) + (Gp (z −1 ) − Gn (z −1 ) z −m ) Q (z −1 )
−Gp (z −1 ) Q (z −1 )
GFcl n = . (3.54)
Gn (z −1 ) + (Gp (z −1 ) − Gn (z −1 ) z −m ) Q (z −1 )

Chosen as a low-pass filter, Q (z −1 ) ∼

= 1 in the low-frequency range and Q (z −1 ) ∼
=0

63
 Table 3.2: Asymptotic properties of transfer functions with DOB control
Low frequency where Q (z −1 ) ∼
= 1 High frequency where Q (z −1 ) ∼
=0
GFcl Tf b Gn (z −1 ) Gp (z −1 )
GFcl d 0 Gp (z −1 )
GFcl n -1 0

in the high-frequency range. If the delay is small, such that Gn (z −1 ) ∼

= Gn (z −1 ) z −m , then
Equations (3.52)–(3.54) can be approximated to Table 3.2.

From Table 3.2, the DOB control has the following key features:

(1) In low frequency, the control system resembles the nominal plant Gn (z −1 ), and the
disturbance is rejected.

(2) In high frequency, the control system degrades to the original plant Gp (z −1 ), and the
effect of measurement noise is attenuated.

The first feature is particularly useful, because it allows feedback and feedforward controllers
to be designed on the nominal system, regardless of the model uncertainties and disturbances.
For our application, it means that the nonlinearity and hysteresis of the clamping force
can be regulated to a constant linear spring element. Then simple linear controllers can be
used in the outer loop for reference tracking without introducing any complex nonlinear or
adaptation techniques.

The second feature tells us that the bandwidth of Q (z −1 ) must be limited according
to the measurement noise, a typical trade-off in feedback controller design. Moreover, the
design of the Q-filter also closely relates to the robust stability of the uncertain system by
representing the uncertainty as a multiplicative perturbation

Gp z −1 = G∗n z −1 1 + ∆ z −1 , (3.55)

where G∗n (z −1 ) = Gn (z −1 ) z −m .

A sufficient condition of robust stability is given by

∆ e−jωT Q e−jωT < 1, ∀ω. (3.56)

64
 Table 3.3: Uncertain parameters of plant model
Parameters Nominal value Variation (%)
J 1.1 × 10−5 (Kgm2 ) [−10, 10]
Kcl 0.22(kN/rev) [−15, 70]
B 2 × 10−4 (Nms/rad) [−10, 70]
τm 1(ms) [−10, 10]

Figure 3.13: Step response of linear and nonlinear EMB models

A candidate Q-filter suggested in [57] is used in this paper

3τ s + 1
Q (s) = , (3.57)
(τ s + 1)3

where τ is the time constant of the Q-filter. The Q-filter is discretized using bilinear
transformation for implementation.

In order to determine the time constant of the Q-filter, ∆ (z −1 ) is calculated using the
following equation

Gp (z −1 ) − G∗n (z −1 )
∆ z −1 = . (3.58)
G∗n (z −1 )

In Eq. (3.58), Gp (z −1 ) is a zero-order-hold discretization of the continuous plant with

uncertain parameters
Kcl 1
Gp (s) = 2 . (3.59)
Js + Bs + itot Kcl τm s + 1

The nominal value and variation of uncertain parameters in Gp (s) are summarized in
Table 3.3.

G∗n (z −1 ) is defined by taking the nominal values in Table 3.3 and letting τm = 0. Figure
3.13 shows the step response of uncertain model Gp (s), nominal model G∗n (z −1 ), and nonlinear
plant model with an amplitude of 15kN. We can see that the nominal model and the nonlinear

65
 Figure 3.14: Magnitude response of Gp (z −1 ) and ∆ (z −1 )

Figure 3.15: Q-filter bandwidth selections

model have the same steady-state value. However, there is discrepancy in the transient
response. The low stiffness in the small clamping force makes the nonlinear model response
slower than that of the linear model, while high stiffness in the large clamping force makes
its response faster. Although the nonlinear model cannot be represented by a single nominal
model, its response is mostly covered by the set of the uncertain model. Therefore, if the
proposed controller can provide robust performance against the uncertain model, it will also
give similar performance to the nonlinear model. Figure 3.14 shows the magnitude response
of 60 random samples of Gp (z −1 ) and ∆ (z −1 ) as their upper boundary.

According to the stability condition, τ should be selected such that (3.56) is satisfied.
Figure 3.15 shows 3 candidates of Q-filter with comparison to the 1/∆. It is seen that among the
three of them, Q1 (z −1 ) does not meet the condition, while the other two do. The bandwidth
of the Q-filter also should not be too high, due to the noise and possible unmodeled dynamics
in the high frequency. Therefore, Q2 (z −1 ) is applied.

After determining the disturbance observer, the feedback and feedforward controller
design is straightforward. A PI controller is used for feedback control. The loop shaping
design method is applied such that the gain crossover frequency and phase margin meet the

66
specifications, which are written as the following equations

Kf b ejωc T Gn e−jωc T = 0, (3.60)

arg Kf b ejωc T Gn e−jωc T = P M − π, (3.61)





where is Kf b ejωc T the freqeuncy response of the discrete-time PI controller, ωc is the gain

crossover frequency of the loop transfer function, and P M is the phase margin.

Originally introduced by Tomizuka [58], a zero-phase error tracking controller (ZPETC)

is designed as feedforward controller based on the approximated inversion of the closed-loop
system. Denote the closed-loop system with the nominal model and PI controller as

z −d Bcs (z −1 ) Bcu (z −1 )
Gcl z −1 = , (3.62)
Ac (z −1 )

where d = 1 is the relative degree of the transfer function, Bcs (z −1 ) is the part of the numerator
that is suitable to invert, and Bcu (z −1 ) = buc0 + buc1 z −1 + . . . + bucs z −s is the one that is not
suitable to invert, which must include unstable zeros outside the unit circle. All three zeros
of Gcl (z −1 ) are inside the unit circle in our case, but the zero that locates near (−1, 0) is not
inverted in order to avoid high frequency oscillation.

The ZPETC is given by

z d Ac (z −1 ) Bcu (z)
Kf f z −1 = . (3.63)
Bcs (z −1 ) [Bcu (1)]2

Notice that z d and Bcu (z) = buc0 + buc1 z + . . . + bucs z s are not causal. Therefore, a total of d + s
steps preview of the reference input is needed.

An extra reference model Kr (z −1 ) is added in addition to Kf f (z −1 ) to further shape

the required input-output response of the overall system. Figure 3.16 shows the Bode plot of
Kr (z −1 ) and Kr (z −1 ) Kf f (z −1 ) Gcl (z −1 ). It can be seen that the frequency response of the
overall system is very similar to that of the reference model. Note that their phase responses
are essentially identical thanks to the ZPETC.

3.3.4 Unified Control Structure

An underlying assumption that has been made throughout the development of the tracking
controller is that the brake pad already touches the disk at the operation of the controller.

67
 Figure 3.16: Frequency response of Kr (z −1 ) and Kr (z −1 ) Kf f (z −1 ) Gcl (z −1 )

Figure 3.17: High-level schematic of unified control system

However, as we know, there is an initial air gap between the brake pad and disk when the
EMB starts from rest. This gap must be handled carefully; otherwise, a large impact will
occur upon contact and damage the system. One approach that is often used is to apply
an additional position controller to eliminate the initial gap and then switch back to the
force controller [36, 43] at contact. However, the whole control system becomes complicated
due to the extra position controller. Furthermore, smooth transition generally cannot be
automatically achieved, even if each controller is well designed separately.

Here, a unified framework is proposed so that the same DOB-based tracking controller
can be used for gap closing, resulting in a seamless transition. Since the DOB-based controller
takes the force as reference, the motor position must be converted to force in order to utilize
the same controller. Figure 3.17 depicts the high-level schematic of the unified control system.

Two additional blocks are used for the unified system. In the clamping force processor,
the estimated clamping force is augmented by a fictitious force, as shown in Fig. 3.18. As
shown in Fig. 3.18, the actual clamping force is elevated by FG , and a segment of linear

68
 Figure 3.18: Augmented clamping force

fictitious force is inserted by connecting the origin and (θG , FG ), where θG is the initial
gap distance. This fictitious force is responsible for converting the motor position to a
corresponding clamping force. FG is set such that the slop of the fictitious force is equal
to the nominal stiffness to retain the best performance of the force control. Because the
fictitious force Ff ic cannot be actually generated by the plant, it will be added to the motor
torque command to mimic its effect.

The reference processor in Fig. 3.17 modifies the reference force to accommodate the
fictitious force. Equation below gives the modified reference

ω FG t
G θG if θm < θG
Fcl∗ = , (3.64)
F + F
r G else

where Fr is the effective reference force and ωG is the designed travel speed. The larger ωG is,
the faster the gap is eliminated. We see that the proposed method not only unifies the force
and the position control but also provides a simple tuning knob for adjusting the speed of
gap closing.

3.4 Simulation Results

The controller is implemented in Matlab/Simulink to verify its performance. The nonlinear

EMB model in Section 3.2 is used as the plant. Model parameters used in the simulations
are listed in Appendix.

69
 (a) Fitting points

(b) Polynomial fitting for reference curve

Figure 3.19: Identification of applying clamping force

3.4.1 Clamping Force Estimation

The polynomial functions of the reference clamping force are first identified. Figure 3.19
shows the simulation results under the input profile that is given in Fig. 3.8. In Fig. 3.19, we
can see that the applied force from the motor consists of multiple small circles corresponding
to the high-frequency superimposed sinusoids. The data points used for reference curve
identification then can be derived from those small circles, which are indicated as circular
marks. The polynomial function is identified using the least square method based on the
fitting points obtained in Fig. 3.19(a). Figure 3.19(b) shows the identified reference clamping
force compared to the actual one using a 3rd order polynomial function. We can see that the

70
 (a) Fitting points

(b) Polynomial fitting for reference curve

Figure 3.20: Identification of releasing clamping force

two curves are virtually identical.

Similarly, the reference releasing clamping force is identified. Figure 3.20 shows the
derivation of the fitting points, where the high-frequency sinusoid is added to the releasing
operation. Figure 3.20(b) is the result of reference curve fitting using a 7th-order polynomial
function. We can see that the identified reference curve is also very close to the actual one.

The two identified reference clamping forces are used in the clamping force model for
esimation. The performance of this estimation will be evaluated based on the tracking control
simulation.

71
3.4.2 Clamping Force Tracking Control

To validate the tracking performance, simulations are performed with various clamping force
commands. Firstly, the stair-step reference clamping force with large and small amplitudes is
used.

Figure 3.21 shows the simulation results where the step amplitude is 12kN, which is
about half of the maximum clamping force that can be achieved by our prototype. This
condition is considered an extreme situation, which could happen in reality, for example,
when the driver or automatic emergency brake (AEB) system suddenly demands a large
amount of brake to avoid a collision.

We can see in Fig. 3.21(a) that the step response of the clamping force has a settling
time of less than 0.2s with no overshoot and very small steady-state error. The motor torque
input is also below the 2Nm constraint. Notice that the motor torque is not symmetric in the
applying and releasing operations due to the nonlinearity and hysteresis, while the output
clamping force shows very uniform responses whether applying or releasing, which clearly
indicates the nominal model regulation ability of the DOB-based controller. Figure 3.21(a)
also gives the response of fast and slow gap closing, respectively, by simply adjusting the ωG
in Eq. (3.64). No matter what ωG is used, the transition to the force control upon contact is
smooth and the tracking performance is identical.

The clamping force estimation is shown in Fig. 3.21(b). The overall performance of
the estimation is good, with a maximum error of less than 500N. The largest error happens
during the release operation at small clamping force.

The control performance under stair-step reference with small amplitude of 2kN is also
evaluated. This simulation is used to evaluate the control ability when the driver or high-level
control requires a more refined brake control. As shown in Fig. 3.22(a), the clamping force
has a generally consistent response, where a similar settling time and steady-state error is
achieved in each step. Slightly larger steady-state errors can be found in the release operation
at small clamping force. This is due to the less accurate clamping force estimation in this
region, which is shown in Fig. 3.22(b).

Next, simulations on sinusoidal references tracking are performed. This kind of simulation
mimics the modulation of the brake force during operation of the anti-lock brake system

72
 (a) Clamping force tracking. Blue solid line: reference clamp-
ing force; Red dashed line: Actual clamping force with fast
gap closing; Yellow dash-dotted line: Actual clamping force
with slow gap closing

(b) Clamping force estimation

Figure 3.21: Simulation of stair-step reference with large amplitude

73
 (a) Clamping force tracking

(b) Clamping force estimation

Figure 3.22: Simulation of stair-step reference with small amplitude

74
 Figure 3.23: Clamping force tracking of sinusoidal reference with 50% modulation

Table 3.4: Control performance of sinusoidal reference with 50% modulation

50% modulation about 15kN Amplitude (%) Phase lag (◦ )
0.5Hz 100 9.2
1Hz 99.5 16.2
2Hz 85.5 33.1

(ABS) or electronic stability control (ESC).

Figure 3.23 shows the simulation results of the sinusoidal reference with 50% amplitude
modulation about 15kN, whose frequency is respectively 0.5Hz, 1Hz, and 2Hz. We can see
that the response of the clamping force is almost linear, resembling the dynamics of a low-pass
filter. This is again an outcome of the DOB-based controller. Table 3.4 lists the amplitude
and phase lag of the clamping force to the reference.

Figure 3.24 shows the simulation results of the sinusoidal reference with 10% amplitude
modulation about 15kN. Compared to the preceding results, the response with small modula-
tion appears less linear. This is due to the stick-slip friction, which is usually more pronounced
in slow motion. Nevertheless, all the performance are reasonably good, as summarized in
Table 3.5.

Table 3.5: Control performance of sinusoidal reference with 10% modulation

50% modulation about 15kN Amplitude (%) Phase lag (◦ )
0.5Hz 98.3 12.4
1Hz 96 15.5
2Hz 93 34.6

75
 Figure 3.24: Clamping force tracking of sinusoidal reference with 10% modulation

Simulation with more demanding reference input is presented in Fig. 3.25(a), where the
reference is a 10Hz sinusoidal signal. It is shown that clamping force can be tracked with
reasonable amplitude decay and phase lag. Figure 3.25 also shows a comparison with cases
in which either the ZPET or the disturbance observer is absent to further demonstrate the
benefits of the proposed control strategy. In Fig. 3.25(a), we can see that the amplitude and
phase error of tracking with ZPET is much smaller than without it. The benefit of ZPET
in reducing the phase delay is clearly shown. Figure 3.25(b) shows the benefit of using the
disturbance observer. Without the help of the disturbance observer, the PID controller alone
cannot track the reference signal at all, since it must also take care of the disturbance by
itself.

3.4.3 Robust Performance Evaluation

In this section, we are going to evaluate the robustness of the control system performance
under different conditions. The reference input is 2Hz sinusoid in all cases. Firstly, the effect
of time delay is shown in Fig. 3.26. In Fig. 3.26, 2ms and 4ms time delays are respectively
added to the measurement feedback during simulation. The discrete-time DOB controller
can accommodate time delay easily by adjusting the delay steps m of z −m in Fig. 3.12[59].
By respectively setting m = 1 and 2, we can see that the time delay shows only minor impact
on control performance. With the longer time delay, tracking error becomes slightly larger
and more oscillatory at reverse points.

Secondly, the plant model parameters are varied to investigate their effect on control
performance. Figure 3.27 shows the simulation results with the variations of +100% bearing
friction, +50% motor time constant, and -50% pad mass, respectively. We can see from Fig.

76
 (a) With and without ZPET

(b) With and without disturbance observer

Figure 3.25: Comparison with different controller

77
 Figure 3.26: Effect of time delay

3.27 that the control performance is almost unchanged under variations in model parameters.
It shows that the proposed control system provides robust performance.

Finally, noise is injected into the measurements in the simulation to investigate how the
control performance is affected. Figure 3.28 shows the simulation results when motor speed
measurement is corrupted by white noise where measurement delay is also included. The
standard deviation of the noise is chosen as 10% of the amplitude of the clean signal. The
noisy speed measurement is then integrated to be used as a position measurement. Although
the tracking error shows some small oscillation and offset due to noise and delay, it shows
from Fig. 3.28 that the error magnitude is maintained similarly to that of nominal case.

78
 (a) +100% bearing friction

(b) +50% motor time constant

(c) -50% pad mass

Figure 3.27: Simulation with model parameters variation

79
 (a) Motor speed measurement with white noise

(b) Control performance

Figure 3.28: Simulation with noisy measurement

80
3.5 Conculsions

In this chapter, the clamping force control system for EMB is developed. Firstly, a novel
clamping force model is proposed that is able to capture both the nonlinearity and hysteresis
of the clamping force using only two reference polynomial functions. An estimation algorithm
that requires only available measurements is given to identify these reference polynomial
functions. Secondly, a contact detection and gap distance estimation algorithm is developed
based on the IMC-based controller and RTS smoother. Simulation results show that the
proposed algorithm gives an accurate air gap distance. Thirdly, a clamping force tracking
controller is designed based on a disturbance observer with PI and ZPET controllers. In
addition, a unified position/force control system is proposed to eliminate the initial gap before
the force control. Simulation results show that the clamping force is estimated accurately.
The tracking controller also performs well in all types of references.

81
 Part II

Optimization-based Control for

High-Level Motion Planning and
Trajectory Tracking

82
 Chapter 4

NMPC based Trajectory Tracking Control for Vehicle with 4WS and 4WID

Being a key component in the autonomous vehicle, the vehicle motion control system is
responsible for guiding the vehicle along the high level reference path while maintaining
driving safety. We will focus on the control algorithm development for vehicle motion control
system in this chapter and the following one. The motion control systems are of the higher
level than the actuator control system discussed in the previous chapters. Compared to
the low-level actuator controller, the design of vehicle motion control faces some additional
challenges as follows:

(1) The vehicle motion control system is a nonlinear multi-input-multi-output (MIMO)

system. Therefore, the control algorithm must be able to coordinate different actuators
to achieve the desired outputs of the system.

(2) The control system is subject to various constraints in terms of the physical limits
of actuators, safety and performance constraints on system states. So, the control
algorithm should also be capable of handling the constraints in a systematic way.

When considering all these requirements, it becomes clear that model predictive control
(MPC) is probably one of the most suitable control methodology that checks all the boxes. It
is well known that MPC provides a very flexible framework which enables the handling of
multi-variable nonlinear models with constraints. However, MPC also suffers from serious
computation demand when applied to fast and nonlinear systems. Therefore, the MPC formu-
lation and its numerical implementation must be chosen carefully with system characteristics
and control goals in mind. In the following two chapters, we will present in detail how the
Nonlinear-MPC (NMPC) technique can be applied to automotive motion control systems
practically with good performance.

83
 In this chapter, we will tackle the trajectory tracking problem for a vehicle equipped with
4-wheel-steering (4WS) and 4-wheel-independent-driving (4WID) system using the NMPC
method.

4.1 Introduction

The application of MPC to (semi-) autonomous vehicle trajectory tracking can be found in
many literature. Raffo et al. [60] presented a model predictive controller structure for solving
the path-tracking problem. Two linear MPC algorithms are compared to determine the most
appropriate implementation strategy. A cascade control structure is used which considers
both kinematic and dynamic control. In [61], a model predictive control based path tracking
algorithm was proposed to achieve accurate and smooth tracking for an autonomous vehicle.
An optimal steering command was calculated by using a QP optimization method based on the
integrated model. A model predictive control (MPC) approach for controlling an active front
steering system in an autonomous vehicle is presented in [62]. NMPC and linear-time-varying
(LTV) MPC are formulated, and their tracking performance and computational complexity are
compared. The effectiveness of the proposed MPC formulation is demonstrated by simulation
and experimental tests up to 21 m/s on icy roads. Similar LTV MPC frameworks are also used
in [63] and [64]. Gao et al. [65] proposed a computationally efficient path-following control
strategy of autonomous electric vehicles (AEVs) with yaw motion stabilization. A path-
following model, a single-track vehicle model, and a magic formula tire model, is constructed.
The continuation/generalized minimal residual (C/GMRES) algorithm is adopted to reduce
the computational burden. The simulation results show that the proposed strategy can achieve
desirable path following and vehicle stability efficacy, while greatly reducing the computational
burden. In [66], an MPC based path-following controller with steering angle envelopes is
proposed. Constraints regarding road sides and lateral stabilization are formulated as the
steering angle envelopes to achieve better computational efficiency. Xiang et al. [67] proposed
a path tracking coordinated control strategy for a Four-in-Wheel-Motor independent-drive
electric vehicle. The strategy consists of upper level model predictive controller and a low-level
force distributor. The NMPC is based on a 7-degree-of-freedom vehicle model with dynamic
adjustment of control target weight coefficients.

The controller developed in this chapter is used by an over-actuated vehicle with 4WS
and 4WID. In addition, the vehicle will likely operate near the handling limits where tire force
saturates. These facts raise two key questions that have to be addressed during controller
development: first, how to apply the control efforts effectively to the actuators of different
dynamic characteristics, and second, how to solve the nonlinear optimal control problem

84
 Figure 4.1: Schematic of NMPC trajectory tracking control system

efficiently in a real-time feasible manner. For the first question, the frequency shaping
technique will be incorporated into the NMPC framework, which allows us to compute control
inputs with consideration of actuators’ dynamical characteristics. While for the latter one,
we propose a fast NMPC algorithm based on real-time iteration (RTI) scheme with control
parameterization. Figure 4.1 depicts the schematic of the trajectory tracking NMPC control
system. The NMPC controller computes the control inputs to the vehicle based on the
reference signal from the upper-level system and feedback from the vehicle, such that the
vehicle’s actual states follow the reference in an optimal way.

The rest of this chapter is structured as follows. In Section 4.2, the vehicle model used
in the NMPC formulation is described. Section 4.3 presents the detailed formulation of the
NMPC control problem, which includes the frequency shaping method and the fast NMPC
algorithm. In Section 4.4, simulation results of the proposed control algorithm are discussed.
Finally, Section 4.5 summarizes the proposed controller design.

4.2 Vehicle Modeling

This section describes the vehicle model that will be used for the NMPC formulation in the
later sections. Vehicle models with various levels of complexity have been introduced in the
literature for different purposes. Regarding the NMPC based trajectory tracking application,
the model should capture the significant dynamics of the vehicle but also be as simple as
possible to enable real-time feasibility. The following assumptions are made when deriving
the vehicle model:

(1) The vehicle’s lateral acceleration may approach the limit during emergency maneuver,
e.g. obstacle avoidance.

85
 Figure 4.2: Schematic of chassis model

(2) The vehicle longitudinal speed may vary in a large range due to the reference command
from a higher level motion planner.

(3) The roll, pitch, and heave motion of the vehicle are not included. But the effect of
load-transfer is considered.

Based on the above assumptions, a double-track chassis model with longitudinal, lateral, and
yaw degrees of freedom (DOF), incorporating load transfer and a Magic Formula tire model,
is adopted in this study, which strikes a nice balance between fidelity and simplicity. The
following sections describe the derivation of the vehicle model.

4.2.1 Vehicle Model Development

The vehicle dynamics are modeled as a rigid body sitting in a two-dimensional global X-Y
plane, and its corresponding velocities are described in a body-fixed x-y coordinate. Figure
4.2 shows the schematic of the vehicle model. The model dynamics are given in Eqns. (4.1) -
(4.3) [63].

86
 x
mẍ = mẏ ψ̇ + Ff,l x
+ Ff,r x
+ Fr,l x
+ Fr,r , (4.1)

mÿ = −mẋψ̇ + Ff,l y y

+ Ff,r y
+ Fr,l y
+ Fr,r , (4.2)
y y y y x x x x
(4.3)




I ψ̈ = a Ff,l + Ff,r − b Fr,l + Fr,r + c −Ff,l + Ff,r − Fr,l + Fr,r ,

where m is the vehicle mass, I is the vehicle yaw moment of inertia, a and b are respectively
the distance of front and rear axles from the center of gravity, c is the half track, ẋ and ẏ
are respectively the vehicle longitudinal and lateral speed in body-fixed coordinate, ψ is the
vehicle yaw angle, ψ̇ and ψ̈ are respectively the yaw rate and acceleration of the vehicle, F x
and F y are respectively the tire forces along the longitudinal and lateral vehicle axles. We
use two subscript symbols to denote four wheels, where the first one representing front/rear
(f /r) axle, and the second one left/right (l/r) wheel.

The vehicle’s position in global frame is expressed as

Ẋ = ẋ cos ψ − ẏ sin ψ, (4.4)

Ẏ = ẋ sin ψ + ẏ cos ψ. (4.5)

The relationship between tire forces along the vehicle axes and tire forces along the
wheel axes are
F x = Fl cos δ − Fc sin δ, (4.6)

F y = Fl sin δ + Fc cos δ, (4.7)

where Fl and Fc are the tire longitudinal and cornering forces along wheel axes, respectively,
δ is the steering angle.

4.2.2 Tire Slip and Force

When steering the wheel, lateral slip angle α develops. It is the ratio of the wheel’s velocities
along the lateral and longitudinal axes, resolved in the wheel’s local coordinates, which is
illustrated in Fig. 4.3 and can be written as

vc
tan α = − , (4.8)
vl

where v c and v l are the cornering and longitudinal velocities, respectively, which are expressed
as
v l = v y sin δ + v x cos δ, (4.9)

87
 Figure 4.3: Wheel local coordinate system and slip angle

v c = v y cos δ − v x sin δ, (4.10)

where v x and v y are the wheel velocities along vehicle body-fixed axes, which are given as

vfx/r,l = ẋ − cψ̇, (4.11)

vfx/r,r = ẋ + cψ̇, (4.12)

and
y
vf,l/r = ẏ + aψ̇, (4.13)
y
vr,l/r = ẏ − bψ̇. (4.14)

When torque is applied to the wheel, longitudinal slip ratio λ is generated, which is
defined as
Rw ω − v l
λ := ∈ [−1, ∞) (4.15)
vl
where Rw is the wheel effective radius, and ω is the wheel rotational velocity. Fig 4.4 shows
the schematic of wheel longitudinal slip.

While it is possible to treat λ as a system state by including wheel rotational dynamics,

we will directly regard λ as the control input to the vehicle model that will be determined
by the NMPC controller. This setup is valid if we assume there is a low-level controller
responsible for tracking the slip ratio reference passed by the NMPC controller. Such wheel
slip controllers have been studied for decades by many researchers and can be found more
and more common in the modern vehicles [68, 69]. The motivation for doing so is threefold:

(1) The wheel dynamics are usually orders of magnitude faster than the chassis dynamics.

88
 Figure 4.4: Wheel longitudinal slip schematic

If included in the vehicle model, it will render the underlying dynamical system stiff,
which hinders the numerical solving of NMPC.

(2) The wheel dynamics increases the number of states by four, which again makes numerical
computation more difficult.

(3) The wheel dynamics are often subject to many uncertainties, such as road friction
variation, mechanical friction, and external disturbance. Dedicated control algorithms
have been proposed in many literature to address these challenges and achieve good
performance. Therefore, it is wise to let those controllers do the job rather than leave
all the burden on the NMPC controller.

However, caution must be taken when we formulate the NMPC using this configuration.
Since the slip ratio becomes the control input now, we must carefully choose the weighting
parameters and constraints such that its profile will not exceed the capability of the physical
actuator.

The vehicle’s motion is almost exclusively determined by the tire forces. Therefore, it is
very important to have a tire model that realistically reflects the tire force. In this study, we
choose the Pacejka’s Magic Formula model, given by [70]

F0 (m) = µF z sin (C arctan (Bm − E (Bm − arctan (Bm)))) , (4.16)

where m is either λ or α depending on the querying tire force, B, C, E are constant coefficients
obtained from experimental data, µ is the road friction coefficient, and F z is tire load. Figure
4.5 shows the tire force form Eq. (4.16) for different road surface conditions under the nominal

89
 Figure 4.5: Tire force form Magic Formula Model with F z = 5718N

Figure 4.6: Tire force form under combined slip condition with F z = 5718N

tire load when the vehicle is at rest. In Fig. 4.5, the µ = 1, 0.4, and 0.17 receptively for the
asphalt, snowy and icy surface, which are typical values based on the experimental results
[71].

Equation (4.16) only describes the tire force under pure slip condition, i.e. either λ or
α is zero. Since the NMPC controller is supposed to simultaneously perform steering and
braking/driving, it is necessary to model the tire force in a combined slip condition. In this
study, the friction-ellipse method [72] is used. This method computes the cornering force Fc
given the longitudinal force Fl as
s  2
Fl
Fc = F0,c (α) 1− , (4.17)
µl F z

where Fl is computed by Eq. (4.16). Figure 4.6 shows the tire forces under combined slip
conditions using Eq. (4.17).

90
4.2.3 Load Transfer

We can see from Eq. (4.16) that the tire force is affected by tire normal load F z . During
at-the-limit maneuvers such as obstacle avoidance, the vehicle will experience very high lateral
acceleration, thus inducing significant load transfer. In order to include this effect while
avoiding modeling complex suspension dynamics, we will make the following assumption:

The transient load shift effect due to suspension dynamics is not modeled. The load
transfer is calculated algebraically by assuming that the normal forces produce moments
about the C.G. that are in static equilibrium with the C.G. moments due to the forces at the
tire contact patches.

With the above assumption, the steady-state load transfer model is expressed straight-
forwardly as

h
2∆F̄xz = Ff,lx x
+ Ff,r x
+ Fr,l x
+ Fr,r , (4.18)
a+b

h
2∆F̄yz = Ff,ly y
+ Ff,r y
+ Fr,l y
+ Fr,r , (4.19)
2c
where h is the height between the vehicle’s C.G. and ground. Note Eqns. (4.18) and (4.19)
create algebraic loops between load transfer ∆F̄x/y z
and tire force Fl/c as they depend on each
other mutually. In principle, this is not a issue to NMPC formulation if we consider the
system as differential-algebraic equations (DAEs) [73]. However, doing so will generally make
the optimization more complicated to solve. Therefore, we will relax the algebraic loops by
introducing first-order lags for load transfer:

z
τL ∆Ḟx/y z
+ ∆Fx/y z
= ∆F̄x/y , (4.20)

where τL is the time constant.

Finally, the tire normal load is given by

z
Ff,l z
= F̄f,l − ∆Fxz − ∆Fyz , (4.21)

z
Ff,r z
= F̄f,r − ∆Fxz + ∆Fyz , (4.22)
z
Fr,l z
= F̄r,l + ∆Fxz − ∆Fyz , (4.23)
z
Fr,r z
= F̄r,r + ∆Fxz + ∆Fyz , (4.24)

91
where F̄∗z is the nominal load when vehicle is at rest.

4.3 Nonlinear Model Predictive Control of Trajectory Tracking

The objective of NMPC is to compute in real-time the optimal control inputs, minimizing
the discrepancy between the actual vehicle trajectory and the reference. While doing this,
the controller must also take into account the physical limitations of actuators, such as
amplitude, rate, and bandwidth. In this section, the objective function is formulated with
system dynamics and control inputs as equality and inequality constraints. We will also
discuss how to solve the formulated optimal control problem (OCP) using real-time iteration
scheme (RTI) and control parameterization, and how to incorporate actuator bandwidth
limitation using frequency shaping.

4.3.1 Formulation of the Optimal Control Problem

The vehicle model presented in the previous section is written as

ξ˙ (t) = f (ξ (t) , u (t)) , (4.25)

η (t) = Cξ (t) , (4.26)

h iT
where f (ξ (t) , u (t)) are given by Eqns. (4.1)-(4.3), (4.4)-(4.5), and (4.20), η = ẋ ψ Y X
h iT
is the control output, ξ = ẏ ẋ ψ ψ̇ Y X ∆Fx ∆Fy z z is the state vector, C =
 
0 1 0 0 0 0 0 0
 
0 0 1 0 0 0 0 0 h iT
  is the output matrix, and u = δf δr sf f sf l srl srr is
0 0 0 0 1 0 0 0
 
0 0 0 0 0 1 0 0
the control input.

92
 The OCP is then formulated over the time horizon t ∈ [t0 , tf ] as
ˆ tf
1 1
minimize ∥η (t) − ηr (t)∥2Q + ∥u (t)∥2R dt + ∥η (tf ) − ηr (tf )∥2P (4.27)
ξ(t),u(t) 2 t0 2
subject to ξ (t0 ) = ξ0 (4.28)
ξ˙ (t) = f (ξ (t) , u (t)) (4.29)
η (t) = Cξ (t) (4.30)
uL ≤ u (t) ≤ uU (4.31)
u̇L ≤ u̇ (t) ≤ u̇U . (4.32)

Equation (4.28) denotes the initial condition constraint, where ξ0 ∈ R8 is the initial state
vector. We assume ξ0 is available either through measurement or estimation. Equation (4.27)
is the objective function that penalizes the reference tracking deviation as well as the control
effort, where ∥·∥ denotes the Euclidean norm and its subscript represents the weighting
matrix. ηr ∈ R4 is the reference control output that is given by the upper-level controller.
Equations (4.31) and (4.32) are the control amplitude and rate constraints, respectively.
The subscriptions L and U denote the lower and upper bounds, respectively. In our control
architecture, it is the upper-level controller’s responsibility to generate a dynamically feasible
trajectory while avoiding obstacles. Therefore, the sole goal of the tracking controller is to
follow that reference state trajectory, and only simple box constraints on the control inputs
are considered here. Though the algorithms in this chapter can be easily extended to cases
where complex path constraints are presented.

The continuous-time OCP defined in Eqns. (4.28)-(4.32) is solved repeatedly online

at each time step k. The optimal control trajectory during t ∈ [t0 , t0 + Ts ] is then applied
to the system, where Ts is the sampling period of the NMPC. Due to the nonlinearity and
nonconvexity of the OCP, its solution can not be found trivially. We introduce numerical
methods based on RTI and control parameterization in the following to solve this problem.

4.3.2 Objective Function with Frequency Shaping

In the OCP (4.28)-(4.32), the control input is constrained by its amplitude (4.31) and rate
(4.32). This is the common way to incorporate the physical limits of the actuators in the MPC
framework. However, the time-domain description might not be the best choice in some cases.
In the case of vehicle chassis control, actuators’ characteristics are better described in the
frequency-domain since many of them, such as steering and braking, are electromechanical
systems which are often designed to work at certain frequency ranges. Furthermore, the

93
steering or slip control themselves might also possess different actuator dynamics for the
front and rear axles. To effectively incorporating different frequency characteristics of the
actuators into NMPC framework, a frequency shaping mechanism is adopted in the NMPC
formulation.

The frequency shaping technique was first introduced in the context of LQG control
[74] and later applied to automotive control in [75]. Recently, it has been used in the MPC
scheme in the field of robotic control [76]. The basic idea of frequency shaping is based on
Parseval’s theorem, which states the following relationship between the time and frequency
domain descriptions of a signal
ˆ ∞ ˆ ∞
1
T
x (t) x (t) dt = X T (−jω) X (jω) dω, (4.33)
0 2π 0

where x (t) is a square integrable signal defined on t ∈ [0, ∞), and X (jω) = F {x (t)}
represents its continuous Fourier transform.

Now we consider an infinite horizon objective function

ˆ ∞
1
J= ∥η (t) − ηr (t)∥2Q + ∥u (t)∥2R dt. (4.34)
2 0

Equation (4.34) can be convert into frequency domain using Eq. (4.33)
ˆ ∞
1
J= H T (−jω) QH (jω) + U T (−jω) RU (jω) dω, (4.35)
4π 0

where H (jω) = F {η (t) − ηr (t)}, and U (jω) = F {u (t)}. We can see that the objective
function of conventional MPC formulation corresponds to constant weighting over the whole
frequency range. In order to enable frequency shaping on the control effort, the constant
weighting matrix R is replaced by the frequency-dependent matrix R e (jω). The modified
objective function now becomes
ˆ ∞
1
Jf = H T (−jω) QH (jω) + U T (−jω) Rf (jω) U (jω) dω, (4.36)
4π 0

with R
e (ω) defined as
   
T
Rf,1 (−jω) Rf,1 (jω)
 ..   .. 
Rf (jω) := 
 . R
  . ,

T
Rf,n u
(−jω) Rf,nu (jω)

94
where Rf,1 (jω) , . . . , Rf,nu (jω) are the weightings that shape the frequency response of each
control input separately. Apply Parseval’s theorem again to Eq. (4.34), we can transform the
frequency domain objective function back into time domain
ˆ ∞
1
Jf = ∥η (t) − ηr (t)∥2Q + ∥uf (t)∥2R dt, (4.37)
2 0

where uf (t) is the filtered control input given as

  
Rf,1 (jω)

 

...

uf (t) = F −1  (4.38)

 U (jω) .
   
 
 Rf,nu (jω) 

We can see that instead of penalizing the control inputs directly, the frequency-shaped
objective function takes into account the filtered ones. Thus, by adjusting the diagonal
elements of the weighting function Rf , we are able to shape the frequency response of each
control input independently.

To accommodate the new objective function into the NMPC formulation, we first
approximate the infinite horizon function to a finite one
ˆ tf
1 1
Jf ∼
= ∥η (t) − ηr (t)∥2Q + ∥uf (t)∥2R dt + ∥η (tf ) − ηr (tf )∥2P . (4.39)
2 t0 2

Here the cost during t ∈ [tf , ∞) is lumped into a terminal term 12 ∥η (tf ) − ηr (tf )∥2P . In
practice, this approximation will have little impact on control performance if the weighting P
is chosen properly. It can even be neglected if the prediction horizon is long enough. In this
dissertation, we use P = kQ, where k is a scaling factor to ensure the stability and tuned by
trial and error based on the simulation results. We then augment the prediction model with
the following dynamics
żR = AR zR + BR u, (4.40)

uf = CR zR + DR u, (4.41)

which is the state-space representation of the frequency-dependent weighting Rf . The

dimension of the state zR is the sum of the orders of Rf,1,...nu (jω). The matrices AR , BR ,
CR , and DR are obtained by the Matlab function tf2ss.

95
 In summary, the frequency-shaped NMPC is formulated as
ˆ tf
1 1
minimize ∥ηa (t) − ηa,r (t)∥2Qf + ∥u (t)∥2Rf + 2ξaT (t) Su (t) dt + ∥η (tf ) − ηr (tf )∥2P
ξ(t),u(t) 2 t0 2
(4.42)
subject to ξa (t0 ) = ξa,0 (4.43)
ξ˙a (t) = fa (ξa (t) , u (t)) (4.44)
ηa (t) = Ca ξa (t) (4.45)
uL ≤ u (t) ≤ uU (4.46)
u̇L ≤ u̇ (t) ≤ u̇U , (4.47)
" # " #
Q 0
where Qf := , Rf := DR T
RDR and S := are the augmented
CRT RCR CRT RDR
" # " #
ξ η
weighting matrices, ξa := is the augmented states, ηa := is the augmented outputs,
zR zR
" #
f (ξ (t) , u (t)) h i
fa (ξa (t) , u (t)) := is the augmented dynamics, and Ca := C I is the
AR zR + BR u
augmented output matrix.

We can see that the frequency-shaped NMPC includes an additional cross term in the
objective function as well as augmented filter states and dynamics in the constraints. These
impose no fundamental difficulty in solving the NMPC. They do introduce small computation
overhead because of the increase of problem dimension. However, since the frequency shaping
filter is linear time-invariant and decoupled from the nonlinear plant, the additional constraint
relating to it can be fulfilled easily in the optimization routine with little effort.

Finally, we want to stress some differences between the frequency shaping technique
and actuator model embedding method. The actuator model embedding is a very common
approach to incorporate the actuator dynamics into MPC formulation. The basic idea is to
augment the original plant with actuator models. In this way, the input to the plant is the
output of the actuator model and the input command to the actuator model is treated as the
decision variable of OCP. The difference between these two schemes is depicted in Fig. 4.7.
In actuator embedding method, the actuator model is connected to the plant in series. While
for frequency shaping, the filter and plant are positioned in parallel. This decoupled feature
not only benefits numerical computation, but also provides more freedom in controller tuning.
Unlike the actuator model, the shaping filter can be designed freely according to the control

96
 (a) Actuator model embedding (b) Frequency shaping

Figure 4.7: Comparison between actuator embedding and frequency shaping methods

objective, regardless of the actual physical characteristics of the actuators. Moreover, the
frequency shaping method can be seamlessly applied to state weighting, while actuator model
embedding only works on control input.

Together with the conventional time domain constraint such as amplitude and rate,
the frequency shaped objective function allows us to more comprehensively represent the
dynamics of low-level actuators.

4.3.3 Fast NMPC based on RTI and Control Parameterization

Numerical methods for solving the continuous-time OCP can be generally classified into three
categories: (1) dynamic programming; (2) indirect approaches; and (3) direct approaches [77].
Although being able to solve the OCP up to global optimality, dynamic programming suffers
seriously from the so-called “curse of dimensionality”, meaning the computational complexity
increases exponentially in the dimension of the system states [78]. Therefore, it is rarely used
in practice for real-time nonlinear systems. The more commonly used methods for solving the
continuous-time OCP are the indirect and direct approaches. The former one involves solving
a boundary value problem that stems from the first-order necessary conditions of optimality.
Modern BVP solvers have been proposed to treat this problem effectively [79]. However, it
is still generally difficult to handle inequality constraints and singular arcs when using this
method. On the other hand, thanks to the great advances in nonlinear programming (NLP)
algorithms in recent years, direct approaches have become more and more popular. In the
direct methods, the continuous-time system is transcribed into discrete-time system using
some numerical method, which entails a finite-dimensional NLP. Then, the resulted NLP is
solved by a numerical solver, which is able to handle inequality path constraints nicely.

There exist two distinct types of direct approaches, namely the sequential method and
the simultaneous method. The sequential method, also known as single shooting, uses control
inputs as decision variables together with a numerical integrator to simulate the whole state

97
trajectory. NLP solver is then employed to determine the optimal control input sequence.
Since the control inputs are the only decision variables, the NLP form sequential method is
usually small scale. However, the drawback of the sequential method is that the nonlinearity
of the dynamics also propagates via the simulation process, which makes the solution very
sensitive to the initial guess. The simultaneous method provides a remedy to this issue.
Simultaneous method itself consists of two approaches, i.e. direct multiple shooting and
direct collocation [77]. Direct multiple shooting divides the whole time horizon into multiple
short segments, and simulates the system separately inside each segment using an embedded
integrator. The continuity of dynamics between adjacent segments is enforced by equality
constraints on the system states. In this way, the nonlinearity is confined to each segment,
rendering the NLP easy to treat. Direct collocation takes a step further, discretizing the
control and state variables completely. The state equations are often discretized based on an
orthogonal collocation method using polynomial approximations. Direct collocation is fully
simultaneous and requires no nested integrator. Both direct methods entail large-scale NLPs
but have the advantages of handling unstable systems. The choice between these two methods
should be made by considering the problem’s nature as well as the solver available. In this
chapter, direct multiple shooting is adopted for solving the tracking NMPC. In the next
chapter, we will discuss the application of direct collocation to the motion planing NMPC.

Using direct multiple shooting, the OCP of Eqns. (4.42)-(4.47) is transcribed as

N −1
Ts X
1
minimize
s0 ,s1 ,...,sN ,
∥ηi − ηr,i ∥2Q + ∥ui ∥2R + 2sTi Sui + ∥ηN − ηr,N ∥2P (4.48)
u0 ,u1 ,...,uN −1
2 i=0 2
subject to ξ0 − s0 = 0 (4.49)
si+1 − Fi (si , ui ) = 0, ∀i ∈ {0, . . . , N − 1} (4.50)
ηi = Csi (4.51)
uL ≤ ui ≤ uU , ∀i ∈ {0, . . . , N − 1} (4.52)
∆uL ≤ ui − ui−1 ≤ ∆uU , ∀i ∈ {1, . . . , N − 1} . (4.53)

For the clarity of presentation, we have omitted the subscription a and f in the states and
weighting matrix. Equations (4.48)-(4.53) is a finite-dimensional NLP whose decision variables
are system states si as well as control inputs ui . The time horizon [t0 , tf ] is divided into
t
N = Tfs segments of equal length defined as [ti , ti+1 ] for i = 0, . . . , N − 1. The control inputs
ui is piecewise constant discretization of its continuous counterpart over these segments, i.e.
u (t) = ui for t ∈ [ti , ti+1 ]. Equation (4.48) is discretized objective function, which is simply
the Riemann sum of the original integration. Equation (4.49) is the initial value constraint. In

98
principle, this constraint can be easily eliminated from the NLP and results in one less decision
variable. However, it is kept here explicitly to achieve the feature of the tangential predictor,
also called in the literature, the initial value embedding [80]. Equation (4.53) is the constraint
on the control input change. Equation (4.50) is the continuity constraint of system dynamics,
where Fi (si , ui ) is a simulator based on some numerical integration method which outputs the
states at the next time step. Depending on the system considered, many numerical methods
can be used, such as matrix exponential, Runge-Kutta (RK) methods, and multistep methods.
Moreover, multiple integration steps can be implemented in one discrete interval to achieve
higher accuracy if necessary. In our case, since the system dynamics consists of both the
nonlinear plant and linear filters, we use two different integration schemes. For the nonlinear
plant model, we use the popular 4th-order RK scheme to construct a simulator, which is
described by Algorithm 4.1. For the linear filter, the discrete-time dynamics is obtained via

Algorithm 4.1 Computing Fi (si , ui ) using 4th-order RK scheme

Input: si , ui , Ni , Ts
System dynamics: f
Output: s+ := Fi (si , ui )
1: s+ ← si
Ts
2: h ← N
i
3: for j = 1, . . . , Ni do
4: k1 ← f (s+ , ui )
5: k2 ← f (s+ + h2 k1 , ui )
6: k3 ← f (s+ + h2 k2 , ui )
7: k4 ← f (s+ + hk3 , ui )
8: s+ ← s+ + h6 (k1 + 2k2 + 2k3 + k4 )
9: end for

matrix exponential.
zR,k+1 = Ad zR,k + Bd uk , (4.54)

where
Ad = eAR Ts , (4.55)
ˆ Ts
Bd = eAR (Ts −τ ) BR dτ. (4.56)
0

In theory, the NLP of (4.48)-(4.53) can be solved by any general-purpose NLP slover,
e,g. Matlab’s fmincon. From a practical point of view, however, the numerical algorithm for
solving the NLPs of NMPC must be tailored for real-time applications. The development
of fast NMPC has been a hot topic in the research community, and different methods have

99
been proposed. There are several possible ways to classify these methods. Here, we divide
them into three categories: the first one relays on specialized optimization solvers exploiting
the problem structure directly, the second one reformulates the problem’s structure, and the
third one solves only sub-optimal solutions.

The examples that belong to the first category include HPIPM, OOQP, FORCES PRO,
and so on [81–83]. These tailored solvers exploit the particular structure of the MPC problems,
usually by the interior point method, to achieve excellent performance. The development of
numerical solvers is beyond the scope of this study, and we will not discuss further on this
topic.

The second category of methods manipulates the formulations of NLP to reduce the
computational complexity. They are useful when sophisticated solvers are not available due
to hardware limitations or cost. Some common approaches belonging to the second category
can be found in the context of linear MPC control. For example, the control horizon can be
chosen shorter than the prediction horizon by fixing the last control input until the end of
the prediction horizon [84]. In this way, the number of decision variables is reduced. In the
same spirit, the move blocking technique is also used to reduce the problem dimension by
constraining the control input to be constant over several discretization steps [85]. Another
similar method is so-called nonuniform grid scheme, where the system is discretized with
varying step sizes [86]. Typically, the discretization grid is dense at the beginning and becomes
more sparse towards the end of the prediction horizon. By this way, fewer decision variables
is required while remaining a relatively long prediction horizon.

While the above methods are based on the idea of reducing the optimization scale and
hoping it can be solved in one sample period, the third kind of method takes a completely
different route. These methods are called time-distributed optimization (TDO) for NMPC [87].
Instead of seeking a fully converged solution to the NLP in each time step, an approximated
solution is obtained by using only limited iterations of the optimization routine, and this
solution is consistently improved over the whole operation duration of the system. In such
a way, a much faster sampling time can be used since the controller doesn’t need to wait
for the NLP to converge, and the feedback can be updated more frequently to cope with
model mismatch or external disturbance. Several methods have been proposed based on
TDO [88–90]. The most notable one among them is the real-time iteration scheme [91].

In this study, a fast NMPC algorithm is developed incorporating the distinctive features
of the second and third types of approaches. In the proposed algorithm, the RTI with

100
 Figure 4.8: Fast NMPC algorithm timeline

condensing is used to enable fast sampling rate, then control parameterization technique is
applied on top of the condensed QP to further reduce the optimization complexity. Figure
4.8 shows the timeline of the algorithm in one sample period. The algorithm is divided
into a feedback phase and a preparation phase. The feedback phase starts at tk when the
measurement/estimation of current states ξk,0 is available. In the feedback phase, a dense but
small-scale QP is formulated by a second condensing, which is then solved quickly by a QP
solver to obtain the control input uk,0 . This control input is sent to the plant immediately
to achieve a short feedback delay. For the rest of the sample period, a preparation phase is
conducted to prepare the data needed for the next sample. Firstly, the solution of dense QP
is expanded to full scale to generate the initial guess for the QP to be solved in the next
sample period. Most of the data needed to construct this QP is already available even before
the new state feedback comes in. The scale of this sparse QP is then partially reduced by
an efficient condensing procedure, and further reduced by control parameterization. At tk+1 ,
feedback phase is repeated again to formulate a fully reduced QP when the new state ξk+1,0
arrives. The major steps of the algorithm are explained in detail in the following.

4.3.3.1 Sparse QP Construction

At each sample step, a sparse QP is obtained from the original NLP (4.48)-(4.53) based on
the sequential quadratic programming (SQP) method. Due to the partial separability of the

101
Lagrangian function, the QP problem can be written as
 " #T #T 
−1
N
" # "
 1 ∆si Hi ∆si + ∆si gi  + 1 ∆sTN P ∆sN + ∆sTN gN
X
minimize (4.57)
∆si ,∆ui
i=0
2 ∆ui ∆ui ∆ui 2

subject to ξ0 − sg0 − ∆s0 = 0 (4.58)

ri + ∆si+1 − Ai ∆si − Bi ∆ui = 0, ∀i ∈ {0, . . . , N − 1} (4.59)
uL ≤ ugi + ∆ui ≤ uU , ∀i ∈ {0, . . . , N − 1} (4.60)
∆uL ≤ ugi − ugi−1 + ∆ui − ∆ui−1 ≤ ∆uU , ∀i ∈ {1, . . . , N − 1} (4.61)

where ∆si = si − sgi , i = 0, . . . , N and ∆ui = ui − ugi , i = 0, . . . , N − 1 are the full Newton
steps to be computed by the QP. sgi and ugi are the initial guesses of the QP. Hi is the ith
diagonal block of the Hessian of the original NLP’s Lagrangian. gi is the gradient of the
h iT
original objective function with respect to ∆si ∆ui . Equations (4.58)-(4.61) are the
T T

original constraints linearized at the initial guess. ri = sgi+1 − Fi (sgi , ugi ) is the residue of the
continuity constraint. Ai and Bi are the Jacobian matrices of the dynamics to ∆si and ∆ui ,
respectively.

The initial guess of QP takes the values of the shifted solution at the previous time step,
also known as warm-start:
sgi = s̃i+1 , i = 0, . . . N − 1, (4.62)

ugi = ũi+1 , i = 0, . . . N − 2, (4.63)

sgN = Fi sgn−1 , ugi−1 , (4.64)

where the ∼ denotes the variables at previous sample time. s̃i = s̃gi + ∆s̃i and ũi = ũgi + ∆ũi
are the previous initial guess updated by the full Newton step. To avoid the expensive
calculation of exact Hessian of Lagrangian, we use Gauss-Newton Hessian approximation [92].
In particular, since the original objective function is already in the linear quadratic form, the
diagonal block of Hessian can be simply written as
" #
C T QC S
Hi = , i = 0, . . . N − 1, (4.65)
S R

which is a constant block diagonal matrix.

102
 The gradient gi is given as
" #" #
CT Q S Csgi − ηr,i (ξ0 )
gi = , i = 0, . . . N − 1, (4.66)
S R ugi

gN = C T P (CsgN − ηr,N (ξ0 )) . (4.67)

When writing Eqns. (4.66) and (4.67),we stress that the reference input ηr,i , i = 0, . . . , N
are the functions of measured/estimated system states. To be more specific, the reference
input ηr,i is obtained by interpolation based on the trajectory given by the high-level planner
and the latest system states. Therefore, the computation of gi must wait until the new
states arrive, causing an additional delay in the feedback phase. To release this overhead, we
replace ξ0 in Eqns. (4.66) and (4.67) with the previous QP solution at the first time step,
i.e. se1 , so that gi , i = 0, . . . , N can be updated during the preparation phase without the
states feedback. The effect of this approximation will be small if the actual plant tracks the
reference trajectory closely.

4.3.3.2 The First Condensing

With Guass-Newton approximation, Equations (4.57)-(4.61) describe a convex QP with a

special structure. To see it more clearly, the QP is rewritten as

1 1
minimize ∆wsT T s ∆ws + ∆wuT T u ∆wu + ∆wsT T c ∆wu + ∆wsT τ s + ∆wsT τ u (4.68)
∆ws ,∆wu 2 2
subject to A∆ws = B∆wu + r (4.69)
E∆wu ≤ e (4.70)

h iT
where we partition the decision variables into ∆ws := ∆sT1 ∆sT2 · · · ∆sTN and ∆wu :=
h iT
∆s0 ∆u0 · · · ∆uN −1 . The weighting matrices and gradients in the objective function
T T T

are defined as
 
C T QC
C T QC
 
s s s s
 
T := diag (T1 , T2 , . . . , TN ) = 

...
,

 
CT P C

103
    
T0u T0,1
u
C T QC S
 u
T1,0 T1u
  T
  S R

u

T :=  ..
=
..
,

 .  
  . 

TNu R
   
0 0 T1c 0 0 S
.
 .. ..
.   ... ..
.
  

T c :=  = ,
TNc −1   S
   

0 ··· 0 0 ··· 0
   
τ1s C T Q (Csg1 − ηr,1 ) + Sug1
 τ2   C Q (Csg2 − ηr,2 ) + Sug2
 s  T 
s

τ :=  ..  =  ..
  ,
 .   .


τNs C T P (CsgN − ηr,N ) + SugN
and    
τ0u C T Q (Csg0 − ηr,0 ) + Sug0
S (Csg0 − ηr,0 ) + Rug0
 u 
 τ1  

u

 ..  = 
τ :=  ..
  .
 .   .


g g


τNu S CsN −1 − ηr,N −1 + RuN −1
Equation (4.69) describes the relationship between ∆ws and ∆wu , which is derived form the
continuity constraint Eq. (4.59) with the definitions
 
I
−A1 I
 

A :=  .. ..
,

 . . 

−AN −1 I
 
A0 B0
B1
 
 
B :=  ..
,

 . 

BN −1
and  
r0
r1
 
 
r := 
 ..
.
.

 
rN −1

104
Equation (4.70) assembles the remaining initial condition constraints and control effort
constraints, where the matrices E and e can be trivially obtained from Eqns. (4.58), (4.60),
and (4.61).

We can see that the scale of QP can be greatly reduced by eliminating ∆ws using the
linear relationship of Eq. (4.69), due to the fact that A is always invertible. The so-called
condensing step yields the following dense QP

1
minimize ∆wuT Tc ∆wu + ∆wuT τc (4.71)
∆w2 2
E∆wu ≤ e (4.72)

where  
T
T0,0 · · · TN,0
 . ... .. 
. u T s
(4.73)
Tc :=  . . 
 = T + M T M,
TN,0 · · · TN,N
 
τ0
 . 
.  u T s s
(4.74)
 .  = τ + M (τ + T m) ,
τc := 
τN
 
M1,1 M1,2
 M2,1 M2,2 M2,3
 
 = A−1 B, (4.75)

 ..
M :=  .. .. ...
 . . .


MN,1 MN,2 MN,3 · · · MN,N +1
 
m1
 . 
.  −1
(4.76)
 .  = A r.
m := 
mN
Furthermore, thanks to the sparse structure of A and B, the condensing can be done efficiently
with complexity of O (N 2 ) using block-wise forward/backward substitution. Following the
similar idea in [93], the condensing procedure is given in Algorithms 4.2-4.5.

4.3.3.3 Control Parameterization

After the first condensing, the dimension of QP is reduced form N (nu + nx ) + nx to N nu + nx .

The scale of QP grows linearly with respect to the prediction horizon, which means the
QP can become quite large if the prediction horizon is long. However, a long prediction

105
Algorithm 4.2 Computing M
Input: A, B
Output: M
1: M1,1 ← A0
2: for k = 2, . . . , N do
3: Mk,1 ← Ak−1 Mk−1,1
4: end for
5: for i = 2, . . . , N + 1 do
6: Mi−1,i ← Bi−2
7: for k = i − 1, . . . , N do
8: Mk,i ← Ak−1 Mk−1,i
9: end for
10: end for

Algorithm 4.3 Computing m

Input: A, r
Output: m
1: m1 ← r0
2: for i = 2, . . . , N do
3: mi ← Ai−1 mi−1 + ri−1
4: end for

horizon is usually necessary in vehicle motion control. The long prediction horizon helps to
stabilize the system. It is also crucial to ensure a feasible solution when the vehicle encounters
obstacles or sharp turn. Therefore, in order to utilize a long prediction horizon without
complicating the optimization, the control parameterization method is used in our study.
Control parameterization for fast MPC control was first introduced in [94], and was adopted
in several real-world applications [95–98]. Open source toolbox pNMPC is also developed
by [99], which formulates the OPC in single shooting method and relies on derivative-free
optimization. Here, we embed control parameterization into multiple shooting RTI framework.

Control parameterization can be seen as a more general form of move blocking. Unlike
move blocking which assumes constant control input over each blocking segment, control
parameterization utilizes customized functions to characterize the control input. In general,
the parameterization of a piece-wise constant control profile can be defined as a map

uk = Upwc (k, p) , (4.77)

where p ∈ P ⊆ Rnp is a set of parameters in some admissible space, which is usually

low-dimensional compared to control input. This map is then substituted into the original

106
Algorithm 4.4 Computing Tc
Input: M , T u , T s , T c
Output: Tc
s
1: WN,1 ← TN MN,1
2: for k = N, . . . , 2 do
T
3: Tk,0 ← Bk−1 Wk,1
T s
4: Wk−1,1 ← Ak−1 Wk,1 + Tk−1 Mk−1,1
5: end for
6: T1,0 ← B0T W1,1
7: T0,0 ← T0u + AT0 W1,1
8: for i = 1, . . . , N do
9: WN,i+1 ← TNs MN,i+1
10: for k = N, . . . , i + 1 do
T
11: Tk,i ← Bk−1 Wk,i+1
12: Wk−1,i+1 ← ATk−1 Wk,i+1 + Tk−1
s
Mk−1,i+1
13: end for
T
14: Ti,i ← Bi−1 Wi,i+1 + Tiu
15: end for
16: for k = 1, . . . , N − 1 do
17: for i = 1, . . . , k + 1 do
18: Tk+1,i ← Tk+1,i + TkcT Mk,i
19: end for
20: end for
u
21: T1,0 ← T1,0 + T1,0

Algorithm 4.5 Computing τc

Input: M , m, T s , T c , τ u , τ s
Output: τc
s
1: WN ← τN + TNs mN
2: for i = N, . . . , 2 do
T
3: τi ← Bi−1 Wi + τiu + TicT mi
4: Wi−1 ← ATi−1 Wi + τi−1 s s
+ Ti−1 mi−1
5: end for
6: τ1 ← B0T W1 + τ1u + T1cT m1
7: τ0 ← AT0 W1 + τ0u

107
 Figure 4.9: Linear interpolating parameterization with np = 4

formulation of OCP to yield a reduced NLP with decision variables of p. The parameterization
schemes are very diverse depending on the problems. It is shown that the impact on control
performance will be minor while faster computation can be achieved, if the parameterization
scheme is chosen carefully [94].

To accommodate the control parameterization method for the RTI scheme, we choose to
parameterize the Newton steps of condensed QP, i.e. ∆uk , instead of the direct control input
uk . A linear interpolating parameterization is used, as illustrated in Fig. 4.9. A number of
control inputs at some intermediate sample instants within the prediction horizon are chosen
as parameters
h iT h iT
p := pT0 pT1 · · · pTnp−1 = ∆uTN0 ∆uTN1 · · · ∆uTNnp −1 ,

where N0 := 0 and Ni ∈ {1, 2, . . . , N } for i = 1, 2, . . . , np − 1. We also define Nnp := N .

The controls in between these instants are obtained by the affine combination of the
predecessor and successor nodes as
 
i − Nj i − Nj
∆ui = 1 − ∆uNj + ∆uNj+1 , (4.78)
Nj+1 − Nj Nj+1 − Nj

where j = max {j ∈ {0, . . . , np } |Nj ≤ i }. The advantage of using such parameterization is

108
that the relationship between u and p can be conveniently expressed as a linear map
 
∆u0
∆u1
 
(4.79)
 
 ..  = Πu p,
.
 
 
∆uN −1

where
 
I
 
 Π1,0 Π 1,1

.. ..
 
. .
 
 
 
Π Π 
 0N −1,0 N0 −1,1 
 

 I 

 ΠN1 +1,1 ΠN1 +1,2 
. .
 
.. ..
 
 
 
ΠN2 −1,1 ΠN2 −1,2
 
 ∈ RN nu ×np nu
 
Πu := 

.. ..
 . . 

 

 I 

ΠNnp −2 +1,np −2 ΠNnp −2 +1,np −1 
 

 .. .. 

 . . 

 

 Π N np −1 −1,np −2 Π Nnp −1 −1,np −1 


 I 


 .
..


 
I

is a constant matrix and each block of it is a diagonal matrix, which renders the whole matrix
very sparse. The matrix Πu can be computed by Algorithm 4.6.

Substitute Eq. (4.79) into the condensed QP (4.71)-(4.72), we get a parameterized QP

1
minimize ∆wpT TΠ ∆wp + ∆wpT τΠ (4.80)
∆wp 2
EΠ ∆wp ≤ e (4.81)

where h iT
∆wp := ∆sT0 p T ,

109
Algorithm 4.6 Computing Π∆
Input: nu , np , N , N0 , N1 , . . . , Nnp −1
Output: Πu
1: for k = 0, . . . , np − 2 do
2: for i = Nk + 1, . . . , Nk+1 − 1 do
i−Nk
3: Πi,k+1 ← Nk+1 −Nk
I
4: Πi,k ← I − Πi,k+1
5: end for
6: ΠNk ,k ← I
7: end for
8: for i = Nnp −1 , . . . , N do
9: Πi,np −1 ← I
10: end for

" #T " #
I I
TΠ := Tc ,
Πu Πu
" #T
I
τΠ := Tc ,
Πu
and " #
I
EΠ := E .
Πu
The QP has degrees of freedom np nu + nx , which is completely independent of the length
of the prediction horizon. It should be noted that the parameterization process creates
additional computation overhead. But its effect should be small due to the sparsity of the
matrix Πu .

4.3.3.4 The Second Condensing

The computation process described until now can be performed before the arrival of state
feedback. They account for the majority of the computation load of NMPC. When the state
feedback ξ0 is actually available, we can quickly conduct a second condensing to further
reduce the QP size by eliminating ∆s0 from ∆wp . The resulted QP is given as

1 T
minimize p Tpp p + pT (τp + Tps (ξ0 − sg0 )) (4.82)
p 2
Ep p ≤ e − Es (ξ0 − sg0 ) (4.83)

110
 " # " #
Tss Tsp τs
where Tpp and Tps are the partitions of TΠ := , τp is the partition of τΠ := ,
Tps Tpp τp
" #
Es
and Ep is the partition of EΠ := .
Ep

The QP (4.82)-(4.83) is now fully condensed without any sparsity. It can be efficiently
solved by a generic QP solver. Once the solution p∗ is obtained, the control input to the
plant can be constructed immediately as u∗0 = ug0 + p∗0 .

4.3.3.5 Solution Expansion

The final step is to expand the condensed QP solution so that it can be used as the initial
guess for the QP of the next time step. The process is fairly straightforward. Since the
Gauss-Newton Hessian approximation is used, the multipliers become irrelevant, and all we
need are the primal variables. We first expand the p into ∆wp by
" #
ξ0 − sg0
∆wp = , (4.84)
p

then ∆wu is obtained as " #

I
∆wu = ∆wp , (4.85)
Πu
finally, the full Newton step is given as
" # " #
∆ws M ∆wu + m
= . (4.86)
∆wu ∆wu

4.4 Simulation Results

We implement the vehicle model in Matlab. The plant model includes 1st order transfer
functions to reflect the effect of low-level actuator control systems. The NMPC is formulated
using the symbolic framework of the open-source tool CasADi. The model and controller
parameters are listed in Appendix A. The condensing and control parameterization are
implemented in Matlab with the required derivatives provided by algorithmic differentiation
of CasADi [100]. The open-source qpOASES is used for solving the condensed QP [101]. All
simulations were performed on a desktop computer with a 3.50GHz Intel Core i5-6600K CPU
and 8GB RAM.

111
4.4.1 High µ Braking

In the high µ braking simulation, the vehicle brakes at an initial speed of 100km/h on an
asphalt road with µ = 1 until it stops. This simulation is used to evaluate if the slip ratio
profile from NMPC is realistic for level-level actuators. We use prediction step N = 20, and
sample time Ts = 25ms. The control inputs are parameterized using linear interpolation. By
some trial and error based on simulation results, we choose the parameterization points to be
Ni = [0, 1, 11, 15].

We first determine the amplitude and rate constraints on the slip ratio. Although some
literature put hard constraints on slip ratio to be within the positive slope of tire force [102],
we will relax the slip ratio constraint by allowing it to vary from 0 to -1 in order to avoid
infeasibility due to tracking error of actuator control. Since the tire nonlinearity is considered
explicitly in the NMPC, we expect the controller to be able to avoid excessive slip by itself
without explicit constraints. For the rate constraint, we set the −1 ≤ λ̇ ≤ 1 to match the
actuator limit. The derivation of rate constraint is included in Appendix B. A high-pass
frequency shaping filter Rf (s) = αα12 s+1
s+1
is applied to wheel slip inputs to avoid high-frequency
excitation, whose parameters are given in Appendix A.

In the simulations, we command the controller to track a constant target deceleration

with entry speed of 100km/h and no steering action. Figure 4.10 shows the simulation results
when the target deceleration equals 1g. Figure 4.10(a) shows the time history of vehicle speed,
deceleration, as well as the slip ratio commands. We can see that the target deceleration is
achieved, and the slip ratio command is also smooth. It is also interesting to see that the
controller put more effort on the front axle than the rear one because of the fact that the
front axle braking is more effective due to the load transfer. Figure 4.10(b) shows the tire
brake force profiles. We can see that both front and rear tires reach their peak values but
stay out of the negative slop region. It shows the controller is able to exploit the maximum
tire force while keeping the vehicle operating in stable region.

4.4.2 Low µ Braking

Now we are going to investigate the braking control performance on a snowy surface, where
µ = 0.4. In particular, we use the prediction model in the NMPC controller that corresponds
to µ = 1. Furthermore, the target deceleration is remained to be 1g despite the maximum
achievable deceleration is only 0.4g. The discrepancy between the prediction model, target

112
 (a) Vehicle responses and wheel slip inputs

(b) Tire force profiles

Figure 4.10: High µ (µ = 1) braking with target deceleration of 1g

113
and the physical limit might occur due to the difficulty on real-time road friction estimation.
We want to see how the controller performs under this discrepancy. Figure 4.11 shows
the vehicle response and wheel slip inputs of the simulation. For comparison, we present
two simulation results, where Fig. 4.11(a) corresponds to the response using NMPC with
frequency shaping and Fig. 4.11(b) is the response using NMPC without frequency shaping.
We can see from Fig. 4.11 that the vehicles achieve the similar amount of deceleration
in both cases. Although the decelerations are similar, from Fig. 4.11(a), we see that the
wheel slip command of NMPC with frequency shaping is smooth. However, as shown in Fig.
4.11(b), the control input of NMPC without frequency shaping suffers serious high frequency
chattering during the steady state, which is very undesirable for low-level actuator. The
control input frequency characteristics of the two cases can also be observed for the power
spectral density plots of Fig. 4.11(c). We can see that the control inputs of no frequency
shaping case show high frequency peak in the high frequency, while those of frequency shaping
case show consistent low density in the high frequency range.

Without frequency shaping, the only way to remove the chattering is to increase the
weighting on the control input. Figure 4.12 shows the simulation result where the control
input weighting is increased. We can see that the chattering of control input is diminished
with larger weighting. On the other hand, the deceleration also becomes smaller due to the
increased weighting on control input. The tradeoff between control performance and input
can be explained by Fig. 4.13. For the NMPC without frequency shaping, the weighting over
the different frequency is adjusted as a whole. Therefore, the low frequency performance
must be sacrificed when we increase the weighting to eliminate the high frequency chattering.
Whereas for the NMPC with frequency shaping, we have two tuning knobs to independently
adjust the weighting in the low frequency and high frequency. Therefore, the low frequency
control performance will not be compromised by large weighting in the high frequency.

4.4.3 Split µ Braking

In the split µ braking, the tires on the left hand side operate on asphalt with µ = 1, and the
right hand side tires operate on a snowy surface with µ = 0.4. It is a challenging task because
the discrepancy in road friction between left and right might create a large yaw moment
disturbance that destabilizes the vehicle. A good controller should achieve a short barking
distance while keeping the vehicle stable. For the trajectory tracking purpose, it should also
keep the lateral and yaw angle deviation as small as possible. To better examine the slip
control performance, we will not use steering control in the simulation.

114
 (a) NMPC with frequency shaping

(b) NMPC without frequency shaping

(c) Power spectral density

Figure 4.11: Low µ (µ=0.4) braking with target deceleration of 1g

115
Figure 4.12: Low µ (µ=0.4) braking with no frequency shaping and increased control input
weighting

Figure 4.13: Control input weighting

116
 Figure 4.14 shows the simulation results with target deceleration of 0.4g which is the
maximum possible deceleration on the snowy surface. We can see form Fig. 4.14(a) that the
vehicle tracks the reference path closely with very small lateral and yaw angle deviations.
Figure 4.14(b) shows that the deceleration reaches the target value and the control inputs
are smooth. Especially, the slip ratio of the high µ side is smaller than that of the low µ side
such that the longitudinal forces of two sides are balanced. From Fig. 4.14(c), we can see
that the tire forces of the right tires are saturated to achieve the maximum deceleration, and
the forces on the left tires are kept similar to those of right tires to avoid generating large
yaw moment disturbance.

We also conduct a simulation with target deceleration of 0.3g. Figure 4.15 shows the
results. We can see that the vehicle is able to achieve the desired deceleration using smooth
control inputs. And the tracking deviations are even smaller than the previous case.

4.4.4 Sine with Dwell

We have evaluated the slip control performance of NMPC in the previous two simulations.
Now we will assess the full control capability of NMPC both 4WS and 4WID. We first
perform the sine with dwell (SD) test. The SD test defined by FMVSS 126 regulation [103] is
a standard procedure used to evaluate the performance of the ESC system on vehicle lateral
stabilization. Here we utilize a similar procedure with small modifications to evaluate the
trajectory tracking control performance.

Firstly, we generate a SD trajectory using an uncontrolled vehicle. The open-loop front

wheel steering profile is shown in Fig. 4.16, which is obtained according to the standard of
FMVSS 126. Figure 4.17 shows the uncontrolled vehicle trajectory under this steering input
with an entry speed of 80km/h. Due to the lack of stability control and large steering angle,
we can clearly see the oversteer tendency towards the end of the simulation.

To evaluate the performance of our control algorithm, the open-loop trajectory is used
as a reference to the NMPC control for tracking. In the closed-loop simulation, we evaluate
two sets of configurations, one with 4WS and slip control and the other with only FWS and
slip control. Figure 4.18(a) shows the trajectory tracking performance of the controllers. We
can see that both control configurations are able to mitigate the oversteering of vehicle and
improve the vehicle stability. The vehicle trajectory is also close to the reference one in both
cases. The 4WS with wheel slip control shows superior tracking performance than the FWS.
Figure 4.18(b) shows the vehicle speed, lateral acceleration, and side slip angle comparison.

117
 (a) Vehicle path and tracking errors

(b) Vehicle speed, deceleration, and wheel slip inputs

(c) Tire force profiles

Figure 4.14: Split µ (left: µ = 1; right: µ = 0.4) braking with target deceleration of 0.4g
118
 (a) Vehicle path and tracking errors

(b) Vehicle speed, deceleration, and wheel slip inputs

(c) Tire force profiles

Figure 4.15: Split µ (left: µ = 1; right: µ = 0.4) braking with target deceleration of 0.3g
119
 Figure 4.16: Front wheel steering angle of sine with dwell simulation

Figure 4.17: Open-loop trajectory of sine with dwell simulation. The green solid line denotes
vehicle C.G. trajectory, and the black polygon denotes vehicle’s shape.

We can see that the 4WS with slip control achieves smaller vehicle speed variation and
side slip angle. The lateral accelerations in both cases are close to 0.9g which indicates the
vehicle is operating at the handling limit. Figure 4.18(c) shows the vehicle yaw rate response.
According to the standard test criteria, the yaw rate should decay to 35% and 20% of the
peak value at 2.929s and 3.679s, respectively. The simulation result shows that the controlled
vehicle satisfy these criteria in both cases.

Figure 4.19 shows the control input profiles of the two control setups. We can see
that both control setups generate smooth input profiles. For the 4WS case, rear steering is
controlled to be in-phase with the front steering to reduce the side-slip angle. In the FWS
case, the front wheel steering shows a similar profile to that of 4WS. However, due to the
lack of rear steering, the slip control becomes more active in order to generate additional yaw
moments, which also results in large vehicle speed variation.

4.4.5 High Speed Double Lane Change

In this simulation, the vehicle is required to make a double lane change (DLC) maneuver
at high speeds with different road friction conditions. For 4WS, rear-wheel is only used for
micro steering by penalizing its high frequency maneuvers using frequency shaping. For
comparison purposes, we also perform the same simulations using the vehicle with front wheel
steering (FWS) and 4WID vehicle. The DLC track is set up according to the ISO 3888-1
standard [104]. The vehicle speed in each simulation is selected as the highest one at which

120
 (a) Trajectory tracking

(b) Vehicle states

(c) Yaw rate response

Figure 4.18: Closed-loop control of sine with dwell (µ = 1)

121
Figure 4.19: Control inputs profiles of sine with dwell. The solid line denotes the response of
4WS and the dashed line denotes the response of FWS.

the maneuver can be completed within the boundary by the vehicle equipped with 4WS and
4WID. The reference path is given by functions of displacement in X direction

dy1 dy2
Yref = (1 + tanh z1 ) − (1 + tanh z2 ) ,
2 2
 
1 1.2 1 1.2
ψref = arctan dy1 − dy2 ,
cosh2 z1 dx1 cosh2 z2 dx2
where z1 = a1 (X − b1 ) − c1 , z2 = a2 (X − b2 ) − c2 . a1,2 , b1,2 , c1,2 , dy1,2 , and dx1,2 are constant
parameters which are given in the Appendix A. The reference longitudinal velocity ẋ is same
as the vehicle’s entry speed.

We first conducted the simulation on an asphalt road (µ = 1) with a vehicle entry speed
of 140km/h. The vehicle tracking performance is shown in Fig. 4.20. We can see that both
vehicles are able to finished the maneuver while remaining stable and inside the boundary.
The vehicle with 4WS shows better tracking performance with slightly smaller lateral and
yaw angle deviations. Figure 4.21 shows the vehicle speeds, lateral accelerations, and side slip
angles of the two vehicles. It can be seen that the vehicle with FWS has larger fluctuations
in speed tracking. The lateral accelerations are very similar. We can see that the maximum
lateral accelerations are approaching to 0.9g in both cases, which indicates that the vehicle
really reaches its handling limit in simulation. We can also see the vehicle with 4WS has
smaller side slip angle. The control inputs of two vehicles are shown in Fig. 4.22. We can
see that all the control inputs are smooth. For the 4WS case, the rear steering is of similar

122
Figure 4.20: DLC on asphalt at 140kph (µ=1). Top plot: the vehicle with 4WS is outlined
by black polygon, and the one with FWS is outlined by red polygon.

Figure 4.21: Vehicle state responses of DLC on asphalt (µ=1) at 140kph

123
Figure 4.22: Control input demands of DLC on asphalt (µ=1) at 140kph. The solid line
denotes 4WS, and the dashed line denotes FWS.

amplitude and phase as those of the front steering. We can also observe that the FWS vehicle
has larger control actions in slip ratios due to the absence of rear steering control. These
larger actions of slip ratio also cause more oscillations in vehicle speed.

Next, simulations are executed on a snowy surface (µ = 0.4) at an entry speed of 70km/h.
The vehicle tracking performance is shown in Fig. 4.23. Similarly to the previous simulation,
the 4WS vehicle shows slightly better tracking performance. Figure 4.24 shows the vehicle
speed, lateral accelerations, and side slip angles comparisons. Again, larger speed variations
and side slip angle are shown in the FWS vehicle, and lateral accelerations are similar. Control
inputs are smooth as shown in Fig. 4.25.

We then perform the simulation on an icy surface (µ = 0.17) at a speed of 50km/h. This
is the most challenging case, as the tire fore can become highly nonlinear and easily saturated.
The results are presented in Fig. 4.26, 4.27, and 4.28. Both vehicles complete the maneuver
successfully. We can see that there is a little chattering in front steering due to the fact
that vehicle becomes considerably more difficult to stabilize on icy surface. From the lateral
acceleration response, we can see that the vehicle has reach its physical limit with ay ∼
= 0.15g.
Nevertheless, the overall control performance is satisfactory in this extreme situation.

Finally, we check the real-time feasibility of the controller. Figure 4.29 shows computation
time for the three DLC simulations presented previously. For comparison purposes, we also
show the computation time of the NMPC without control parameterization. We can see
that the computation time of the controller with control parameterization is less than those

124
Figure 4.23: DLC on snowy surface (µ=0.4) at 70kph. Top plot: the vehicle with 4WS is
outlined by black polygon, and the one with FWS is outlined by red polygon.

Figure 4.24: Vehicle state responses of DLC on snowy surface (µ=0.4) at 70kph

125
Figure 4.25: Control input demands of DLC on snowy surface (µ=0.4) at 70kph. The solid
line denotes 4WS, and the dashed line denotes FWS.

Figure 4.26: DLC on icy surface (µ=0.17) at 50kph. Top plot: the vehicle with 4WS is
outlined by black polygon, and the one with FWS is outlined by red polygon.

126
 Figure 4.27: Vehicle state responses of DLC on icy surface (µ=0.17) at 50kph

Figure 4.28: Control input demands of DLC on icy surface (µ=0.17) at 50kph. The solid line
denotes 4WS, and the dashed line denotes FWS.

127
 (a) 140km/h on asphalt

(b) 70km/h on snowy surface

(c) 50km/h on icy surface

Figure 4.29: Computation times of DLC simulations

128
 Table 4.1: Tracking performance comparison
Road surface Asphalt Snowy Icy
Control Parameterization with without with without with without
Maximum lateral deviation (m) 0.76 0.79 0.39 0.37 0.39 0.34
Maximum angular deviation (◦ ) 5.79 4.86 2.22 2.18 2.11 2.24

without and smaller than the sampling period (25ms). Table shows the tracking performance
of the controller with and without control parameterization. We can see that there are only
minor differences in performance between the two controllers.

4.5 Conclusions

A NMPC controller is proposed for autonomous vehicle trajectory tracking using 4WS and
4WID. The controller is based on a nonlinear double track model with a Magic Formula tire
model. Frequency shaping is used to allocate control inputs for steering and slip control. A
fast NMPC algorithm is proposed based on RTI with condensing and control parameterization.
The proposed control system is tested through simulations. It is shown that the vehicle
achieves good braking distance with smooth control inputs. Sine with dwell and double lane
change simulations show that the controller is able to stabilize the vehicle at the handling
limit and achieve good tracking performance under various road frictions.

129
 Chapter 5

Motion Planning for Autonomous Vehicle based on NMPC

In this chapter, we develop a motion planning algorithm for autonomous vehicles. The
motion planning algorithm aims to generate a reference trajectory for the lower-level tracking
controller. The algorithm should generate maneuvers like braking, acceleration, lane keeping
or lane change. The trajectory should be dynamically feasible so that it can be realized by
the vehicle within physical limitations. In addition, other requirements regarding driving
safety, comfort, and traffic rules might also be enforced to the trajectory planning.

5.1 Introduction

MPC has been widely used for vehicle motion planning mainly because of its ability to
handle different constraints systematically and explicitly. In [105], a trajectory generation
algorithm for unmanned ground vehicles is proposed based on MPC. Information on obstacles
is incorporated online into the NMPC framework and then solved online with nonlinear
programming. Lee et al. [106] presented a scenario-based MPC trajectory planning algorithm
that consists of spatial planning with embedded temporal optimization that leverages waypoint
information on the road. Trajectory candidates are generated by solving spatial and temporal
optimization problems for each feasible scenario, and then integrated together. In [107], a
collision imminent steering system is developed using nonlinear model predictive control
to perform a lane change at high speed in a highway environment. Two formulations are
presented based on minimum distance formulation and minimum slip formulation, respectively.
Taherian et al. [108] proposed an integrated trajectory planning with collision avoidance
based on MPC. A torque vectoring controller is also designed for lateral-yaw stabilization
under low friction and crosswinds on highways.The steering input computed from the MPC
is used to steer the vehicle along the reference trajectory while the torque vectoring controller
provides additional lateral-yaw stability. In [109], authors presented a vehicle motion planning
and control framework based on MPC. Collision avoidance is ensured by transforming the

130
predicted motion and uncertainty of other road users into constraints within the MPC
formulation. Hajiloo et al. [110] proposed a emergency collision avoidance system using
steering and differential braking. MPC is used for developing a combined path planning and
tracking controller with a hierarchical structure that considers collision avoidance, vehicle
stability, and path tracking. Quirynen et al. [111] presents a hierarchical control framework
for obstacle avoidance in autonomous driving systems. A particle-filtering is used for low-rate,
long-term motion planning. A NMPC-based vehicle controller is used to track the reference
trajectory and avoid obstacles. In [112], a two-stage nonlinear nonconvex control approach for
autonomous vehicle driving is developed. An outer-loop nonlinear model predictive control
is adopted for generating the collision-free trajectory with the resultant input based on a
simplified vehicle model. The inner loop is a simple linear feedback controller based on an
optimal preview distance.

In this chapter, we will present the application of NMPC to autonomous vehicle motion
planning. Note that the underlying control framework is the same as that used in the previous
chapter. Nevertheless, the algorithm in this chapter differs significantly from the previous one
in terms of problem formulation as well as numerical method. Furthermore, we incorporate
control barrier function into the NMPC scheme for safety guarantee.

The rest of this chapter is organized as follows. In Section 5.2, the vehicle model for
motion planning is described. Section 5.3 gives the definitions of motion planning problem.
The NMPC formulation and numerical method are discussed in Section 5.4. In Section 5.5,
we present simulation results of proposed motion planning algorithm. Finally, Section 5.6
concludes the chapter.

5.2 Vehicle Modeling

In order for the NMPC to generate a feasible trajectory, the vehicle model must be able
to reflect essential physical properties of the real system. On the other hand, the vehicle
model should not be over-complicated, since the optimization of motion planning is often
computationally expensive due to the various kinds of complex path constraints. In this work,
a spatial domain nonlinear point-mass model is used.

5.2.1 Point-Mass Vehicle Model

Figure 5.1 shows the schematic of the model used in this work. The vehicle is represented by
a particle, and its motion is described with respect to the road adherent Frenet frame. n

131
 Figure 5.1: Schematic of point mass model

and t are the two axis of the Frenet coordinate which respectively point to the normal and
tangential direction of the reference path. This reference path could represent the center line
of the lane or any reference path from the decision making module at a high level. The origin
of the Frenet coordinate coincides with the projection of the vehicle along the n direction to
the reference path.

The equations of motion of the vehicle are given by

1
ṡ = v cos ψe , (5.1)
1 − ye κ (s)

ẏe = v sin ψe , (5.2)

κ (s)
ψ̇e = ω − v cos ψe , (5.3)
1 − ye κ (s)
v̇ = a, (5.4)

where v is the vehicle speed, a is the vehicle acceleration. ψe is the angular deviation between
v and t axis, ye is the lateral deviation of vehicle to the reference path along n axis, s is the
arc length traveled by the vehicle’s projection on the reference path. ω is the angular speed
of vehicle’s heading with respect to global axis, and κ is the curvature of the reference path,
which is a function of s. Note that the model implies 1 − ye κ (s) ̸= 0, which is valid under
normal driving conditions where ye ≪ 1/κ.

132
 Equations (5.1)-(5.4) are compactly written as

ξ˙ (t) = fκ(s) (ξ (t) , u (t)) , (5.5)

h iT h iT
where ξ := ye ψe v s and u := ω a . The subscript κ indicates that the curvature
enters the dynamics as a varying parameter. We assume the state of system and road
curvature are known from measurement or estimation.

5.2.2 Spatial Transformation

In this section, we perform a spatial transformation to the time-dependent dynamics (5.5).

The procedure is similar to what is given in [113]. This reformulation changes the independent
variable of the model from time t to arc length s. Assuming ṡ > 0, i.e. vehicle is always
driving forward, we have the following relationship

dξ dξ dt ξ˙
ξ ′ := = = . (5.6)
ds dt ds ṡ

We will use the superscript ′ to denote the derivative of a variable with respect to s in the
rest of the chapter. Apply Eq. (5.6) to (5.5), we obtain the vehicle model in the spatial
domain as
fκ(s) (ξ (t) , u (t))
ξ ′ (s) = s
:= fκ(s) (ξ (s) , u (s)) . (5.7)
ṡ
Now the dynamics of state variables are functions of s. Later, we will see that the spatial
formulation renders road boundaries into simple box constraints on the state. It also makes the
parameter κ an explicit function of the independent variable, thus simplifying the simulation
of the system. Note that the independent spatial variable s now becomes a clock state with
dynamics s′ = 1 embedded in the system (5.7).

5.2.3 Convert Cartesian Coordinates to Frenet Coordinates

The sensor measurements of the vehicle are usually given by the global Cartesian coordi-
nates. In addition, the trajectory tracking control system is also formulated in Cartesian
coordinate frame. Therefore, the conversion between Cartesian coordinates and Frenet coor-
dinates must be performed such that the tracking and planning control systems can work
together. Assuming that the reference path is represented by a discrete set of coordinates,
h iT h iT
Xr,i Yr,i sr,i κr,i for i = 0, 1, . . . , Nr , and the vehicle’s coordinates are X Y ψ ,

133
we want to find out the index corresponding to the closest point to the vehicle,

p = arg min (Xr,i − X)2 + (Yr,i − Y )2 . (5.8)

i=0,1,...,Nr

Then we can derive the vehicle’s coordinates in Frenet frame as

s = sr,p , (5.9)

and q
ye = (Xr,p − X)2 + (Yr,p − Y )2 . (5.10)

For the angle coordinate, we first calculate the reference path’s heading angle by the integration

ˆsr,i
ψr,i = κr ds + ψ0 , (5.11)
0

where ψ0 is the initial heading angle of the reference path. This integration can be computed
numerically using the trapezoidal method. Then the angle coordinate of the vehicle can be
expressed as
Ẏ
ψe = arctan − ψr,p ∼ = ψ − ψr,p , (5.12)
Ẋ
where the approximation is valid if the side slip angle β ∼= 0.

Finally, the vehicle speed is given by

p
v= Ẋ 2 + Ẏ 2 . (5.13)

5.3 Motion Planning Problem Definition

In this section, we define the cost function and various constraints, except for collision
avoidance, for the motion planning problem. They will be used later to formulate the NMPC.
The safety-critical collision avoidance constraint will be discussed in detail in the next section.

5.3.1 Objective Function

A quadratic objective function is defined to minimize the deviation from the reference path
and control input.
ˆ sf
1
J= q1 ye2 (s) + q2 ψe2 (s) + q3 (v (s) − vr (s))2 + ∥u (s)∥2R ds, (5.14)
2 s0

134
where q1 , q2 , and q3 are the weighting factors for lateral deviation, angular deviation, and
velocity tracking error, respectively. R is the weighting matrix for control inputs. Note that
the integral is defined form s0 to sf in the spatial domain instead of time domain making the
prediction horizon equals sf − s0 in further distance.

5.3.2 Control Input Constraints

The inputs to the vehicle model are yaw rate ω and acceleration a. These two inputs are
physically bounded by the maximum yaw moment and driving/braking force available to the
vehicle, which depends on both the actuator capability and road friction. The constraints on
actuator capability are expressed by box constraints

ωL ≤ ω (s) ≤ ωU , (5.15)

aL ≤ a (s) ≤ aU . (5.16)

And the constraints due to road friction are represented by the friction circle

(v (s) ω (s))2 + a (s)2 ≤ (µg)2 , (5.17)

where µ is the road friction coefficient. This is a nonlinear constraint with control inputs and
state coupled.

5.3.3 State Constraints

The vehicle’s lateral deviation should stay within the road boundary. Thanks to the spatial
formation, this can be expressed as simple box constraints

ye,L (s) ≤ ye (s) ≤ ye,U (s) , (5.18)

where ye,L (s) and ye,U (s) are road boundaries which are functions of the spatial variable
s obtained by curve fitting based on sensor data. If temporal formation is used instead,
inequality (5.18) will be a nonlinear constraint since the variable s itself is a nonlinear function
of time. Similarly, the yaw angle error is constrained by

ψe,L (s) ≤ ψe (s) ≤ ψe,U (s) . (5.19)

135
The vehicle speed limit is enforced by the constraint

vL (s) ≤ v (s) ≤ vU (s) . (5.20)

where vL and vU may vary with different road segments. Sometimes it is also useful to
constrain the maximum lateral acceleration of the vehicle for driving comfort considerations

ay,L (s) ≤ v (s) ω (s) ≤ ay,U (s) . (5.21)

Note that it is also an input/state coupled constraint.

5.3.4 Waypoint Constraints

The waypoint constraint enforces the vehicle passing by an intermediate point on the route at
a particular time. This type of constraint is useful when there exists higher-level algorithm,
e.g. A∗ or RRT, providing waypoints for the vehicle to follow. It also allows us to handle
tasks with time-scheduling requirements, for example, autonomous intersection management.
The spatial form enables easy formulation of waypoint constraints as t is now a function of
the space coordinates. Recall Eq. (5.1), we can write the dynamics of time in the spatial
domain as
dt 1 1 − ye κ (s)
t′ := = = . (5.22)
ds ṡ v cos ψe
The time dynamics (5.22) is augmented to the point-mass model (5.7), resulting in an
additional state t. Then the waypoint constraint can be expressed as

twp,L ≤ t (swp ) ≤ twp,U , (5.23)

twp,L ≤ ye (swp ) ≤ ywp,U , (5.24)

ψwp,L ≤ ψe (swp ) ≤ ψwp,U , (5.25)

vwp,L ≤ v (swp ) ≤ vwp,U , (5.26)

where swp ∈ Swp is the spatial coordinate from a set of waypoints. Equations (5.23) and
(5.24) respectively defines the constraints of time and lateral displacement of the waypoint.
Equations (5.25) and (5.26) can be used to constrain the heading angle and speed of the
vehicle when traversing waypoints, if necessary.

So far, we have presented the major constraints that motion planning might encounter,
except for collision avoidance. In the next section, we will discuss how to formulate this

136
constraint using control barrier function [114].

5.4 Collision Avoidance Constraints based on Control Barrier Function

We present a new formulation for collision avoidance in motion planning NMPC in this
section. The constraint is based on the recently developed theory of control barrier function.

5.4.1 Conventional Formulation of Collision Avoidance Constraint

Collision avoidance constraints are used to prevent vehicles from crashing into surrounding
obstacles. Under spatial formulation, the easiest way is to treat the obstacle blocking the
lane by tightening the road boundary on one side, as shown in Fig. 5.2. In this way, collision
avoidance becomes a very simple box constraint that can be absorbed into the road boundary
constraint (5.18). However, it also raises the necessity of upper-level logic responsible for
optimally assigning obstacles to either side of the road, which is far from a trivial task [115].

Without the need for additional supervisory logic, another common approach to modeling
the collision avoidance constraint is to use distance constraints based on Euclidean norm

(s − sobs )2 (ye − ye,obs )2

h (ξ) := + − 1 ≥ 0, (5.27)
(2l1 )2 (2l2 )2

where sobs and ye,obs are the spatial coordinates of the obstacle, l1 and l2 are the geometric
parameters of the constraint ellipse. Figure 5.3 illustrates this constraint visually. To see its
connection to the previous formulation under spatial formulation, we rewrite constraint Eq.

Figure 5.2: Collision avoidance by road boundary tightening

137
 Figure 5.3: Collision avoidance using Euclidean norm

(5.27) in the following form given s is an independent clock state




 [ye,L , ye,U ] s ≥ sobs + 2l1 or s ≤ sobs − 2l1






" s #
(s − sobs )2

ye ∈ 2l2 1 − + ye,obs , ye,U ∪ . (5.28)

 (2l1 )2


 " s # sobs − 2l1 < s < sobs + 2l1
(s − sobs )2


ye,L , ye,obs − 2l2 1 −


(2l1 )2



From Eq. (5.28), we can see that the Euclidean distance constraint is equivalent to lateral
displacement constraints on ye , which is similar to the road tightening method. However,
it is also clear that it is a nonconvex constraint due to the fact that it creates two discrete
feasible regions of ye when sobs − 2l1 ≤ s ≤ sobs + 2l1 . Nevertheless, this nonconvexity does
not likely cause serious problems with optimization and control performance in practice, as
local minimum is usually considered good enough for real-time control tasks.

We will use the Euclidean norm formulation in our study due to the lack of supervisory
logic from the high-level decision making layer. The conventional way to embed collision
avoidance constraints into NMPC is to simply enforce Eq. (5.27) on every discretized node of
the prediction horizon, i.e. h (ξk ) ≥ 0 for ∀k ∈ {1, . . . , N }. However, under this formulation,
the collision avoidance constraints are always inactive if the obstacles are beyond the vehicle’s
forward reachable set along the prediction horizon. In other words, safety is only considered
within the prediction horizon. On the other hand, it is crucial for vehicles act proactively
when approaching to obstacles in some situations such as driving at high speed. This means a
long horizon is required so that the NMPC is able to take the faraway obstacles into account,
which adversely increases the computational load of optimization.

Motivated by the aforementioned reason, we will explore a new collision avoidance

138
constraint formulation using control barrier function theory, which guarantees the system’s
safety in the sense of set invariance.

5.4.2 Control Barrier Function for Safety Critical System

Control barrier function (CBF) was firstly defined in [116]. A modern version of CBF is given
in [114], and later refined in [117]. A nice overview of the theory and applications of CBF is
given in [118]. For completeness, the theory of CBF is outlined here, which follows closely to
those given in [118]. The only major difference is that CBF is stated in spatial domain in our
study whereas the derivation in the reference is presented in the time domain.

We consider a nonlinear affine control system

ξ ′ = f (ξ) + g (ξ) u, (5.29)

where ξ ∈ D ⊂ Rn is the system state in some admissible set, functions f : Rn → Rn and

g : Rn → Rn×m are locally Lipschitz, and u ∈ U ⊂ Rm is control input in some admissible
set.

Let u = k (ξ) be a feedback controller which is Lipschitz continuous, the closed-loop

dynamics of Eq. (5.29) are
ξ ′ = f (ξ) + g (ξ) k (ξ) . (5.30)

Due to the locally Lipschitz condition, for any initial condition ξ0 ∈ D there exists a maximum
interval of existence I (ξ0 ) = [s0 , sf ) such that ξ (s) is the unique solution to Eq. (5.30) on
I (ξ0 ).

The notion of safety is formalized in the context of enforcing invariance of a set. The safe
set C is defined as the superlevel set of a continuously differentiable function h : D ⊂ Rn → R:

C := {ξ ∈ D ⊂ Rn : h (ξ) ≥ 0} . (5.31)

Definition 5.1. The set C is forward invariant if for every ξ0 ∈ C, ξ (s) ∈ C for ξ (0) = ξ0
and all s ∈ I (ξ0 ). The system Eq. (5.30) is safe with respect to the set C if the set C is
forward invariant.

Before defining CBF, we denote a function α : R → R to be an extended class K∞

function if it is strictly increasing and with α (0) = 0. This allows us to define:

139
Definition 5.2. Let C ⊂ D ⊂ Rn be the superlevel set of a continuously differentiable
function h : D → R, then h is a control barrier function (CBF) if there exists an extended
class K∞ function α such that for the control system Eq. (5.29):

sup h′ (ξ, u) := sup (Lf h (ξ) + Lg h (ξ) u) ≥ −α (h (ξ)) , (5.32)

u∈U u∈U

for all x ∈ D.

Given a CBF h, we define the set consisting of all control values that satisfy:

Kcbf (ξ) := {u ∈ U : h′ (ξ, u) + α (h (ξ)) ≥ 0} . (5.33)

The main result that relates the existence of a CBF and safety of the control system is given
as

Theorem 5.1. Let C ⊂ D ⊂ Rn be a set defined as the superlevel set of a continuously

differentiable function h : D ⊂ Rn → R. If h is a control barrier function on D and ∂h
∂ξ
̸ 0
(ξ) =
for all ξ ∈ ∂C, where ∂C denotes the boundary of C, then any Lipschitz continuous controller
u (ξ) ∈ Kcbf (ξ) for the system (5.29) renders the set C safe.

Theorem 5.1 enables intuitive controller synthesis based on optimization. Suppose we

have a nominal feedback controller u = k (ξ) for control system (5.29), which is not necessarily
safe. Then a safety-guarantee controller can be constructed by the following minimum norm
problem.

1
u (ξ0 ) = arg min ∥u − k (ξ0 )∥2R (CBF-OCP)
u∈U 2
s.t. Lf h (ξ0 ) + Lg h (ξ0 ) u + α (h (ξ0 )) ≥ 0 (5.34)

where ξ0 is the current state measurement or estimation. Note that the constraint (5.34) is
linear in control at a given system state. Therefore, in case U is a convex polytope, CBF-OCP
becomes a QP, and can be solved quickly online.

However, constraint (5.34) can not be applied directly to our application. This is due to
the fact that h defined in (5.27) has relative degree of 2 with respect to the vehicle model
(5.7), so that u does not explicitly appears in h′ . We state the following definition of relative
degree given in [119].

140
Definition 5.3. The relative degree of a continuously differentiable function h : D ⊂ Rn → R
with respect to system (5.29) is the number of times we need to differentiate it along the
dynamics of (5.29) until the control u explicitly shows.

It can be easily shown that the collision avoidance constraint (5.27) is of relative degree 2.
Therefore, an alternative CBF formulation compatible with high relative degree systems must
be used. In [120], an exponential control barrier function based on input-output linearization
was first introduced for high relative degree systems. Then, a more generalized formulation
was proposed in [119] in the form of high order CBF (HOCBF). In this study, we use the
latter one and slightly modify it for the spatial domain.

For a continuously differentiable function h with relative degree of p, we define a series

of functions ψ0 : D → R, ψ1 : D → R, ψ2 : D → R, . . ., ψp : D → R such that

ψ0 (ξ) := h (ξ) , (5.35)

ψ1 (ξ) := ψ0′ (ξ) + α1 (ψ0 (ξ)) , (5.36)

ψ2 (ξ) := ψ1′ (ξ) + α2 (ψ1 (ξ)) , (5.37)

..
.
′
ψp (ξ, u) := ψp−1 (ξ) + αp (ψp−1 (ξ)) , (5.38)

where α1 (·), α2 (·), . . ., αp (·) denote extended class K∞ functions.

A series of set C1 , C2 , . . ., Cp associated with Eqns. (5.35)-(5.38) are defined:

C1 := {ξ ∈ D : ψ0 (ξ) ≥ 0} , (5.39)

C2 := {ξ ∈ D : ψ1 (ξ) ≥ 0} , (5.40)
..
.

Cp := {ξ ∈ D : ψp−1 (ξ) ≥ 0} . (5.41)

Note that C1 = C.

The HOCBF is defined using Eqns. (5.35)-(5.41).

Definition 5.4. A function h : D ⊂ Rn → R is a high order control barrier function (HOCBF)
of relative degree p for system (5.29) if there exist differentiable extended class K∞ functions

141
α1 , α2 , . . ., αp such that

ψp (ξ, u) := Lpf h (ξ) + Lg Lp−1 (p−1)

(5.42)


f h (ξ) u + O h (ξ) + αp (ψp−1 (ξ)) ≥ 0,

for all ξ ∈ C1 ∩ C2 ∩ . . . ∩ Cp . In (5.42), O (·) denotes the remaining Lie derivatives with degree
less than or equal to p − 1.

The set of control values that satisfy (5.42) is defined as

Khocbf (ξ) := {u ∈ U : ψp (ξ, u) ≥ 0} . (5.43)

We are now ready to give the following Theorem relating HOCBF to control system
safety.

Theorem 5.2. Given a HOCBF h from Def. 5.4 with the associated sets C1 , C2 , . . ., Cp , if
ξ (s0 ) ∈ C1 ∩ C2 ∩ . . . ∩ Cp , then any Lipschitz continuous controller u (ξ) ∈ Khocbf (ξ) renders
the safe set C forward invariant for system (5.29).

Theorem 5.2 allow us to synthesize an optimization based controller certifying the safety
of high relative degree control system (5.29)

1
u (ξ0 ) = arg min ∥u − k (ξ0 )∥2 (HOCBF-QP)
u∈U 2
s.t. Lpf h (ξ0 ) + Lg Lp−1 (p−1)
(5.44)


f h (ξ0 ) u + O h (ξ0 ) + αp (ψp−1 (ξ0 )) ≥ 0

where constraint (5.44) is again linear in u.

5.4.3 HOCBF based Collision Avoidance Constraint for NMPC

The optimization based controller CBF-OCP or HOCBF-QP are often used together with
control Lyapunov function to achieve stabilizing control with formal safety guarantee [117].
In these implementations, CBF or HOCBF is enforced only at current control step using
state measurement ξ0 . This can potentially lead to over greedy control actions resulting in
poor performance or infeasibility. On the contrary, NMPC optimizes the control actions over
the whole future prediction horizon, which generally achieves long-term optimal performance.

Here we integrate HOCBF constraint to NMPC framework in a multi-step fashion, so

that collision avoidance is certified not only at the current control step but in future steps.

142
We rewrite the definition of h here for convenience

(s − sobs )2 (ye − ye,obs )2

h (ξ) := + − 1, (5.45)
(2l1 )2 (2l2 )2

and the safe set is defined as

C := {ξ ∈ Rn : h (ξ) ≥ 0} . (5.46)

h (ξ) is of relative degree 2 when we differentiate it along the vehicle dynamics. This means we
need to define two extended class K∞ functions α1 (·) and α2 (·) for HOCBF formulation. For
the sake of simplicity, linear functions are adopted, which basically makes HOCBF become
as the exponential CBF
α1 (ψ0 ) := λ1 ψ0 , (5.47)

α2 (ψ1 ) := λ2 ψ1 , (5.48)

where λ1 and λ2 are tuning parameters whose physical meaning will be discussed later.
Furthermore, we assume that λ1 = λ2 := λ for simplicity.

We rewrite the vehicle dynamics (5.7) affine in control.

   v sin ψe (1−ye κ)   
ye v cos ψe
0 0
  1−ye κ
ψe   −κ 0 " #
   
  v cos ψe
d  1−ye κ  ω
(5.49)
    
v= 0 + 0
ds    v cos ψ
 a
e
    
s  1   0 0  |{z}
 
1−ye κ u
t v cos ψe 0 0
| {z } | {z } | {z }
ξ f (ξ) g(ξ)

With Eqns. (5.47), (5.48) and (5.49), HOCBF then can be derived as

ψ2 (ξ) = L2f h (ξ) + Lg Lf h (ξ) u + 2λh′ (ξ) + λ2 h (ξ) . (5.50)

Using Theorem 5.2, the collision avoidance constraint is expressed as

L2f h (ξ) + Lg Lf h (ξ) u + 2λh′ (ξ) + λ2 h (ξ) ≥ 0. (5.51)

Unlike the single-point constraint used in HOCBF-QP, we consider (5.51) over multiple steps

143
within prediction horizon.

L2f h (ξk ) + Lg Lf h (ξk ) uk + 2λh′ (ξk ) + λ2 h (ξk ) ≥ 0, for k = 0, . . . , Ncbf , (5.52)

where Ncbf ∈ {0, . . . , N } is the length of constraint horizon, which can be different from
the prediction horizon N. When Ncbf = 0, constraint (5.52) is linear in control given that
ξ0 is known, which can be handled by the optimization routine trivially. Whereas any
constraint beyond the first step results in a state-input coupled path constraint requiring
more computational effort. Therefore, Ncbf should be chosen carefully to balance control
performance and computational complexity.

5.4.4 HOCBF with Vehicle Speed Scheduling

We have seen how space based NMPC facilitates simpler constraint formulations for motion
planning. However, spatial formulation also creates an issue for motion planning due to the
fixed prediction horizon in space. Figure 5.4 depicts the difference in prediction horizons
of spatial and temporal formations at different vehicle speeds. In Fig. 5.4, hs and ts are
respectively the sample distance and sample time of spatial and temporal formations, which
are both constant values. In the case of time-based formulation, the total prediction length
grows with higher vehicle speed. This behavior is similar to what a human driver does, and
it is crucial for the stability and feasibility of collision avoidance at high speed. In the case of
space based formulation, however, the total prediction length remains constant regardless of
vehicle’s speed. Therefore, if the controller is tuned for low speed, it might have difficulty
avoiding obstacles when the vehicle is driving at a higher speed. While if tuned for high
speed, it will have an unnecessary long prediction horizon at low speed.

A naive remedy is to adjust the sample distance hs or the number of prediction steps N on
the fly according to vehicle speed. However, this requires reformulating the structure of OCP
in real-time, which not only creates additional computation overhead but also undermines
the effectiveness of warm-starting.

The utilization of HOCBF allows us to propose a more elegant solution to this issue.
Instead of changing the problem structure, we simply adapt the value of the hyperparameter λ
in HOCBF according to vehicle speed. To see the effect of λ, let us write down the definitions
of ψ0 , ψ1 , and ψ2
ψ0 = h, (5.53)

ψ1 = ψ0′ + λψ0 , (5.54)

144
 Figure 5.4: Comparison between space based and time based NMPC formulation

ψ2 = ψ1′ + λψ1 . (5.55)

Using above equations, HOCBF constraint ψ2 ≥ 0 can be expressed in terms of h

h′′ + 2λh′ + λ2 h ≥ 0. (5.56)

Equation (5.56) is a 2nd-order linear differential inequality of h. In order to find the lower
bound of h, let g := h′ + λh and (5.56) can be written as

g ′ + λg ≥ 0. (5.57)

Then, by comparison lemma [121], the solution of g satisfies

g ≥ Ae−λs , (5.58)

where A is a constant value depending on initial condition. Now substituting h back to Eq.
(5.58), we get
h′ + λh ≥ Ae−λs . (5.59)

Again, using comparison lemma, we obtain the bound of h as

h ≥ (As + B) e−λs > 0, (5.60)

where B is also a constant value depending on the initial condition. Equation (5.60) describes
the lower bound of h, which decreases exponentially with the rate of λ. In other words, λ
determines how fast a vehicle is allowed to approach an obstacle. Figure 5.5 shows intuitively

145
 Figure 5.5: Effect of λ on collision avoidance

the effect of λ on collision avoidance trajectory planning. Large λ tolerates faster decreasing
distance between vehicle and obstacle when the vehicle is far away, and starts evasion when
the vehicle is relatively close to the obstacle. On the other hand, small λ confines the vehicle
to approaching obstacles slowly, and also makes evasion earlier, resembling the effect of
increasing prediction horizon.

With the above knowledge in mind, it becomes obvious how we can schedule λ in order
to achieve consistent performance at different vehicle speeds. We want λ to be large at low
speed, and gradually decrease at higher vehicle speeds. In practice, a set of λ is firstly tuned
by simulations at several vehicle speeds, from low to high. Then a lookup table is generated
based on linear interpolation using these values. During real-time control, λ is updated at
each sample instant using the lookup table according to the current vehicle speed.

146
5.5 Summary of NMPC Formulation and Numerical Method

With all the ingredients at hand, we are ready to write down the NMPC formulation for
vehicle motion planning.
ˆ sf
1
minimize q1 ye2 (s) + q2 ψe2 (s) + q3 (v (s) − vr (s))2 + ∥u (s)∥2R ds
ξ(s),u(s), 2 s0
σf ,σξ ,σay ,σwp
2
+ ∥σf ∥2qf + ∥σξ ∥2qξ + σay qay
+ ∥σwp ∥2qwp (5.61)

subject to ξ (s0 ) = ξ0 (5.62)

ξ ′ = f (ξ) + g (ξ) u (5.63)
uL ≤ u (s) ≤ uU , ∀s ∈ [s0 , sf ] (5.64)
(v (s) ω (s))2 + a (s)2 ≤ µg + σf , ∀s ∈ [s0 , sf ] (5.65)
ξL (s) − σξ ≤ ξ (s) ≤ ξU (s) + σξ , ∀s ∈ [s0 , sf ] (5.66)
ay,L (s) − σay ≤ v (s) ω (s) ≤ ay,U (s) + σay , ∀s ∈ [s0 , sf ] (5.67)
ξwp,L − σwp ≤ ξ (swp ) ≤ ξwp,U + σwp , ∀swp ∈ Swp (5.68)
L2f h (ξk ) + Lg Lf h (ξk ) uk + 2λh′ (ξk ) + λ2 h (ξk ) ≥ 0, ∀k ∈ {0, . . . , Ncbf } (5.69)
σf ≥ 0, σξ ≥ 0, σay ≥ 0, σwp ≥ 0. (5.70)

For feasibility consideration, constraints on friction circle, state, lateral acceleration and
waypoint are softened by slack variables with 2-norm.

The NMPC presented above needs to be solved online in a receding horizon fashion.
Real-time feasibility might be an issue due to nonlinear and nonconvex path constraints. In
the previous chapter, we talked about how to solve trajectory tracking NMPC efficiently
based on direct multiple shooting with RTI scheme. In principle, the same algorithm could be
applied here for motion planning NMPC. However, there are two reasons why this algorithm
is not the most suitable choice here.

(1) Multiple shooting transcription usually enforces path constraints only at discretization
nodes,. Thus, the positions of the waypoints are also confined to these discrete points,
i.e. swp ∈ {s0 , s1 , . . . , sN }, which puts great limitations on the high-level decision
making algorithm if discretization is sparse.

(2) RTI scheme hinges on solving linearized OCP with only one QP iteration at each
sample period. Therefore, constraints are not satisfied strictly at each sample period.

147
 Figure 5.6: Direct collocation discretization

Convergence to an exact solution is achieved if the sample rate is fast enough. However,
for motion planning tasks with a relatively slow sample rate, exact solutions are
preferable at every sample step so that the safety critical collision avoidance constraint
is always ensured.

For the above two reasons, we choose to apply direct collocation transcription to the motion
planning NMPC. A noticeable difference between direct collocation and direct multiple
shooting is the discretization method. Direct collocation is based on the orthogonal collocation
discretization method which approximates system state on each collocation interval by a dth-
order polynomial pk (s, vk ) with vk ∈ Rnx (d+1) as coefficients [77]. Each polynomial consists of
a set of d collocation points, whose positions are chosen according to Gauss-Legendre scheme
in this study. Figure 5.6 shows the schematic of direct collocation discretization when d = 3.

The direct collocation allows us to possibly enforce the path constraints at arbitrary
point s by simply evaluating the polynomial pk (s, vk ). Therefore, the waypoint can be chosen
freely at any point within the prediction horizon, i.e. swp ∈ [s0 , sf ]. The NLP result form

148
direct collocation is expressed as

N −1
2
X
minimize lk (vk , ξk , uk ) + ∥σf ∥2qf + ∥σξ ∥2qξ + σay qay
+ ∥σwp ∥2qwp (5.71)
vk ,ξk ,uk
k=0
σ∈{σf ,σξ ,σay ,σwp }

subject to ξ0 − q0 = 0 (5.72)
vk,0 − ξk = 0, ∀k ∈ {0, . . . , N − 1} (5.73)
p′k (sk,1 , vk ) − f (vk,1 ) − g (vk,1 ) uk = 0, ∀k ∈ {0, . . . , N − 1} (5.74)
..
.
p′k (sk,d , vk ) − f (vk,d ) − g (vk,d ) uk = 0, ∀k ∈ {0, . . . , N − 1} (5.75)
pk (sk+1 , vk ) − ξk+1 = 0, ∀k ∈ {0, . . . , N − 1} (5.76)
ξwp,L − σwp ≤ pm (swp , vm ) ≤ ξwp,U + σwp , ∀swp ∈ Swp (5.77)
h (ξk , qk , σ) ≤ 0, ∀k ∈ {0, . . . , N − 1} (5.78)

where q0 is the state measurement or estimation, pm (·) is the collocation polynomial containing
the waypoint. Equations (5.73)-(5.75) are collocation constraints. Equation (5.76) is continuity
constraint. And Eq. (5.78) lumps all the other path constraints together.

Direct collocation typically results in a very large and sparse NLP with a dimension of
N nx (d + 2) + (N − 1) nu . They can be solved efficiently by structure-exploiting solvers. In
this study, the open source nonlinear interior point based IPOPT solver is used [122].

5.6 Simulation Results

Implementation of the motion planning NMPC is done using Matlab and Casadi similar to
that of tracking NMPC in the previous chapter. The model and controller parameters are
given in Appendix A. In the simulations, the trajectory tracking NMPC developed in the
previous chapter is used as a lower-level controller to track the planned trajectory given by
the motion planning NMPC. For the sake of simplicity, we only consider static obstacles in
the following simulations. The same control framework can be used for moving obstacles
with only a few modifications, which we briefly discuss in Appendix C.

Figure 5.7 shows the top level schematic of the motion planning and tracking control
system. The motion planning NMPC takes information and commands such as obstacle
positions, waypoints, and road curvature from higher-level perception and decision making
modules. Together with vehicle state feedback, it computes a reference trajectory for the

149
 Figure 5.7: Motion planning and tracking control system

Figure 5.8: Road configuration for DLC. The blue solid line denotes lane condounary. The
blue dashed line denotes lane center; red dashed line denotes obstacle.

tracking controller. The trajectory tracking NMPC computes steering and slip control inputs
to the vehicle based on reference and state feedback. Note that the motion planning is
formulated in the Frenet frame, while trajectory tracking is in Cartesian frame. Therefore,
conversion blocks are needed to connect the planning and tracking loops.

5.6.1 Obstacle Avoidance with Double Lane Change

We consider the problem of avoiding a static obstacle sitting on a straight road segment with
a double lane change maneuver. The road has two lanes where the obstacle is located on
the bottom lane at 50m along the X direction. The lane width is 3.6m, and the obstacle
is 4.6m long and 2.4m wide. The controlled vehicle enters the road in the bottom lane at
X = 0. Figure 5.8 illustrates the road configuration of the simulation. We first simulate a
vehicle with an entry speed of 70km/h. For comparison, we tuned two NMPC controllers, one
with CBF based collision avoidance constraints and the other with the conventional distance
constraint. Figure 5.9 shows the vehicle trajectory of the two controllers. We can see the two
formulations are able to achieve very similar responses with proper tuning. However, there is
a subtle difference in how they handle collision avoidance constraints. In Fig. 5.9 we denote
the boundaries of two collision avoidance constraints, i.e. h = 0. It is shown that in order
to achieve a similar trajectory, the constraint based on distance must be larger than that
based on CBF. This is due to the fact that the trajectory based on distance constraint always
tends to stay on the constraint boundary to achieve a smaller objective function value. While
the trajectory based on CBF will leave a small buffer to the boundary since CBF defines

150
Figure 5.9: Collision avoidance at 70km/h (µ = 1). The red dotted ellipse denotes CBF
constraint. The blue dotted ellipse denotes distance constraint.

Figure 5.10: Collision avoidance at 85km/h (µ = 1). The magenta rectangle denotes the
shape of controlled vehicle.

not only the safe set but the way the vehicle is allowed to approach the set boundary by
hyperparameter λ.

Using the same controller parameters, we now simulate the vehicle at an 85km/h entry
speed. The results are shown in Fig. 5.10. In order to obtain feasible solutions, we have to
formulate the collision avoidance constraint to soft constraint. We can see that the vehicle
with CBF formulation is still able to avoid the obstacle, while the other one collides with the
obstacle due to the soften constraint. This result shows that the CBF based collision avoidance
is more robust against vehicle speed than the conventional distance-based constraint.

Figure 5.11 shows more simulations of NMPC with CBF at different vehicle speeds.
It shows that without retuning the controller parameters, the vehicle is able to avoid the
obstacle with a consistent distance at a wide range of speeds based on the scheduling of λ.

Next, we show the motion planning with additional constraints such as lateral acceleration

151
 (a) 45km/h

(b) 60km/h

(c) 80km/h

(d) 100km/h

Figure 5.11: Collision avoidance at various vehicle speeds (µ = 1)

152
 Figure 5.12: Collision avoidance with lateral acceleration constraints (µ = 1)

and waypoints. Figure 5.12 shows the motion planning with different lateral acceleration
constraints. The vehicle entry speed is 50km/h. We see that the motion planner is able to
generate different trajectories according to the lateral acceleration requirements. It shows
that the maximum lateral accelerations are reduced according to the constraints, although
they are slightly over the limit due to the softened constraints. However, although the lateral
acceleration does meet the limit, the controller tends to fulfill it by decelerating the vehicle
speed, which is not very preferable in terms of the driver’s comfort. To avoid an over decrease
in speed, we simply increase the lower bound of the speed constraint. The lower bound of
speed is set to be 80% of the entry speed. The result is shown in Fig. 5.13. Now, instead
of reducing the vehicle speed, the motion planner actually modifies the shape of trajectory
to have a smaller curvature such that the lateral acceleration limit is satisfied, thus really
improving the driver’s comfort.

We now show the motion planning with waypoint constraints. The two waypoints are set
as swp = {30, 55}, ye,wp = {3, 2.5}, and twp = {1.5, 2.7} . The simulation is presented in Fig.
5.14. We can see from the simulation that the motion planner generates a feasible trajectory
that pass both the spatial and temporal points with only small deviations due to the soft
constraint slackness. The vehicle speed is automatically scheduled by the motion planner to
meet the temporal constraints.

5.6.2 Oval Circuit Tracking with Obstacles

We have shown the motion planning on a straight road with an obstacle. Now we will
simulate the vehicle with the motion planner on an oval circuit with multiple obstacles. The

153
Figure 5.13: Collision avoidance with lateral acceleration constraints and additional speed
constraints (µ = 1)

(a) Vehicle trajectory

(b) Trajectory of lateral deviation and time (c) Vehicle speed

Figure 5.14: Motion planning with waypoint constraints (µ = 1)

154
vehicle target speed is 70km/h. Figure 5.15 shows the simulation results with two different
configurations of objective weighting. In Fig. 5.15, the blue solid line represents the trajectory
based on heavy weighting on lateral deviation, while the red dash-dotted line is the result
of the tuning that emphasizes yaw angle deviation. The values of weighting can be found
in Appendix A. The rest of the controller parameters are the same in both cases. Lateral
acceleration and waypoint constraints are not used in the simulations. We can see from
the results that the vehicle is able to complete the circuit without collision in both setups.
Both trajectories have maximum lateral accelerations of around 0.9g, which demonstrates
the capability of the proposed controller to handle the vehicle at the limit. However, the
two set of trajectories have subtle differences. When the lateral deviation is weighted more,
the vehicle always tries to return back to the center line whenever it is possible. Therefore,
the lateral deviation is generally smaller, but the yaw angle deviation suffers relatively large
fluctuations. It also shows a larger speed variation. On the other hand, the trajectory with
larger weight on yaw angle deviation is much smoother with less speed variation despite bigger
lateral deviation. Note that the two trajectories take different homotopies when passing
the obstacle on the left half of the circuit, reflecting the nonconvex nature of the collision
avoidance constraint.

We also investigated how the speed weighting affects the motion planning. Figure 5.16
shows the trajectories with two different speed weightings, those values can be found in
Appendix A.. We see that the motion planner with large speed weighting tends to generate a
trajectory with smaller curvatures in order to accommodate the constraints on the friction
ellipse. While the controller with smaller weighting is able to make large curvatures by
reducing the vehicle speed. It also shows a generally smaller lateral acceleration with a
smaller weighting on vehicle speed.

155
 (a) Vehicle paths

(b) Vehicle states

Figure 5.15: Oval circuit tracking with heavy weighting on lateral deviation and heading
angle deviation (µ = 1)

156
 (a) Vehicle paths

(b) Vehicle states

Figure 5.16: Oval circuit tracking with large and small weightings on speed (µ = 1)

157
5.7 Conclusions

A motion planning algorithm based on the NMPC framework has been presented. The
NMPC generates reference trajectories with numerical optimization based on the spatial
domain point mass model and various input and state constraints. We incorporated CBF
based collision avoidance constraints into the NMPC formulation so that it receives formal
safety certification. Various simulations are conducted to evaluate the performance of the
hierarchical motion planning and tracking control system. The simulation results have shown
the great capability and flexibility of the proposed method for autonomous vehicle motion
planning and control.

158
 Chapter 6

Summary and Future Work

6.1 Summary

In this dissertation, the applications of model-based control design approaches for autonomous
vehicles are presented. The four applications are divided into two categories according to the
multi-layer schematic of the autonomous vehicle control system. For the low-level actuation
layer, control development of EPAS and EMB systems are presented. While for the high-level
motion control layer, trajectory tracking and motion planning controller design are discussed.
The following summarizes the findings and contributions of these four cases.

6.1.1 EPAS Steering Feel Control

Drivers are sensitive to the steering feel of a vehicle and significant tuning efforts are placed to
set the desired steering feel. Chapter 2 discusses a closed loop torque overlay control system
to improve the steering feel of an electric power-assisted steering (EPAS) system. This system
has a reference model, a rack force estimator, and a tracking controller. A target steering feel
is generated from a reference model in which it uses an estimate of rack force determined by
a Kalman filter based rack force estimator to reflect the actual vehicle operating conditions.
A sliding mode controller is designed for tracking the target steering feel. The performance
of the proposed control system is evaluated through simulation and a hardware-in-the-loop
test. The simulation and experimental results show that the EPAS system with the proposed
control method can successfully improve the steering feel under various operation conditions.

6.1.2 EMB Clamping Force Control

The EMB is a key component of the brake-by-wire (BBW) system, generating the clamping
force between the brake pad and disk through an electric motor and mechanical transmission.
Control of the clamping force has a significant impact on the brake performance. In Chapter 3,
an EMB clamping force control system is developed that addresses several major challenges in

159
practical implementation. Firstly, a nonlinear EMB model including the DC motor, planetary
gear set, ball screw, and clamping force model is built. In particular, a novel clamping force
model is introduced based on a linear transform of two polynomial functions. For the control
system design, a clamping force estimation algorithm is proposed that requires only the
existing measurements of motor torque, angular speed, and position. The contact point
of the brake pad and the disk is also estimated using an internal model controller and a
Rauch–Tung–Striebel smoother. The tracking controller is based on the disturbance observer
structure, with a PI feedback controller and a zero-phase error tracking feedforward controller.
Furthermore, a unified architecture is proposed such that a smooth transition between gap
closing and clamping force tracking is realized. Finally, the performance of the entire control
system under various conditions is evaluated based on simulation.

6.1.3 NMPC based Trajectory Tracking Control

Trajectory tracking is an essential component of the autonomous motion control system.

The main objective of the controller is to control the vehicle to closely follow the reference
trajectory by manipulating various control inputs. In Chapter 4, we proposed a NMPC
controller for trajectory tracking. The subject vehicle is assumed to be equipped with 4WS
and 4WID. A continuous-time NMPC is first formulated with a quadratic objective function
balancing the tracking error and control effort. Constraints on control amplitude and rate
are also embedded in the NMPC formulation. The standard formulation is extended by the
frequency shaping technique to take into account different frequency domain requirements for
actuators. In order to efficiently solve the NLP in real-time, an algorithm that incorporates
the RTI scheme and control parameterization is proposed. The resulted QP problem is much
smaller and its dimension is independent of the prediction horizon. Simulation results show
that the NMPC effectively controls the vehicle to track the reference up to the handling
limit under various road friction conditions. All the control inputs are smooth thanks to the
frequency shaping. The control algorithm is also real-time feasible.

6.1.4 NMPC based Motion Planning

Motion planning is the high-level controller on the motion control layer of autonomous vehicles.
It is responsible for generating a local trajectory as a reference to the tracking controller.
The problem is formulated in the spatial domain, which entails simple representation of road
boundary constraints. The trajectory is subjected to control input constraints presented by
amplitude, rate, and friction circle. Different path constraints, such as road boundary, lateral
acceleration, and waypoints are also taken into consideration by the planner. HOCBF is used

160
to formulate the safety critical collision avoidance constraint. The numerical method is based
on direct collocation transcription and an interior point solver, which enables efficient solving
by sparsity exploitation and flexible choice of waypoint constraints. The motion planner is
simulated together with the trajectory controller on a straight road and oval circuit with
obstacles. The results show that the NMPC based motion planner successfully computes a
dynamically feasible trajectory satisfying the control inputs and path constraints, and the
tracking controller is able to control the vehicle to follow the trajectory closely.

6.2 Future Work

In this section we aim to give some directions for future work that are closely related to the
topics in this dissertation.

6.2.1 EPAS Control

Chapter 2 discussed EPAS steering feel control which utilizes the reference model and rack
force estimation. Based on this structure, a more flexible steering feel can be realized. The
estimated rack force can be further modified according to the driver’s preference and the
vehicle’s driving condition before it is used by the reference model. For example, the rack
force estimation can be increased on low µ surface and decreased on high µ surface to achieve
a more consistent steering feel. It can also be increased within a small steering angle and
decreased within a larger steering angle to emphasize on-center steering feel while reducing
steering effort within a large steering angle.

Furthermore, the same framework of steering feel control can be applied to steering
position control in the future since the steering angle is used as the control variable. The
compatibility of steering feel control and position control also enables a shared control strategy.
For example, the control system can focus on the steering feel control when the vehicle is in
the safe region, and gradually switch into position control when the vehicle approaches the
unsafe region.

6.2.2 EMB Control

The clamping force estimation algorithm is presented in Chapter 3, and its performance is
validated via experimental testing. However, in real-world operation, temperature variation
and wear will affect the clamping force significantly. Therefore, mechanisms for adapting
to the temperature and wear conditions should be developed. One possible way is to use
a scaling factor to modify the the envelop of the clamping force. The values of the scaling

161
factor must be calibrated using experimental tests. The machine learning based approaches
are also promising for the modeling of nonlinear clamping force.

From the perspective of vehicle dynamics control, the EMB clamping force control system
can be incorporated with wheel slip controller and simulated with vehicle model to validate its
performance on ESC or torque vectoring. Different control algorithms can also be developed
for the clamping force tracking. For example, H∞ control or LPV control can be used to
achieve robust performance against disturbance and uncertainty.

6.2.3 Trajectory Tracking Control

The trajectory tracking NMPC control structure can be easily extended to include more
actuators, such as active suspension and active anti-roll bar to control the vehicle’s vertical
and roll motions. With more actuators considered, the computational demand will also
increase. Therefore, the controller parameters should be tuned more carefully to balance the
control performance and computation time. In particular, alternative control parameterization
schemes can be explored, such as exponential parameterization, polynomial parameterization,
and Fourier parameterization. Furthermore, based on the low-level actuators, different
parameterizations can be used independently.

The deterministic NMPC framework can be robustified against model uncertainty and
disturbance by using techniques such as tube MPC, stochastic MPC, or adaptive MPC.
However, computational demand will become an issue for these methods. Although robust
control within the NMPC framework might be challenging, it has been investigated in the
context of other control methodologies, e.g. H∞ control, µ synthesis, and sliding mode
control. Therefore, the combination of NMPC control with other robust control techniques
can be a promising approach. In particular, since the vehicle trajectory control is based on a
over-actuated system, we can choose some actuators, such as steering, for trajectory tracking
using the NMPC method, and other actuators, such as wheel slip, for disturbance rejection
using robust control methods.

The frequency shaping method is only used for control input in our study. The same
technique can be straightforwardly extended to the system state. For example, we can
constrain the high frequency component of vehicle lateral and longitudinal acceleration to
improve driving comfort. However, the constraints on state might affect the trajectory
tracking performance, which could possibly cause safety issues since collision avoidance is not
considered at the tracking level. Therefore, the controller tuning must be careful.

162
6.2.4 Motion Planning

The feasibility issue of NMPC based motion planning must be studied in the future. Due to
the complexity of the environment and uncertainty of the measurement, a feasible solution
might not always be found. Therefore, a feasibility condition for the motion planning NMPC
should be studied. If a feasible solution cannot be found, a fail-safe mechanism must be
implemented to avoid the collision.

Personalized motion planing is a possible direction for future research. Human drivers
might have various preferences in terms of vehicle motion. For example, some may prefer
to follow the preceding car rather than overtake. Personalization of motion planning is able
to improve the driving experience and satisfy the driver’s expectations. Learning-based
approaches such as reinforcement learning techniques can be used for this purpose.

Only single vehicle motion planning is considered in this dissertation. To achieve better
traffic safety and efficiency, multi-vehicle motion planning or cooperative driving should
be investigated using frameworks such as distributed MPC and game theory. Formation
control, also known as convoy control can also be applied to multi-vehicle motion planning
problem. In formation control, the vehicles drive as a whole and change the formation
structure adaptively according to the environment. It is particularly good at handling the
scenarios such as lane-drop bottlenecks, on-ramp merging, and so on.

Finally, in order to implement the motion control system on a real vehicle, a companion
vehicle state estimation pipeline must be incorporated alongside with the control system.
Particle filtering and receding horizon estimation are two promising techniques that can be
used for this purpose.

163
 Appendix A

List of Model and Controller Parameters

The vehicle model and controller parameters that are used in Chapter 4 and 5 are listed as
follows.

Table A.1: Vehicle parameters

Parameters Meaning Value
m Vehicle mass 2010 [kg]
I Vehicle yaw inertia 3300 [kg/m2 ]
a Distance from C.G. to front axle 1.05 [m]
b Distance from C.G. to rear axle 1.45 [m]
c Half track width 0.75 [m]
h Height between C.G. and wheelbase 0.4 [m]
τL Time constant of load transfer 0.01 [s]

Table A.2: Tire parameters

Road surface Parameters Longitudinal force Lateral force
µ 1 1
B 11.5 8.6
Asphalt
C 1.6 1.1
E 0.35 -1.2
µ 0.4 0.4
B 10 19
Snowy
C 2 0.55
E 0.6 -2.1
µ 0.17 0.17
B 31 28
Icy
C 1.7 1.5
E 0.7 -1.2

164
 Table A.3: Parameters for the double lane change reference path
Parameters dy1 dy2 dx1 a1 b1 c1 a2 b2 c2
Value 3.7 3.7 25 0.096 19 1.2 0.1 71 1.2

Table A.4: NMPC tracking controller parameters

Parameters Meaning Value
T Prediction horizon 1 [s]
Ts Simple time 0.025 [s]
Ncp Control parametrization nodes {1, 11, 15}
Q Reference tracking weighting diag (0.26, 300, 10, 3.2)
R Control inputs weighting diag (500, 500, 1250, 1250, 1250, 1250)
δf,U/L Front steering angle limit ±30◦
δ̇f,U/L Front steering rate limit ±60◦ /s
δr,U/L Rear steering angle limit ±10◦
δ̇r,U/L Rear steering rate limit ±30◦ /s
λ̇U/L Wheel slip rate limit ±1/s
Rf,f ws Frequency shaping filter for front steering 0.02s+1
0.003s+1
Rf,rws Frequency shaping filter for rear steering 0.05s+1
0.005s+1
Rf,slp Frequency shaping filter for wheel slip 0.01613s+1
0.001613s+1

Table A.5: NMPC motion planning controller parameters

Parameters Meaning Value
M Prediction horizon 30 [m]
h Simple distance 1.5 [m]
Q1 State weighting with large weight on ye diag (1.2, 40, 0.01, 0, 0)
Q2 State weighting with large weight on ψe diag (0.4, 120, 0.01, 0, 0)
Q3 State weighting with large weight on v diag (1.2, 40, 0.05, 0, 0)
Q4 State weighting with small weight on v diag (1.2, 40, 0.001, 0, 0)
R Control inputs weighting diag (1, 0.1)
ωU,L Yaw rate limit ±15◦ /s
aU,L Acceleration limit ±7m/s2
ω̇U,L Yaw rate limit ±30◦ /s2
ȧU,L Acceleration rate limit ±15m/s3
vi Sample speed for λ scheduling {0, 30, 50, 70, 85, 100, 150} [kph]
λi Value of λ at sample speed {0.5, 0.5, 0.3, 0.3, 0.2, 0.15, 0.15}

165
 Appendix B

Determining Rate Constraint of Wheel Slip

Since the wheel slip ratios are used as control inputs of the proposed trajectory tracking
NMPC, their rate constraints should be chosen properly according to the psychical capability
of the low-level actuators. The following presents how we can derive a realistic rate constraint
on slip ratio.

The most deterministic factor that affects how fast slip ratio changes is the maximum
brake torque available, which can be roughly calculated from tire force. Assuming pure slip
condition, the maximum tire force is

Fx,max = µFz , (B.1)

where µ is road friction coefficient, and Fz is tire normal load. Considering load transfer, Fz
on a single front wheel is approximately

h
Fz = Fz0 + max , (B.2)
2L

where Fz0 is the static load, ax is the deceleration, h is the C.G height, and L is the wheel
base. Assuming full brake where ax = µg and µ = 1, the maximum tire force obtained as

h ∼
Fx,max = Fz0 + mg = 7300N. (B.3)
2L

Using the relationship between brake torque and tire force, we have

Tb,max = Fx,max rw ∼
= 2700N m. (B.4)

where rw = 0.37m is the wheel radius. However, due to the tire relaxation length and brake
system dynamics, Tb,max cannot reached immediately. Therefore, a first-order lag with time

166
 Figure B.1: Brake torque profile

Figure B.2: Slip ratio response

constraint of 0.01s is used to generate the profile of Tb,max , as shown in Fig. B.1.

The brake torque trajectory in Fig. B.1 is then apply to a quarter car model to generate
the slip ratio response, which is considered as the upper bound of achievable slip ratio by the
actuator. The quarter car model is expressed as

mv̇ = −Fx , (B.5)

J ω̇ = rw Fx − Tb sgn (ω) , (B.6)

v − ωrw
λ= , (B.7)
v
where m is the mass of the quarter car, v is the longitudinal speed, and ω is the wheel
rotational speed. Figure B.2 shows the response of the slip ratio. In Fig. B.2, we only show
the response where λ is within the range of positive tire force slop, as the NMPC controller
will not let the tire operate in negative slop. Neglecting the initial transition part, we can see
that the λ̇ approaches the maximum of 4s−1 and gradually reduces to 1s−1 at the end. Based

167
on the simulation results, the rate constraint of λ are set as 1s−1 . For reference, a sliding
mode based slip controller for passenger cars is designed in [69] and is able to reach λ = 0.1
in around 0.05 seconds, which means an equivalent rate of 2s−1 . Therefore, our choice should
be totally achievable and pretty conservative.

168
 Appendix C

Moving Obstacle Collision Avoidance Constraint

Let us first rewrite the HOCBF h for collision avoidance here

(s − sobs )2 (ye − ye,obs )2

h (ξ) := + − 1, (C.1)
(2l1 )2 (2l2 )2
h iT
where ξ := ye ψe v s t is the state of vehicle model.

The constraint of HOCBF based collision avoidance is given as

L2f h (ξ) + Lg Lf h (ξ) u + 2λh′ (ξ) + λ2 h (ξ) ≥ 0. (C.2)

We can see that the constraint (C.2) is expressed in terms of the derivatives of h along the
system dynamics. When obstacles are static, sobs and ye,obs are constant values. Therefore,
the derivatives can be calculated fairly straightforwardly. If obstacles are moving, however,
sobs and ye,obs become time varying, and thus their derivatives also become nonzero. In
order to find obstacle’s derivatives, a common way is to use a kinematic model that assumes
constant speed and acceleration within the prediction horizon,

1
sobs (t) = s0 + vs t + as t2 , (C.3)
2
1
ye,obs (t) = ye,0 + vy t + ay t2 , (C.4)
2
where the initial position s0 and ye,0 , speed vs and vy , and acceleration as and ay are assumed
known by perception system. Plug Eqns. (C.3) and (C.4) into (C.1), we obtain h for moving
obstacle, which now depends on an additional state t. Now the derivatives of h can be
calculated.

169
 References

1. V. Ivanov, D. Savitski, and B. Shyrokau, “A Survey of Traction Control and Antilock

Braking Systems of Full Electric Vehicles With Individually Controlled Electric
Motors,” IEEE Trans. Veh. Technol., vol. 64, no. 9, pp. 3878–3896, Sep. 2015. DOI:
10.1109/TVT.2014.2361860

2. L. Zhai, T. Sun, and J. Wang, “Electronic Stability Control Based on Motor

Driving and Braking Torque Distribution for a Four In-Wheel Motor Drive Electric
Vehicle,” IEEE Trans. Veh. Technol., vol. 65, no. 6, pp. 4726–4739, Jun. 2016. DOI:
10.1109/TVT.2016.2526663

3. M. Doumiati, O. Sename, L. Dugard, J.-J. Martinez-Molina, P. Gaspar, and Z. Szabo,

“Integrated vehicle dynamics control via coordination of active front steering and rear
braking,” European Journal of Control, vol. 19, no. 2, pp. 121–143, Mar. 2013. DOI:
10.1016/j.ejcon.2013.03.004

4. T. Gordon, M. Howell, and F. Brandao, “Integrated Control Methodologies for Road

Vehicles,” Vehicle System Dynamics, vol. 40, no. 1-3, pp. 157–190, Jan. 2003. DOI:
10.1076/vesd.40.1.157.15877

5. K. Bengler, K. Dietmayer, B. Farber, M. Maurer, C. Stiller, and H. Winner, “Three

Decades of Driver Assistance Systems: Review and Future Perspectives,” IEEE Intell.
Transport. Syst. Mag., vol. 6, no. 4, pp. 6–22, 2014. DOI: 10.1109/MITS.2014.2336271

6. J. Ziegler, P. Bender, M. Schreiber, H. Lategahn, T. Strauss, C. Stiller, Thao Dang,

U. Franke, N. Appenrodt, C. G. Keller, E. Kaus, R. G. Herrtwich, C. Rabe, D. Pfeiffer,
F. Lindner, F. Stein, F. Erbs, M. Enzweiler, C. Knoppel, J. Hipp, M. Haueis,
M. Trepte, C. Brenk, A. Tamke, M. Ghanaat, M. Braun, A. Joos, H. Fritz, H. Mock,
M. Hein, and E. Zeeb, “Making Bertha Drive—An Autonomous Journey on a Historic
Route,” IEEE Intell. Transport. Syst. Mag., vol. 6, no. 2, pp. 8–20, 2014. DOI:
10.1109/MITS.2014.2306552

7. W. Schwarting, J. Alonso-Mora, and D. Rus, “Planning and Decision-Making for

Autonomous Vehicles,” Annu. Rev. Control Robot. Auton. Syst., vol. 1, no. 1, pp.
187–210, May 2018. DOI: 10.1146/annurev-control-060117-105157

8. C. Katrakazas, M. Quddus, W.-H. Chen, and L. Deka, “Real-time motion planning

methods for autonomous on-road driving: State-of-the-art and future research

170
 directions,” Transportation Research Part C: Emerging Technologies, vol. 60, pp.
416–442, Nov. 2015. DOI: 10.1016/j.trc.2015.09.011

9. Z. He and M. Gu, “Dynamic Research on Control Strategy of Electric Power Steering

System,” in SAE 2012 World Congress & Exhibition, Apr. 2012, pp. 2012–01–0212.
DOI: 10.4271/2012-01-0212

10. I. A. Badiru, “Customer Focus in EPS Steering Feel Development,” SAE Int. J. Passeng.
Cars - Mech. Syst., vol. 7, no. 3, pp. 1009–1015, Apr. 2014. DOI: 10.4271/2014-01-0148

11. K. D. Norman, “Objective Evaluation of On-Center Handling Performance,” in SAE

International Congress and Exposition, Feb. 1984, p. 840069. DOI: 10.4271/840069

12. M. Bröcker, “New control algorithms for steering feel improvements of an electric
powered steering system with belt drive,” Vehicle System Dynamics, vol. 44, no. sup1,
pp. 759–769, Jan. 2006. DOI: 10.1080/00423110600885780

13. I. Yamazaki, I. Kushiro, and Y. Kunihiro, “Electronic Power Steering Compensating

Control for Influence of Vehicle Dynamics on Steering Torque,” SAE Int. J. Passeng.
Cars – Mech. Syst., vol. 2, no. 1, pp. 239–246, Apr. 2009. DOI: 10.4271/2009-01-0049

14. Tao Yang, “A New Control Framework of Electric Power Steering System Based on
Admittance Control,” IEEE Trans. Contr. Syst. Technol., vol. 23, no. 2, pp. 762–769,
Mar. 2015. DOI: 10.1109/TCST.2014.2325892

15. A. Marouf, M. Djemai, C. Sentouh, and P. Pudlo, “A New Control Strategy of an

Electric-Power-Assisted Steering System,” IEEE Trans. Veh. Technol., vol. 61, no. 8, pp.
3574–3589, Oct. 2012. DOI: 10.1109/TVT.2012.2209689

16. C. C. de Wit, S. Guegan, and A. Richard, “Control design for an electro power steering
system: Part I the reference model,” in 2001 European Control Conference (ECC).
Porto: IEEE, Sep. 2001, pp. 3611–3616. DOI: 10.23919/ECC.2001.7076494

17. M. Moradkhani, M. R. Hairi-Yazdi, and F. R. Salamsi, “Hinf loop-shaping control

of an electric power steering system,” in 2007 IEEE/ASME international conference
on advanced intelligent mechatronics. Zurich, Switzerland: IEEE, 2007, pp. 1–6. DOI:
10.1109/AIM.2007.4412513

18. D. Lee, K. Yi, S. Chang, B. Lee, and B. Jang, “Robust steering-assist torque control
of electric-power-assisted-steering systems for target steering wheel torque tracking,”
Mechatronics, vol. 49, pp. 157–167, Feb. 2018. DOI: 10.1016/j.mechatronics.2017.12.007

19. M. Dohring, E. Lee, and W. Newman, “A load-dependent transmission friction model:

theory and experiments,” in [1993] Proceedings IEEE International Conference on
Robotics and Automation. Atlanta, GA, USA: IEEE Comput. Soc. Press, 1993, pp.
430–436. DOI: 10.1109/ROBOT.1993.292210

171
20. F. Wilhelm, T. Tamura, R. Fuchs, and P. Mullhaupt, “Friction Compensation Control
for Power Steering,” IEEE Trans. Contr. Syst. Technol., vol. 24, no. 4, pp. 1354–1367,
Jul. 2016. DOI: 10.1109/TCST.2015.2483561

21. K. Johanastrom and C. Canudas-de Wit, “Revisiting the LuGre friction model,” IEEE
Control Syst., vol. 28, no. 6, pp. 101–114, Dec. 2008. DOI: 10.1109/MCS.2008.929425

22. E. Garcia, P. Gonzalez de Santos, and C. Canudas de Wit, “Velocity Dependence in the
Cyclic Friction Arising with Gears,” The International Journal of Robotics Research,
vol. 21, no. 9, pp. 761–771, Sep. 2002. DOI: 10.1177/0278364902021009877

23. Y. Li, T. Shim, D. Wang, and T. Offerle, “Study on parameters affecting steering
feel of column assist electric power steering,” IJVD, vol. 77, no. 3, p. 153, 2018. DOI:
10.1504/IJVD.2018.098941

24. Jose Velazquez Alcantar, “Improving Steering Feel in Electric Power Steering Systems:
A Model Reference Feedback Controller Approach,” Master’s thesis, University of
California, Davis, 2014.

25. A. Balachandran and J. C. Gerdes, “Designing Steering Feel for Steer-by-Wire Vehicles
Using Objective Measures,” IEEE/ASME Trans. Mechatron., vol. 20, no. 1, pp. 373–383,
Feb. 2015. DOI: 10.1109/TMECH.2014.2324593

26. Y. Li, T. Shim, D. Wang, and T. Offerle, “Comparative Study of Rack Force Estimation
for Electric Power Assist Steering System,” in Volume 3: Vibration in Mechanical
Systems; Modeling and Validation; Dynamic Systems and Control Education; Vibrations
and Control of Systems; Modeling and Estimation for Vehicle Safety and Integrity;
Modeling and Control of IC Engines and Aftertreatment Systems; Unmanned Aerial
Vehicles (UAVs) and Their Applications; Dynamics and Control of Renewable Energy
Systems; Energy Harvesting; Control of Smart Buildings and Microgrids; Energy
Systems. Tysons, Virginia, USA: American Society of Mechanical Engineers, Oct. 2017,
p. V003T33A005. DOI: 10.1115/DSCC2017-5255

27. ——, “Enhancement of Steering Feel of Electric Power Assist Steering System Using
Modeling Reference Control,” in 2018 Annual American Control Conference (ACC).
Milwaukee, WI: IEEE, Jun. 2018, pp. 3257–3262. DOI: 10.23919/ACC.2018.8431923

28. T. Weiskircher, S. Fankem, and B. Ayalew, “Rack Force Estimation for Electric
Power Steering,” in Volume 3: 17th International Conference on Advanced Vehicle
Technologies; 12th International Conference on Design Education; 8th Frontiers in
Biomedical Devices. Boston, Massachusetts, USA: American Society of Mechanical
Engineers, Aug. 2015, p. V003T01A007. DOI: 10.1115/DETC2015-46228

29. Y. Li, T. Shim, D. Wang, and T. Offerle, “Effect of Load-Dependent Friction on the
Estimation of Rack Force in Electric Power-Assisted Steering System,” Journal of
Dynamic Systems, Measurement, and Control, vol. 141, no. 11, p. 111005, Nov. 2019.
DOI: 10.1115/1.4044181

172
30. C. Maron, T. Dieckmann, S. Hauck, and H. Prinzler, “Electromechanical Brake System:
Actuator Control Development System,” in SAE International Congress and Exposition,
Feb. 1997, p. 970814. DOI: 10.4271/970814

31. R. Schwarz, R. Isermann, J. Böhm, J. Nell, and P. Rieth, “Modeling and Control of an
Electromechanical Disk Brake,” in SAE International Congress & Exposition, Feb. 1998,
p. 980600. DOI: 10.4271/980600

32. C. F. Lee, “Brake force control and judder compensation of an automotive electromechan-
ical brake,” Ph.D. dissertation, Department of Mechanical Engineering, The University
of Melbourne, 2013.

33. C. Line, C. Manzie, and M. Good, “Control of an Electromechanical Brake for

Automotive Brake-By-Wire Systems with an Adapted Motion Control Architecture,” in
SAE 2004 Automotive Dynamics, Stability & Controls Conference and Exhibition, May
2004, pp. 2004–01–2050. DOI: 10.4271/2004-01-2050

34. E. Holweg, R. Klomp, J. Klaassens, and E. Lomonova, “Modeling and Inverse

Model-Based Control of an Electro-Mechanical Brake Actuator,” IFAC Proceedings
Volumes, vol. 33, no. 26, pp. 39–44, Sep. 2000. DOI: 10.1016/S1474-6670(17)39118-8

35. C. Line, C. Manzie, and M. Good, “ROBUST CONTROL OF AN AUTOMOTIVE

ELECTROMECHANICAL BRAKE,” IFAC Proceedings Volumes, vol. 40, no. 10, pp.
579–586, 2007. DOI: 10.3182/20070820-3-US-2918.00078

36. G. Park and S. B. Choi, “Clamping force control based on dynamic model estimation
for electromechanical brakes,” Proceedings of the Institution of Mechanical Engineers,
Part D: Journal of Automobile Engineering, vol. 232, no. 14, pp. 2000–2013, Dec. 2018.
DOI: 10.1177/0954407017738394

37. C. Line, C. Manzie, and M. Good, “Electromechanical Brake Modeling and Control:
From PI to MPC,” IEEE Trans. Contr. Syst. Technol., vol. 16, no. 3, pp. 446–457, May
2008. DOI: 10.1109/TCST.2007.908200

38. C. F. Lee and C. M. Chris Line, “Explicit Nonlinear MPC of an Automotive

Electromechanical Brake,” IFAC Proceedings Volumes, vol. 41, no. 2, pp. 10 758–10 763,
2008. DOI: 10.3182/20080706-5-KR-1001.01824

39. C. F. Lee and C. Manzie, “Near-time-optimal tracking controller design for an

automotive electromechanical brake,” Proceedings of the Institution of Mechanical
Engineers, Part I: Journal of Systems and Control Engineering, vol. 226, no. 4, pp.
537–549, Apr. 2012. DOI: 10.1177/0959651811422164

40. S. Kwon, S. Lee, J. Lee, and D. Kum, “Accurate State Estimation for Electro-Mechanical
Brake Systems,” J. Electr. Eng. Technol., vol. 14, no. 2, pp. 889–896, Mar. 2019. DOI:
10.1007/s42835-019-00124-x

173
41. S. Formentin, G. Rallo, and S. M. Savaresi, “Data-Driven Clamping Force Control for
an Electric Parking Brake Without Speed Measurement,” in 2018 IEEE Conference on
Decision and Control (CDC). Miami Beach, FL: IEEE, Dec. 2018, pp. 5128–5133. DOI:
10.1109/CDC.2018.8619170

42. R. Hoseinnezhad, A. Bab-Hadiashar, and T. Rocco, “Real-Time Clamp Force

Measurement in Electromechanical Brake Calipers,” IEEE Trans. Veh. Technol., vol. 57,
no. 2, pp. 770–777, Mar. 2008. DOI: 10.1109/TVT.2007.906374

43. C. Jo, S. Hwang, and H. Kim, “Clamping-Force Control for Electromechanical

Brake,” IEEE Trans. Veh. Technol., vol. 59, no. 7, pp. 3205–3212, Sep. 2010. DOI:
10.1109/TVT.2010.2043696

44. Y. Li, T. Shim, D.-H. Shin, S. Lee, and S. Jin, “Effective Clamping Force Control for
Electromechanical Brake System *,” in 2020 IEEE/ASME International Conference on
Advanced Intelligent Mechatronics (AIM). Boston, MA, USA: IEEE, Jul. 2020, pp.
643–648. DOI: 10.1109/AIM43001.2020.9158796

45. J. Brecht, A. Elvenkemper, J. Betten, U. Navrath, and J. B. Multhoff, “Elastic

Properties of Friction Materials,” in 21st Annual Brake Colloquium & Exhibition, Oct.
2003, pp. 2003–01–3333. DOI: 10.4271/2003-01-3333

46. S. Saric, A. Bab-Hadiashar, and R. Hoseinnezhad, “Clamp-Force Estimation for a

Brake-by-Wire System: A Sensor-Fusion Approach,” IEEE Trans. Veh. Technol., vol. 57,
no. 2, pp. 778–786, Mar. 2008. DOI: 10.1109/TVT.2007.905251

47. G. Park, S. Choi, and D. Hyun, “Clamping force estimation based on hysteresis
modeling for electro-mechanical brakes,” Int.J Automot. Technol., vol. 18, no. 5, pp.
883–890, Oct. 2017. DOI: 10.1007/s12239-017-0086-5

48. W.-F. Xie, “Sliding-Mode-Observer-Based Adaptive Control for Servo Actuator With
Friction,” IEEE Trans. Ind. Electron., vol. 54, no. 3, pp. 1517–1527, Jun. 2007. DOI:
10.1109/TIE.2007.894718

49. R. Schwarz, R. Isermann, J. Böhm, J. Nell, and P. Rieth, “Clamping Force Estimation
for a Brake-by-Wire Actuator,” in International Congress & Exposition, Mar. 1999, pp.
1999–01–0482. DOI: 10.4271/1999-01-0482

50. S. Saric, “Development of an Intelligent Perception System for an Automotive Brake-by-

Wire System,” Ph.D. dissertation, Swinburne University of Technology, 2009.

51. Y. F. Fu, X. H. Hu, W. R. Wang, and Z. Ge, “Simulation and Experimental

Study of a New Electromechanical Brake with Automatic Wear Adjustment
Function,” Int.J Automot. Technol., vol. 21, no. 1, pp. 227–238, Feb. 2020. DOI:
10.1007/s12239-020-0022-y

52. Bohm, Jurgen, S. Martin, B. Jurgen, B. Karlheinz, and S. Peter, “System for controlling
or adjusting an electromechanical brake,” U.S. Patent 6 279 694.

174
53. B. Francis and W. Wonham, “The internal model principle of control theory,”
Automatica, vol. 12, no. 5, pp. 457–465, Sep. 1976. DOI: 10.1016/0005-1098(76)90006-6

54. D. Simon, Optimal state estimation: Kalman, H [infinity] and nonlinear approaches.
Hoboken, N.J: Wiley-Interscience, 2006, oCLC: ocm64084871.

55. M. White, M. Tomizuka, and C. Smith, “Improved track following in magnetic disk
drives using a disturbance observer,” IEEE/ASME Trans. Mechatron., vol. 5, no. 1, pp.
3–11, Mar. 2000. DOI: 10.1109/3516.828584

56. C. Kempf and S. Kobayashi, “Disturbance observer and feedforward design for a
high-speed direct-drive positioning table,” IEEE Trans. Contr. Syst. Technol., vol. 7,
no. 5, pp. 513–526, Sep. 1999. DOI: 10.1109/87.784416

57. Xu Chen and M. Tomizuka, “Optimal plant shaping for high bandwidth disturbance
rejection in discrete disturbance observers,” in Proceedings of the 2010 American
Control Conference. Baltimore, MD: IEEE, Jun. 2010, pp. 2641–2646. DOI:
10.1109/ACC.2010.5531256

58. M. Tomizuka, “Zero Phase Error Tracking Algorithm for Digital Control,” Journal of
Dynamic Systems, Measurement, and Control, vol. 109, no. 1, pp. 65–68, Mar. 1987.
DOI: 10.1115/1.3143822

59. C. Kempf and S. Kobayashi, “Discrete-time disturbance observer design for systems
with time delay,” in Proceedings of 4th IEEE International Workshop on Advanced
Motion Control - AMC ’96 - MIE, vol. 1. Mie, Japan: IEEE, 1996, pp. 332–337. DOI:
10.1109/AMC.1996.509428

60. G. Raffo, G. Gomes, J. Normey-Rico, C. Kelber, and L. Becker, “A Predictive Controller

for Autonomous Vehicle Path Tracking,” IEEE Trans. Intell. Transport. Syst., vol. 10,
no. 1, pp. 92–102, Mar. 2009. DOI: 10.1109/TITS.2008.2011697

61. E. Kim, J. Kim, and M. Sunwoo, “Model predictive control strategy for smooth path
tracking of autonomous vehicles with steering actuator dynamics,” Int.J Automot.
Technol., vol. 15, no. 7, pp. 1155–1164, Dec. 2014. DOI: 10.1007/s12239-014-0120-9

62. P. Falcone, F. Borrelli, J. Asgari, H. E. Tseng, and D. Hrovat, “Predictive Active

Steering Control for Autonomous Vehicle Systems,” IEEE Trans. Contr. Syst. Technol.,
vol. 15, no. 3, pp. 566–580, May 2007. DOI: 10.1109/TCST.2007.894653

63. P. Falcone, H. Eric Tseng, F. Borrelli, J. Asgari, and D. Hrovat, “MPC-based yaw and
lateral stabilisation via active front steering and braking,” Vehicle System Dynamics,
vol. 46, no. sup1, pp. 611–628, Sep. 2008. DOI: 10.1080/00423110802018297

64. P. Falcone, M. Tufo, F. Borrelli, J. Asgari, and H. E. Tseng, “A linear time varying
model predictive control approach to the integrated vehicle dynamics control problem
in autonomous systems,” in 2007 46th IEEE Conference on Decision and Control. New
Orleans, LA, USA: IEEE, 2007, pp. 2980–2985. DOI: 10.1109/CDC.2007.4434137

175
65. N. Guo, X. Zhang, Y. Zou, B. Lenzo, and T. Zhang, “A Computationally Efficient
Path-Following Control Strategy of Autonomous Electric Vehicles With Yaw Motion
Stabilization,” IEEE Trans. Transp. Electrific., vol. 6, no. 2, pp. 728–739, Jun. 2020.
DOI: 10.1109/TTE.2020.2993862
66. Q. Cui, R. Ding, C. Wei, and B. Zhou, “Path-tracking and lateral stabilisation for
autonomous vehicles by using the steering angle envelope,” Vehicle System Dynamics,
pp. 1–25, Jun. 2020. DOI: 10.1080/00423114.2020.1776344
67. C. Xiang, H. Peng, W. Wang, L. Li, Q. An, and S. Cheng, “Path tracking coordinated
control strategy for autonomous four in-wheel-motor independent-drive vehicles with
consideration of lateral stability,” Proceedings of the Institution of Mechanical Engineers,
Part D: Journal of Automobile Engineering, vol. 235, no. 4, pp. 1023–1036, Mar. 2021.
DOI: 10.1177/0954407020946884
68. R. de Castro, R. E. Araujo, and D. Freitas, “Wheel Slip Control of EVs Based on
Sliding Mode Technique With Conditional Integrators,” IEEE Trans. Ind. Electron.,
vol. 60, no. 8, pp. 3256–3271, Aug. 2013. DOI: 10.1109/TIE.2012.2202357
69. Taehyun Shim, Sehyun Chang, and Seok Lee, “Investigation of Sliding-Surface Design on
the Performance of Sliding Mode Controller in Antilock Braking Systems,” IEEE Trans.
Veh. Technol., vol. 57, no. 2, pp. 747–759, Mar. 2008. DOI: 10.1109/TVT.2007.905391
70. H. B. Pacejka and I. Besselink, Tire and vehicle dynamics, 3rd ed. Oxford Waltham:
Butterworth-Heinemann Elsevier, 2012.
71. C. S. Ahn, “Robust Estimation of Road Friction Coefficient for Vehicle Active Safety
Systems,” Ph.D. dissertation, University of Michigan, Ann Arbor, Michigan, 2011.
72. M. Jonasson, J. Andreasson, B. Jacobson, and A. S. Trigell, “Global force potential of
over-actuated vehicles,” Vehicle System Dynamics, vol. 48, no. 9, pp. 983–998, Sep.
2010. DOI: 10.1080/00423110903243232
73. H. G. Bock, M. M. Diehl, D. B. Leineweber, and J. P. Schlöder, “A Direct Multiple
Shooting Method for Real-Time Optimization of Nonlinear DAE Processes,” in
Nonlinear Model Predictive Control, F. Allgöwer and A. Zheng, Eds. Basel: Birkhäuser
Basel, 2000, pp. 245–267. DOI: 10.1007/978-3-0348-8407-51 4
74. N. K. Gupta, “Frequency-shaped cost functionals - Extension of linear-quadratic-
Gaussian design methods,” Journal of Guidance and Control, vol. 3, no. 6, pp. 529–535,
Nov. 1980. DOI: 10.2514/3.19722
75. H. Peng and M. Tomizuka, “Preview Control for Vehicle Lateral Guidance in Highway
Automation,” Journal of Dynamic Systems, Measurement, and Control, vol. 115, no. 4,
pp. 679–686, Dec. 1993. DOI: 10.1115/1.2899196
76. R. Grandia, F. Farshidian, A. Dosovitskiy, R. Ranftl, and M. Hutter, “Frequency-Aware
Model Predictive Control,” IEEE Robot. Autom. Lett., vol. 4, no. 2, pp. 1517–1524, Apr.
2019. DOI: 10.1109/LRA.2019.2895882

176
77. L. T. Biegler, Nonlinear programming: concepts, algorithms, and applications to chemical
processes, ser. MOS-SIAM series on optimization. Philadelphia: Society for Industrial
and Applied Mathematics : Mathematical Programming Society, 2010.
78. D. P. Bertsekas, Dynamic programming and optimal control. volume 1, fourth edition ed.
Belmont, Mass: Athena Scientific, 2017.
79. M. Grötschel, S. O. Krumke, and J. Rambau, Eds., Online optimization of large scale
systems. Berlin ; New York: Springer, 2001.
80. M. Diehl, H. J. Ferreau, and N. Haverbeke, “Efficient Numerical Methods for Nonlinear
MPC and Moving Horizon Estimation,” in Nonlinear Model Predictive Control,
M. Morari, M. Thoma, L. Magni, D. M. Raimondo, and F. Allgöwer, Eds. Berlin,
Heidelberg: Springer Berlin Heidelberg, 2009, vol. 384, pp. 391–417, series Title:
Lecture Notes in Control and Information Sciences. DOI: 10.1007/978-3-642-01094-13 2
81. G. Frison and M. Diehl, “HPIPM: a high-performance quadratic programming
framework for model predictive control,” IFAC-PapersOnLine, vol. 53, no. 2, pp.
6563–6569, 2020. DOI: 10.1016/j.ifacol.2020.12.073
82. E. M. Gertz and S. J. Wright, “Object-oriented software for quadratic programming,”
ACM Trans. Math. Softw., vol. 29, no. 1, pp. 58–81, Mar. 2003. DOI:
10.1145/641876.641880
83. A. Zanelli, A. Domahidi, J. Jerez, and M. Morari, “FORCES NLP: an efficient
implementation of interior-point methods for multistage nonlinear nonconvex programs,”
International Journal of Control, vol. 93, no. 1, pp. 13–29, Jan. 2020. DOI:
10.1080/00207179.2017.1316017
84. J. M. Maciejowski, Predictive control: with constraints. Harlow, England ; New York:
Prentice Hall, 2002.
85. R. Cagienard, P. Grieder, E. Kerrigan, and M. Morari, “Move blocking strategies in
receding horizon control,” Journal of Process Control, vol. 17, no. 6, pp. 563–570, Jul.
2007. DOI: 10.1016/j.jprocont.2007.01.001
86. R. Quirynen, M. Vukov, and M. Diehl, “Multiple Shooting in a Microsecond,” in
Multiple Shooting and Time Domain Decomposition Methods, T. Carraro, M. Geiger,
S. Körkel, and R. Rannacher, Eds. Cham: Springer International Publishing, 2015,
vol. 9, pp. 183–201, series Title: Contributions in Mathematical and Computational
Sciences. DOI: 10.1007/978-3-319-23321-57
87. D. Liao-McPherson, M. M. Nicotra, and I. Kolmanovsky, “Time-distributed optimization
for real-time model predictive control: Stability, robustness, and constraint satisfaction,”
Automatica, vol. 117, p. 108973, Jul. 2020. DOI: 10.1016/j.automatica.2020.108973
88. T. Ohtsuka, “A continuation/GMRES method for fast computation of nonlinear
receding horizon control,” Automatica, vol. 40, no. 4, pp. 563–574, Apr. 2004. DOI:
10.1016/j.automatica.2003.11.005

177
89. K. Graichen and A. Kugi, “Stability and Incremental Improvement of Suboptimal MPC
Without Terminal Constraints,” IEEE Trans. Automat. Contr., vol. 55, no. 11, pp.
2576–2580, Nov. 2010. DOI: 10.1109/TAC.2010.2057912

90. L. Grüne and J. Pannek, “Analysis of unconstrained NMPC schemes with incomplete
optimization *,” IFAC Proceedings Volumes, vol. 43, no. 14, pp. 238–243, Sep. 2010.
DOI: 10.3182/20100901-3-IT-2016.00124

91. M. Diehl, H. G. Bock, and J. P. Schlöder, “A Real-Time Iteration Scheme for Nonlinear
Optimization in Optimal Feedback Control,” SIAM J. Control Optim., vol. 43, no. 5,
pp. 1714–1736, Jan. 2005. DOI: 10.1137/S0363012902400713

92. M. Diehl, H. Bock, J. P. Schlöder, R. Findeisen, Z. Nagy, and F. Allgöwer,

“Real-time optimization and nonlinear model predictive control of processes governed by
differential-algebraic equations,” Journal of Process Control, vol. 12, no. 4, pp. 577–585,
Jun. 2002. DOI: 10.1016/S0959-1524(01)00023-3

93. J. Andersson, “A General-Purpose Software Framework for Dynamic Optimization,”

Ph.D. dissertation, Faculty of Engineering, KU Leuven, Leuven, Belgium, 2013.

94. M. Alamir, Stabilization of Nonlinear Systems Using Receding-horizon Control Schemes,

ser. Lecture Notes in Control and Information Sciences. London: Springer London,
2006, vol. 339. DOI: 10.1007/978-1-84628-471-7

95. A. Murilo, M. Alamir, and D. Alberer, “A General NMPC Framework for a Diesel
Engine Air Path,” International Journal of Control, pp. 1–21, Apr. 2014. DOI:
10.1080/00207179.2014.905708

96. A. Murilo, R. Rodrigues, E. L. S. Teixeira, and M. M. D. Santos, “Design

of a Parameterized Model Predictive Control for Electric Power Assisted
Steering,” Control Engineering Practice, vol. 90, pp. 331–341, Sep. 2019. DOI:
10.1016/j.conengprac.2019.07.010

97. M. Alamir, A. Murilo, R. Amari, P. Tona, R. Fürhapter, and P. Ortner, “On

the Use of Parameterized NMPC in Real-time Automotive Control,” in Automotive
Model Predictive Control, M. Morari, M. Thoma, L. del Re, F. Allgöwer, L. Glielmo,
C. Guardiola, and I. Kolmanovsky, Eds. London: Springer London, 2010, vol. 402,
pp. 139–149, series Title: Lecture Notes in Control and Information Sciences. DOI:
10.1007/978-1-84996-071-79

98. K. M. Madhavan Rathai, M. Alamir, O. Sename, and R. Tang, “A Parameterized NMPC

Scheme for Embedded Control of Semi-active Suspension System,” IFAC-PapersOnLine,
vol. 51, no. 20, pp. 301–306, 2018. DOI: 10.1016/j.ifacol.2018.11.029

99. K. M. Madhavan Rathai, “Synthesis and real-time implementation of parameterized

NMPC schemes for automotive semi-active suspension systems,” Ph.D. dissertation,
Université Grenoble Alpes, Grenoble, France, 2020.

178
100. J. A. E. Andersson, J. Gillis, G. Horn, J. B. Rawlings, and M. Diehl, “CasADi: a
software framework for nonlinear optimization and optimal control,” Math. Prog.
Comp., vol. 11, no. 1, pp. 1–36, Mar. 2019. DOI: 10.1007/s12532-018-0139-4

101. H. J. Ferreau, C. Kirches, A. Potschka, H. G. Bock, and M. Diehl, “qpOASES: a

parametric active-set algorithm for quadratic programming,” Math. Prog. Comp., vol. 6,
no. 4, pp. 327–363, Dec. 2014. DOI: 10.1007/s12532-014-0071-1

102. S. Chang and T. J. Gordon, “A flexible hierarchical model-based control methodology

for vehicle active safety systems,” Vehicle System Dynamics, vol. 46, no. sup1, pp.
63–75, Sep. 2008. DOI: 10.1080/00423110701882306

103. “Laboratory Test Procedure For FMVSS 126, Electronic Stability Control Systems,”
U.S. Department Of Transportation, National Highway Traffic Safety Administration,
Standard, 2008.

104. “ISO 3888-1:2018 Passenger cars — Test track for a severe lane-change manoeuvre —
Part 1: Double lane-change,” International Organization for Standardization, Standard,
2018.

105. Y. Yoon, J. Shin, H. J. Kim, Y. Park, and S. Sastry, “Model-predictive active steering
and obstacle avoidance for autonomous ground vehicles,” Control Engineering Practice,
vol. 17, no. 7, pp. 741–750, Jul. 2009. DOI: 10.1016/j.conengprac.2008.12.001

106. S. Lee and H. E. Tseng, “Trajectory Planning with Shadow Trolleys for an
Autonomous Vehicle on Bending Roads and Switchbacks,” in 2018 IEEE Intelligent
Vehicles Symposium (IV). Changshu: IEEE, Jun. 2018, pp. 484–489. DOI:
10.1109/IVS.2018.8500498

107. J. Wurts, J. L. Stein, and T. Ersal, “Collision Imminent Steering at High Speed Using
Nonlinear Model Predictive Control,” IEEE Trans. Veh. Technol., vol. 69, no. 8, pp.
8278–8289, Aug. 2020. DOI: 10.1109/TVT.2020.2999612

108. S. Taherian, U. Montanaro, S. Dixit, and S. Fallah, “Integrated Trajectory Planning

and Torque Vectoring for Autonomous Emergency Collision Avoidance,” in 2019 IEEE
Intelligent Transportation Systems Conference (ITSC). Auckland, New Zealand: IEEE,
Oct. 2019, pp. 2714–2721. DOI: 10.1109/ITSC.2019.8917495

109. I. Batkovic, M. Zanon, M. Ali, and P. Falcone, “Real-Time Constrained Trajectory

Planning and Vehicle Control for Proactive Autonomous Driving With Road Users,” in
2019 18th European Control Conference (ECC). Naples, Italy: IEEE, Jun. 2019, pp.
256–262. DOI: 10.23919/ECC.2019.8796099

110. R. Hajiloo, M. Abroshan, A. Khajepour, A. Kasaiezadeh, and S.-K. Chen, “Integrated

Steering and Differential Braking for Emergency Collision Avoidance in Autonomous
Vehicles,” IEEE Trans. Intell. Transport. Syst., vol. 22, no. 5, pp. 3167–3178, May 2021.
DOI: 10.1109/TITS.2020.2984210

179
111. R. Quirynen, K. Berntorp, K. Kambam, and S. Di Cairano, “Integrated Obstacle
Detection and Avoidance in Motion Planning and Predictive Control of Autonomous
Vehicles,” in 2020 American Control Conference (ACC). Denver, CO, USA: IEEE, Jul.
2020, pp. 1203–1208. DOI: 10.23919/ACC45564.2020.9147820

112. U. Rosolia, S. De Bruyne, and A. G. Alleyne, “Autonomous Vehicle Control: A

Nonconvex Approach for Obstacle Avoidance,” IEEE Trans. Contr. Syst. Technol.,
vol. 25, no. 2, pp. 469–484, Mar. 2017. DOI: 10.1109/TCST.2016.2569468

113. R. Verschueren, S. De Bruyne, M. Zanon, J. V. Frasch, and M. Diehl, “Towards

time-optimal race car driving using nonlinear MPC in real-time,” in 53rd IEEE
Conference on Decision and Control. Los Angeles, CA, USA: IEEE, Dec. 2014, pp.
2505–2510. DOI: 10.1109/CDC.2014.7039771

114. A. D. Ames, J. W. Grizzle, and P. Tabuada, “Control barrier function based quadratic
programs with application to adaptive cruise control,” in 53rd IEEE Conference on
Decision and Control. Los Angeles, CA, USA: IEEE, Dec. 2014, pp. 6271–6278. DOI:
10.1109/CDC.2014.7040372

115. M. Graf Plessen, D. Bernardini, H. Esen, and A. Bemporad, “Spatial-Based Predictive

Control and Geometric Corridor Planning for Adaptive Cruise Control Coupled With
Obstacle Avoidance,” IEEE Trans. Contr. Syst. Technol., vol. 26, no. 1, pp. 38–50, Jan.
2018. DOI: 10.1109/TCST.2017.2664722

116. P. Wieland and F. Allgöwer, “CONSTRUCTIVE SAFETY USING CONTROL

BARRIER FUNCTIONS,” IFAC Proceedings Volumes, vol. 40, no. 12, pp. 462–467,
2007. DOI: 10.3182/20070822-3-ZA-2920.00076

117. A. D. Ames, X. Xu, J. W. Grizzle, and P. Tabuada, “Control Barrier Function Based
Quadratic Programs for Safety Critical Systems,” IEEE Trans. Automat. Contr., vol. 62,
no. 8, pp. 3861–3876, Aug. 2017. DOI: 10.1109/TAC.2016.2638961

118. A. D. Ames, S. Coogan, M. Egerstedt, G. Notomista, K. Sreenath, and P. Tabuada,

“Control Barrier Functions: Theory and Applications,” in 2019 18th European
Control Conference (ECC). Naples, Italy: IEEE, Jun. 2019, pp. 3420–3431. DOI:
10.23919/ECC.2019.8796030

119. W. Xiao and C. Belta, “Control Barrier Functions for Systems with High Relative
Degree,” in 2019 IEEE 58th Conference on Decision and Control (CDC). Nice, France:
IEEE, Dec. 2019, pp. 474–479. DOI: 10.1109/CDC40024.2019.9029455

120. Q. Nguyen and K. Sreenath, “Exponential Control Barrier Functions for enforcing high
relative-degree safety-critical constraints,” in 2016 American Control Conference (ACC).
Boston, MA, USA: IEEE, Jul. 2016, pp. 322–328. DOI: 10.1109/ACC.2016.7524935

121. H. K. Khalil, Nonlinear systems, 3rd ed. Upper Saddle River, N.J: Prentice Hall, 2002.

180
122. A. Wächter and L. T. Biegler, “On the implementation of an interior-point filter
line-search algorithm for large-scale nonlinear programming,” Math. Program., vol. 106,
no. 1, pp. 25–57, Mar. 2006. DOI: 10.1007/s10107-004-0559-y

181