published by

International Federation of Robotics

Frankfurt, Germany
February 2022

POSITION PAPER

Artificial
Intelligence
in Robotics
 Artificial Intelligence in Robotics

Contents
Executive Summary....................................................................................... 3
Introduction .................................................................................................... 4
Definitions ...................................................................................................... 4
Artificial Intelligence ................................................................................ 4
Robots..................................................................................................... 6
Artificial intelligence in robotics ............................................................... 6
Applications ................................................................................................... 8
Sense and respond ................................................................................. 8
Performance Optimization .................................................................... 14
The impact of AI in robotics on work and jobs ...................................... 17
Future trends ............................................................................................... 17
Market adoption .................................................................................... 17
Research directions .............................................................................. 18
Safety standards and certification for AI in robotics .................................... 19
Safety .................................................................................................... 19
Certification ........................................................................................... 20
Conclusion ................................................................................................... 20
Case studies and videos ............................................................................. 21

Cover picture: MechMind

2 Information Paper | February 2022

 Artificial Intelligence in Robotics

 Performance optimization in which AI is

Executive Summary used to optimize process design and
robot programming as well as to improve
Artificial Intelligence (AI) holds great potential quality inspection and maintenance.
for robotics, enabling a range of benefits in
sectors as diverse as manufacturing and
healthcare. Though AI is already making its Pick-and-place applications are the most
mark on robotics, it is at a much slower pace widely adopted type of sense-and-respond
and in a far narrower field of application than is application in manufacturing and logistics. Pick-
commonly assumed. This paper summarizes and-place covers a broad range of applications
the most common applications of AI in robotics including palletizing, packaging and machine
currently in commercial use and provides an tending. Other manufacturing applications in
overview of market potential over the next 5 to very early stages of AI adoption are assembly
10 years. and welding.
The main aim of using AI in robotics is to better AI has considerable potential to speed up
manage variability and unpredictability in design and programming of robotic
the external environment, either in real-time, automation, though this is in early stages of
or off-line. This offers benefits for development. AI could help significantly lower
manufacturers, logistics providers and retailers the overall cost of the installation and re-
dealing with frequently changing products, programming of a robotic application – by 50%
orders and stock in so-called ‘high mix/low- according to some experts. This helps make
volume’ environments. It also helps robots to automation economically viable for small-to-
function in public environments – from medium sized manufacturers and larger
supermarkets to hospitals - which are inherently companies with high variability. The use of AI in
unpredictable. design and programming of robot applications
AI is not necessarily a prerequisite for dealing is still at a very early stage however. Currently,
with variability and unpredictability. For the most established performance optimization
example, simple pick and place applications application is robotic quality inspection. AI
with variance in product placement, but not in enables faults that may be undetectable for
the product itself, can be achieved without AI. humans to be identified at each stage of the
Also, robot mobility does not require AI. production cycle. This improves product quality
However, the greater the variability and and also minimizes waste as faulty parts can be
unpredictability of the environment, the more taken out of the line before being worked on
likely it is that AI algorithms will provide a cost- further.
effective and fast solution – for example for The manufacturing and logistics sectors are
manufacturers or wholesalers dealing with leading the use of AI in robotic applications.
millions of different products that change on a However, there are rapid developments in the
regular basis. AI is also useful in environments
use of AI in healthcare and professional service
in which mobile robots need to distinguish
assistance robots.
between the objects or people they encounter
and respond differently. Safety in robotic applications using AI is
currently ensured through hard-coded,
We differentiate between two main categories
deterministic algorithms that always take
of application:
priority, ensuring the robot stops when
• Sense-and-respond applications in which encountering an obstacle. This may change as
the robot identifies and responds AI becomes more ubiquitous in robotics, and
autonomously to its external environment in researchers are currently looking at ways to
a real-time closed loop enable the use of ‘trustworthy’ AI in safety-
critical components of robotic applications.

Position Paper | January 2022 3

 Artificial Intelligence in Robotics

Figure 1: Pick and place using AI, image credit: ABB.

factories and warehouses, are booming. These

Introduction environments are characterized by high levels
of variability and unpredictability, where AI can
There is much confusion around the role of offer significant benefits, enabling robots to
artificial intelligence (AI) in robotics. Images adapt to a changing external environment in
from Hollywood movies from ‘Transformer’ to real-time without recoding. AI is also starting to
‘Ex Machina’ portray a future in which robots be used in optimizing process design and robot
have achieved human-like intelligence and programming and is increasingly adopted in
physical capabilities. Terms such as ‘robotic robotic quality inspection.
process automation’ are used to describe This paper looks at the most common uses of
software programs that have no physical AI in commercial robotic applications and
robotic component. And there is no standard discusses market trends. It also discusses
definition of which algorithms constitute AI. safety standards and regulation for AI in
AI is still in its infancy in real-world applications robotics.
of robotics. Traditionally, automotive and
electronics manufacturers - the main adopters Definitions
of industrial robots – have looked for precision,
speed and predictability from their robot.
Achieving these characteristics does not Artificial intelligence
require AI. Increasingly however,
manufacturers operate in an environment of Existing definitions of AI vary in scope and there
high variability, producing small quantities of a is no standard. For example, researchers Stuart
large range of products, often ordered at short Russell and Peter Norvig 1 define AI as ‘an
notice. Re-programming robots each time to fit intelligent agent, where "Agent" means a
new production or packaging lines is costly. software system which perceives its
Meanwhile, robots have moved out of factories environment through sensors and acts upon
and into public domains such as hospitals and that environment through actuators and
supermarkets. And logistics robots, which "Intelligence" means the ability to select an
fetch-and-carry parts and packages around action that is expected to maximize a


1 Stuart Russell and Peter Norvig. Artificial Intelligence: A Modern Approach. Pearson Education, 3rd edition,
2009

4 Information Paper | February 2022

 Artificial Intelligence in Robotics

performance measure. The European performance optimization count as AI.

Commission, meanwhile, defines AI as Definitions have changed over time. Algorithms
‘Systems that display intelligent behavior by that were considered AI twenty years ago are
analyzing their environment and taking actions now not, leading to the joke among AI experts
– with some degree of autonomy – to achieve that ‘AI is just algorithms we don’t understand
specific goals’ 2. yet.’
There is also no agreement on which Within the category of machine learning, we
algorithms or methods constitute AI. Part of the refer in this paper to four types described
difficulty with setting a concrete definition is that below:
AI algorithms and models (such as neural  Supervised: The algorithm learns from
networks) are a combination of mathematical being presented with hundreds or
functions that, alone, could be considered
thousands of labelled examples of what it is
‘traditional’ algorithms. At a certain point, when
tasked with identifying.
these functions are combined – such as in a
 Semi-supervised: The algorithm trains
neural network - they achieve a level of
autonomy that can be characterized as AI, or itself from a combination of labelled
they could simply be considered as a examples and unlabeled examples.
combination of traditional mathematical  Self-supervised: The algorithm uses labels
functions. that already exist in the input data (for
example, specific sounds or colors) to
There is general agreement that machine
learning counts as AI. In contrast to a create a training model.
deterministic algorithm, a machine-learning  Reinforcement learning: The algorithm is
algorithm learns and improves itself through given a goal and is rewarded for making
experience. There is less agreement on which steps towards that goal and so learns how
algorithms for other areas such as reasoning to complete a task effectively through trial-
and decision-making, path planning and and-error learning.

Table 1: Robot definitions according to ISO 8373-2021 3.

Term Definition
Industrial  A automatically controlled, reprogrammable multipurpose manipulator,
Robot programmable in three or more axes, which can be either fixed in place or fixed
to a mobile platform for use in automation applications in an industrial
environment
 Industrial robots include the manipulating portion(s) of mobile robots, where a
mobile robot consists of a mobile platform with an integrated manipulator or
robot.

Service Robot  Robot in personal use or professional use that performs useful tasks for humans
or equipment. (Tasks in personal use include handling or serving of items,
transportation, physical support, providing guidance or information, grooming,
cooking and food handling, and cleaning. Tasks in professional use include
inspection, surveillance, handling of items, person transportation, providing
guidance or information, cooking and food handling, and cleaning.)


2 https://digital-strategy.ec.europa.eu/en/library/definition-artificial-intelligence-main-capabilities-and-scientific-
disciplines
3 ISO 8373:2021 - Robotics — Vocabulary

Position Paper | January 2022 5

 Artificial Intelligence in Robotics

Mobile Robot  Robot able to travel under its own control

 A mobile robot can be a mobile platform with or without manipulators. In addition
to autonomous operation, a mobile robot can have means to be remotely
controlled.

Robots Artificial intelligence in robotics

Robot definitions are provided by the The main aim of using AI in robotics is to better
International Standardization Organization manage variability and unpredictability in the
(ISO) in ISO 8373-20213 (see Table 1). IFR external environment, either in real-time, or off-
mainly distinguished between industrial and line.
service robots. Mobile robots are considered as
subset of service robots, including those used
in manufacturing settings.

Figure 2: Differentiation and overlap between Robotics and Artificial Intelligence2, 3.

It’s important to note that: software programs as well as hardware

 AI is not per se required to enable the such as sensors and cameras to execute
above tasks but can bring significant the required task. In contrast to software
performance benefits and enable specific applications such as recommendation
functionality. A general rule of thumb is that engines, AI is only ever one component of
the more variability and unpredictability a robotic application. This is particularly
there is in the external environment, the relevant to the question of robot safety,
more useful AI will be. which to date is generally governed by
 In real-world robotic applications, AI hard-coded, deterministic algorithms.
algorithms are combined with non-AI

6 Information Paper | February 2022

 Artificial Intelligence in Robotics

Background Information

Autonomy in Robotics
Until recently, most robots were hard-coded to ment – for example adjusting force and to a
execute a task according to a pre-defined flexible part or in a sanding process where the
trajectory and with a pre-defined level of force. amount of force required reduces during the
These robots are oblivious to their external process as the surface becomes smoother.
environment. This means that the object the robot
The past five years have seen strong growth in
works on (such as a car part) must always be
autonomous robots which are able to adjust to
presented in exactly the same position, and the
far greater variability in their external
robot is not able to adjust its force or motion – for
environment. For example, as we discuss in the
example to stop if something or someone gets in
IFR paper ‘A Mobile Revolution: How Mobility is
its way.
Reshaping Robotics’, autonomous mobile robots
Over the last decade, ‘co-bots’, with either built-in can not only stop if they encounter an object in
or add-on force-torque sensors, have given their path, they can also re-plan their route and
robots a limited ability to sense and respond to adjust their path in real-time. Autonomy does not
their external environment. For example, co-bots necessarily require AI. However, the higher the
can recognize movement in a designated zone level of autonomy, the greater the chances of AI
and adjust speed – or stop – accordingly. This algorithms being employed to categorize an
has enabled robots to be integrated into unfamiliar environment and to determine the
production lines alongside humans, with benefits best way to interact with that environment to
we discuss in the IFR paper ‘Demystifying achieve the application’s goal (for example
Collaborative Robots’. picking up a bottle from an unsorted bin and
placing it in a rack).
Force-torque sensors also enable co-bots to
adjust to minor variance in the external environ-
The IFR has defined the following five levels of autonomy in robotics:

THE FIVE LEVELS OF AUTONOMY IN ROBOTS

NO AUTONOMY, NO AUTONOMY, NO NO AUTONOMY BUT

SENSE AND RESPOND AUTONOMY ADVANCED AUTONOMY
REMOTE CONTROL SENSE AND RESPOND
DESCRIPTION

Robot can sense force

Robot is operated Robot is not aware of and motion and Robot can sense its Robot can identify
manually and its internal state or respond according to external environment characteristics of its
remotely external surroundings. pre-determined and calculate a external environment
response
parameters

Pick-and-place robot
EXAMPLES

Collaborative Assistant robot that

Remote-operated Weldingrobot in feeding a machine
assembly robot in responds differently to
surgical robot automotivefactory with parts from an
manufacturing a child than an adult
unsorted bin

Hard code overrides

Hard code overrides
Settings for force and probabilistic
SAFETY

User ensures safety Sensors on cages speed

probabilistic
algorithms
algorithms

NO AI DETERMINISTIC ALGORITHMS PROBABILISTIC ALGORITHMS HIGH AI

Information Paper | February 2022 7

 Artificial Intelligence in Robotics

Figure 3: Packaging boards for IKEA Slovakia, image credit: Photoneo.

correctly at the target destination. These

Applications algorithms may or may not be AI – or will most
likely be a combination.
In this section, we look at some of the most
common applications of AI in robotics in Generally, the more variability there is in the
commercial use. We differentiate between two input (‘sensing’) data, the more useful AI will
types of application: be in planning the appropriate response, such
as the correct path and force with which to
 Sense-and-respond applications in which approach and grip an object that is randomly
the robot identifies and responds placed. The algorithms for sensing and
autonomously to its external environment responding may run on the robot or robot
in a real-time closed loop: Pick-and-place system (for example, a robot and vision
applications are the most common robotic system), or in some form of higher control
applications in this category system, or in a mixture of these. If the
 Performance optimization in which AI is response does not require coordination with
used to optimize process design and robot other robots, the algorithm may run on the
programming, as well as in quality robot. However, in decisions involving
inspection and maintenance. multiple robots, the algorithm will run in the
control system. In mobile robots, for example,
fleet management systems are used to
Sense and respond control groups of robots. AI might be used in
these systems to determine which robot to
This covers applications in which the robot send to perform which task.
identifies and responds in real-time to its
external environment to complete a given Below are the most common sense-and-
task. respond applications.

The ‘sensing’ part requires on-robot and/or Pick-and-place

external sensors and may include cameras to
help the robot locate either itself, or an object As the name suggests, pick-and-place
it is tasked with manipulating. Machine applications require the robot to pick up an
learning algorithms of some form are typically object from one location and place it in
employed here. The ‘respond’ part generally another. This can be done with traditional,
includes some kind of reasoning and decision- deterministic, programming using exact
making algorithms to determine the best coordinates if the object to be picked is
course of action – such as the correct force always the same, arrives in the same position
and gripper position with which to pick up an and is placed in the same position, for
object, and the correct path to place it in example on a feed conveyor. Deterministic

8 Information Paper | February 2022

 Artificial Intelligence in Robotics

programming can also be used to pick objects There are a number of applications within the
that are the same but are unsorted. broad category of ‘pick-and-place’ which
require slightly different capabilities:
 Pick-and-drop: The robot picks items
from one unsorted bin and drops them into
another
 Put-wall: The robot picks objects from a
mixed bin, scans the barcode on the
object which indicates which cubbyhole it
should be placed in, and then places
accordingly
 Induction sorting: The robot picks
products or packages and places them on
Video 1: Machine tending at Allied Moulded with
ActiNav. an empty spot on a conveyor
 Order fulfilment: The robot picks items
For example, Universal Robots’ ActivNav from multiple bins or cartons and
system uses statistical matching to pick assembles them into the order
similar, but unsorted, parts from bins for
machine tending. The system uses a CAD For example, Obeta, an electrical supply
model of the part to be picked as input. wholesaler based in Berlin, has implemented
Service technicians assess the best pick AI from Covariant to compile orders from
points based on the object and the type of cartons on a conveyor. Obeta’s warehouse
gripper used – a vacuum gripper requires a receives hundreds of new products each
pick-point where suction can be applied over week, so the company needed a system that
the whole surface of the pick-point for could easily adapt to new items without
example. ActiNav uses a laser-range finder to disruption. Products come in different
produce a point cloud that enables packaging and quantities and have to be
identification the object. When an unsorted bin carefully gripped and packed into dispatch
arrives, a 3D picture is taken and a point cloud containers or cardboard boxes, both quickly
generated. Statistical matching determines and continually. Obeta has struggled to find
the likelihood of a potential object identified in enough workers, particularly during the
the point cloud being the object to be picked. COVID-19 pandemic, and the robot, which
Path-planning algorithms determine the best can pick 600 objects per hour, ensures
path for the robot arm to approach and pick continuity of operations.
the object.
The general rule that AI – in this case machine
learning – offers benefits as the level of
variability increases applies here. For
example, transparent, thin or flat objects are
difficult to identify as the 3D camera will
categorize them as flat surfaces rather than as
3D objects. AI also offers benefits in situations
in manufacturing, logistics and retail in which
parts (or parcels) to be picked are mixed and
unsorted in bins,or loaded differently each
time on pallets. AI also enables vision systems Figure 4: Order fulfilment at Obeta.
to correctly identify the outlines of tightly-
packed, similar objects such as packages of
the same color and surface or textured
surfaces such as wood. Finally, AI is useful in
environments with high variance in lighting
conditions.

Information Paper | February 2022 9

 Artificial Intelligence in Robotics

Background Information

Developments in machine learning drive adoption of

sense-and-respond applications
Machine-learning algorithms that power sense- The ultimate aim is to enable an algorithm to
and-respond robotic applications are advancing quickly generalize from what it has already
rapidly. Initially, algorithms had to be trained on learned, applying existing knowledge to
hundreds or thousands of images of each target recognizing and manipulating new objects. For
object in order to recognize it correctly. This is a example, the more exposure an algorithm has
time-consuming process that accounts for a to objects tagged or identified as ‘bottle’, the
substantial portion of the cost of the application. more quickly it can identify and pick bottles of
The algorithm must be retrained with tagged different shapes and materials that it has not
examples each time a new product is introduced. encountered before from a mixed bin, applying
This is not economically viable for companies with the right force and grasp. Training on different
high product variability – for example wholesalers objects, for example clothing, can also
with millions of products, many of which change strengthen the neural network (the underlying
on a daily basis. computing model) for bottles, increasing the
‘generalizability’ of the algorithm.
Significant development effort has been therefore
directed by AI software and systems providers to Though we will therefore see increasingly
enable algorithms to learn with fewer, or no, generalizable algorithms (powered by richer
tagged examples. This is achieved through semi- neural networks) this does not imply that AI-
supervised or self-supervised learning methods. driven robotic applications can be easily
Semi-supervised learning takes a small set of transferred from one environment – such as bin-
tagged images as a base, from which the picking – to another – such as welding. Robots
algorithm trains itself. Self-supervised learning is comprise many different hardware and software
more complex and uses labels that already exist components which must be integrated, and
in the input data (for example, specific sounds or robots are themselves typically integrated with
colors). The knowledge accrued by the system enterprise-wide systems such as fleet
integrator or software provider in their pick-and- management, enterprise-resource planning, or
place applications, and by the end-user, naturally manufacturing execution. The interfaces
expands over time, reducing the training and between components of the robot and external
installation time of an AI-based pick-and-place systems must all be programmed or adapted for
application for a specific domain, for example a new application.
packaged food.

Picking technology is advancing rapidly in described above, tailored to the specific

terms of both speed and reliability. To be application requirements.
economically viable, picking robots need to be
faster than human workers, while most Palletizing, de-palletizing and packaging
companies require an accuracy of around
Robots are used to take objects off pallets on
99.5%. It is still difficult for robots to pick
arrival, and load pallets and packages with
objects that are not rigid – for example, goods
objects for shipment. Robots have
in plastic wrapping, fruit and vegetables in
traditionally been used to palletize and de-
nets, or floppy materials – with an accuracy
palletize objects of standard sizes, while
and speed that is commercially viable, but we
employees unpack and load non-standard
can expect this to change over the next 5
goods. Using AI algorithms for object
years.
recognition and path planning, robots are
Additional robotic applications using AI use starting to be used for palletizing and de-
elements of the pick-and-place functionality palletizing of non-standard goods. Palletizing
objects of variable weight is particularly tricky

10 Information Paper | February 2022

 Artificial Intelligence in Robotics

as it requires the algorithm work out the best more sophisticated convolutional neural
order of placement, so that heavier items, network (CNN) algorithm can be used to
even if smaller, are loaded on the bottom layer recognize boxes with problematic surfaces,
of a mixed pallet. including varying textures, shiny or reflecting
For example, LQ Group, a fast-moving material, protruding tapes, patterns, or
consumer goods company, is using AI cartons with black covering, which can also
technology from MechMind to de-palletize cause problems. A command is then sent to
goods and place them on a conveyor. LQ the robot which uses a specially developed
receives thousand kinds of cartons of various universal gripper enabling it to pick with an
sizes, colors, patterns, shades, which would accuracy of +-3 mm. This way, it is able to
be costly or impossible to program unload 1,000 boxes in our hour, with 99.7%
individually. In the MechMind solution, once pick-rate accuracy.
the pallet arrives at a certain position, the host
control system sends the details of the cartons
on the pallet including size, weight and
quantities to the robot. As soon as the robot
receives the arrival signal from the host
system, it triggers the camera to take a picture
and sends the picture to the vision control
system, which combines the coordinates and
poses of cartons into grasping points. These
are then processed by AI algorithms which
send instructions back to the robot’s gripper
including how many suction cups have to be
activated in order to pick a specific number of
Video 4: De-palletizing using an ABB robot and
cartons, which specific cartons have to be Photoneo 3D vision system with machine learning
picked on the same layer and the coordinates
for the position of the box when it is put on the Machine tending
conveyor (https://ifr.org/case-studies/smart-
3d-vision-solutions). AI can be used to enable robots to feed parts
from unsorted bins to machines and is
particularly useful when there is high
variability in the application environment, for
example lighting, bin sizes and position and
mixed bins.

Welding
Industrial robots are widely used for welding
when the parts are the same and the weld
points therefore in the same position.
Video 2: AI-assisted palletizing and de-palletizing However, the automation of welding has
with Mech-Mind AI solution.
traditionally been too costly for manufacturers
Similarly, vision systems provider Photoneo with high part or product variability. Machine
uses machine learning for de-palletizing learning can be used to identify welding
cartons of different sizes, that may be tightly points and weld path.
packed or have shiny surfaces. The cartons
can be picked individually, versus the Assembly
traditional process of de-layerization, where AI is also being trialed in the assembly
the pallet contents are picked up as one and process in manufacturing. Assembly tasks
placed on the conveyor. The pallet is scanned such as snapping parts into place, gluing
with a 3D scanner and the algorithm, which is parts or inserting a windscreen into a car, are
pre-trained on 5,000 box shapes and typically carried out with hard-coded
constantly learns new ones it encounters. A

Information Paper | February 2022 11

 Artificial Intelligence in Robotics

parameters but AI can be used to enable spaces, providing information and enabling
these tasks in variable environments. The remote interaction with specialists.
same technology described for pick-and-place Manufacturers and logistics providers with
and welding applications can be applied to high product variability and turnover ideally
identifying the coordinates for snapping a part want to move from fixed racks - where the
into place and planning the correct path and same product is stored in the same place - to
force for example. dynamic packing spaces where the size of the
Assembly tasks are generally more packing area, and the products within it, vary
complicated than pick-and-place and machine continuously and intelligent mobile robots are
tending, often involving multiple steps. It’s key to executing this vision.
likely that AI will first be applied to very specific Autonomous navigation means these robots
tasks - for example, enabling a robot to locate locate themselves within, and simultaneously
the exact position for placing a screw - within create or update, a map of their surroundings.
an assembly process. IFR members believe They also use path planning algorithms to
larger-scale adoption of AI in assembly will determine the best route to their target
take between five and ten years. destination. AI is not needed for autonomous
navigation or path planning but AMR
Recycling providers such as MiR and Fetch Robotics
Robots are starting to be used in the recycling are building AI into their AMRs to improve
of waste materials, improving productivity, and obstacle recognition and response.
saving workers from unpleasant and often AMRs use sensors to detect if an obstacle is
dangerous tasks. Some, such as those from in their path, in which case they will slow or
AMP Robotics, use AI to recognize objects in stop. The speed of slow-down or stop is the
a conveyor of mixed objects, for sorting same for all obstacles. Using 3D cameras and
according to type of waste. machine learning algorithms enables AMRs
to recognize the object in front of them and
tailor their response accordingly. For
example, the AMR would stop and give
priority to an automated guided vehicle that
might either hit it or cause an obstacle if it
stopped in the AMRs path. On the other hand,
it might slow down but not stop if a human
approaches and is able to walk round the
AMR. If it recognizes another AMR, it can
predict that vehicle’s motion and adjust its
Figure 5: Waste recycling, image credit: AMP
Robotics. own path, but not slow down.
Academic researchers are also looking at the
Mobile Robots use of machine learning to optimize path
As we discuss in our information paper, ‘A planning, but this is in early stages.
Mobile Revolution: How mobility is reshaping
Precision agriculture
robotics’ there is currently a boom in mobile
robots, particularly autonomous mobile robots There is a nascent market for robots in
(AMRs) which navigate autonomously and are commercial agriculture. Vision systems using
used in a wide variety of industry sectors machine learning algorithms are a key
including manufacturing, logistics, retail and component of many agricultural robotic
healthcare. These robots perform a range of systems. There are a number of start-up
tasks, from fetching and carrying goods and companies in precision agriculture focused
parts in manufacturing, transporting linens on planting, weeding, watering and the
and medicines in hospitals, tracking stock in targeted application of fertilizer and pesticide.
supermarkets and serving as ‘mobile Machine learning is used to distinguish
assistants’ in shops, hospitals and public between weeds and plants so that only the

12 Information Paper | February 2022

 Artificial Intelligence in Robotics

weeds are sprayed by the robot. The extent to which AI is used in these
Manufacturers of these systems claim applications depends on the exact
precision spraying reduces carbon emissions functionality required. Even within one sub-
as it uses 95% less herbicide than task, AI algorithms are likely to be combined
conventional spraying, which covers the with traditional algorithms to achieve the
whole cultivated area 4. required sense-and-respond functionality.
For example, the following video shows the
Lio personal assistance robot which supports
care-home residents and staff in a number of
tasks such as greeting patients, grasping and
carrying objects, offering drinks, reminding
patients of, and accompanying them to,
upcoming appointments.

Figure 6: The Avo precision weeding robot from

Ecorobotix.

The use of robots to harvest produce is still in

very early stages but AI plays a key role.
Harvesting applications are a form of pick and
place except the produce is no longer in a bin,
but unstructured on a plant. Machine learning
is used to identify the produce to harvest, as
well as to determine whether it is ripe. AI is
also used for path planning for the robot’s Video 3: Lio assistance robot
gripper, enabling it to pick the produce with the
These tasks use AI algorithms as part of
right grasp and force and place it in a bin or
larger programs to accomplish specific tasks.
other container.
For example, the task of serving a drink
comprises a number of sub-functions
Assistance robots
including: navigation; obstacle avoidance;
Robots are increasingly making an object recognition to distinguish between a
appearance in public environments such as person and a table; facial recognition to serve
airports, supermarkets and shopping malls, the right drink to the right resident; placement
hospitals and care homes, providing a range tasks – recognizing and locating the table on
of assistive functions. Robots in retail spaces which the drink is to be placed, locating where
can provide production information and lead to place and planning the path of the robot
shoppers to the right aisle for the product in arm and; general behavioral tasks such as
question, for example. In airports, robots can recovering from unplanned obstacles or
update passengers on gate information, and activity.
also guide them where necessary through the F&P Robotics, which has developed the Lio,
airport. In hospitals, robots can enable remote uses AI within some but not all of these sub-
consultations and undertake simple diagnostic functions. For example, AI algorithms are
procedures such as measuring temperature or used within the object recognition model.
blood oxygen levels. These applications all Open-source, pre-tagged object models are
involve mobile robots and are described in used to train the algorithm to recognize
more detail in the IFR paper ‘A Mobile common objects such as a person and a
Revolution: How mobility is reshaping table, while others are trained specifically for
robotics’. the application, and the robot is programmed
to ignore objects it cannot recognize. AI is


4 See https://www.ecorobotix.com/en/avo-autonomous-robot-weeder/

Information Paper | February 2022 13

 Artificial Intelligence in Robotics

used for facial recognition, trained on video (CAD) programs of the machines involved.
data of the resident to ensure the input to the Machines, including robots, generally perform
algorithm comprises images from different differently in the real-world to in model
angles. The Lio also has speech synthesis environments due to external effects such as
and voice recognition capability, using AI to vibration from other machines in the
enable the robot to understand the instruction production cell. Traditionally, programmers
given. The placement function – which bears have to observe simulations when transferred
some similarity to the pick-and-place to the real world and adjust the simulation in
functionality described above – also uses AI an iterative loop until the real-world
for placement identification, but not for applications works properly. AI can be used
planning the path of the robot arm to the in designing the initial simulation to predict the
identified spot on which to place the drink. AI optimal robot path and other parameters, and
is used extensively in the behavioral model then to identify the best performing simulation
which enables the robot to respond from multiple prototypes. Simulations can be
appropriately to the situation it encounters. For run in parallel, and speeded up, providing the
example, if the robot identifies a crowd of AI algorithms with far more data, much faster,
people, it might say something, but not if it than could be achieved by relying only on
cannot identify a person in the vicinity. If the data from real-world tests. This reduces the
robot approaches a door, AI algorithms are number of iterations and fine-tuning required
used to enable it to identify whether the door in real-world testing. Companies can also
is open and then take the correct course of simulate multi-robot systems in which AI is
action – opening it if closed, for example, used to optimize robot programs based on the
starting with locating the handle. data from all of the robots in the system.
In many of the above tasks, some of the Using AI for simulating automation scenarios
required functionality could be ‘hard-coded’ – is still at a very early stage but holds potential,
for example, the question of how to respond to both in industrial robot applications and also
encountering a crowd of people could be in service robot applications, where human-
coded through traditional ‘if/ else’ algorithms. robot interaction can be simulated in a
However, AI can provide greater flexibility for completely risk-free environment.
classifying a given situation and determining
the appropriate response. The choice of an AI
algorithm is therefore specific to the task and
influenced by the level of complexity in the
environment and the number of potential
response options.

Performance optimization
AI is used to optimize process design, robot
programming and robot maintenance, as well Figure 7: Order picking simulation, image credit:
as in robotic quality inspection. In general, MechMind.
these applications run off-line – the data is
collected from the robot and other machines, Robot programming
analyzed, and then the robot program is
Experts see promise in using AI to help
adjusted. We explore these in more detail
program robots. Programming and
below.
integration account for 50-70% of the cost of
Process design a robot application. Re-programming costs
have traditionally made robotic automation
Simulation is often used to design optimal too costly for many manufacturers with short
automation processes before they are production runs of a wide range of products.
implemented. Simulation programs are Methods to enable faster programming and
typically based on computer-aided-design program re-use are therefore key to robot

14 Information Paper | February 2022

 Artificial Intelligence in Robotics

adoption in variable environments. Experts used in future to help match the right program
estimate that AI could halve the resources from a suite of applications to a task specified
required for programming. by the user.
While industrial robot applications requiring
extremely high levels of precision and speed
are typically coded by a programmer, many
collaborative robot applications that do not
have very high cycle time or precision
requirements can also be programmed by
demonstration.
The robot arm is moved through the steps to
be performed to create a program which is
then fine-tuned through a tablet-based Figure 8: Intuitive tabled-based creation and
interface with intuitive touch-based controls. optimization of record paths, image credit: Neura
AI can be used in the fine-tuning process to Robotics.
determine the required movements of the
robot to perform the task optimally. Ultimately, Optimizing existing programs
this means a robot operator with no robot AI is used to optimize robot programs once
programming skills will be able to teach a running. Data on the robot’s movement
robot from any manufacturer. collected from sensors on and around the
Some robot manufacturers are already robot can be analyzed with AI algorithms to
offering this functionality 5. Neura Robotics is detect the optimal movements to achieve the
trialing the use of vision systems to program task. For example, FANUC has introduced ‘AI
by demonstration. For example, Neura path control function’ to improve the precision
Robotics is working on an application for of cutting and welding with robots, adjusting
welding in which an operator simply points to for external factors such as vibration.
the weld locations and the robot’s algorithms FANUC’s AI path control function estimates a
detect the exact welds and plan the weld path. robot's path from an acceleration sensor and
This saves around a day’s programming time. the amount of deviation from the command
Additionally, once the algorithm has been path. It provides appropriate compensation to
trained, it can be applied to parts of different achieve high-precision circular and straight-
sizes and weld positions 6. Neura Robotics line paths – a task that previously required
also uses Natural Language Processing to hours of trial and error by skilled operators 8.
enable voice instruction. The algorithm is ABB is applying AI algorithms to the analysis
trained using a demonstration of the task in of presses and robot behavior in press and
which the demonstrator describes what they stamping lines to minimize equipment waiting
are doing – for example picking up a box of times. Using a holistic approach, a control
biscuits – so that the robot can execute this algorithm identifies bottlenecks and manages
task when given a verbal instruction. the start and stop times of robots and
AI algorithms for tasks such as for path presses, thus making lines more stable and
planning, collision detection and picking are predictable 9.
also embedded by some robot systems
providers, such as MechMind, into their
programming interfaces to enable code-free
or code-light programming 7. AI could also be


5 See for example Universal Robots’ ActiNav.

6 See http://neura-robotics.com/
7 See MechMind’s Mech-Viz Intelligent Robot Programming Environment
8 See https://www.fanuc.co.jp/en/product/new_product/2020/202003_aipathcontrol.html
9 See https://new.abb.com/news/detail/56162/innovation-highlights-2020

Information Paper | February 2022 15

 Artificial Intelligence in Robotics

Finally, AI can be used to analyze data from likely to be effective. Prescriptive

multiple robots performing the same task to maintenance applications are already used
optimize the robot program for all of them. by operators of large plants but are not yet
established in robotics.
Predictive maintenance
Quality inspection
As the name suggests, the goal of predictive
maintenance is to predict when maintenance Quality inspection is, alongside pick-and-
on a part or machine will be required before place, the most frequently adopted use of AI
the part or machine underperforms or fails. in robotics currently. Using 3D vision and
Figure 9: Automotive assembly with an ABB robot machine learning algorithms, robots can
and vision system. inspect parts during the production process.
Faulty parts can be taken out of the
Traditionally, maintenance is conducted either production line before they are moved on to
on a timed basis, regardless of the actual state the next station, saving costs in unnecessary
of the machine, or when a machine breaks further processing.
down. Machine downtime can cost large
It is difficult to code exact parameters that
manufacturers over a million dollars per hour.
would enable the detection of all instances of
In predictive maintenance, sensors on the
flaws such as scratches that are different
machine and externally, if necessary, capture
each time. AI algorithms, in contrast, can
data on a continuous basis about the
generalize from many different examples.
machine’s condition – for example, vibration,
The algorithm continues to fine-tune itself as
noise, and gear speed – which can be
it encounters more examples. In some cases,
analyzed to predict the timing and scope of
3D vision can detect product flaws that are
maintenance. Machine learning is used to
too small for the human eye to identify.
establish a model of ‘normal’ performance so
that variations from this norm can be The same functionality can be applied to
identified, flagged and addressed. Again, AI is error-proofing which reduces wastage by
not required per se to enable predictive detecting and correcting faults at the
maintenance – parameters for values beginning of a run, before a lot of parts are
representing acceptable ranges of produced. FANUC offers an error-proofing
performance could be hard-coded, for application 11 - shown in this video - that
example. However, AI brings benefit in checks, for example, whether a nut that a
identifying these parameters where they are robot should tighten is actually in place.
not known exactly, and when they may vary in
different situations such as external
temperature.
A number of robot manufacturers, such as
ABB 10 offer predictive maintenance services
(sometimes also termed condition-based
maintenance) for their robots.
A more recent development is prescriptive
maintenance in which AI is used to work out
what is most likely to be wrong once a
potential maintenance issue or other
performance anomaly has occurred.
Prescriptive maintenance uses predictive Video 4: Error-proofing with FANUC application.
algorithms to determine the solution most


10 See https://new.abb.com/news/detail/73275/prsrl-abb-launches-condition-based-maintenance-service-for-fleet-and-
individual-robot-assessments
11 See https://www.fanucamerica.com/news-resources/articles/new-artificial-intelligence-error-proofing-features-machine-
learning-technology

16 Information Paper | February 2022

 Artificial Intelligence in Robotics

Some forms of quality inspection are difficult robot providers, was launched in January
to automate – particularly the quality of 2015 to accelerate the development of robotic
finishes such as polish or paint. Though inspection applications for production plants.
scratches can easily be detected using AI, it is
difficult to train an algorithm to distinguish
qualities such as overall smoothness, which The impact of AI in robotics on
looks different in different lighting conditions work and jobs
and for which human quality inspectors would
use touch. Digitalizing this kind of sensory or Despite initial headlines predicting the
haptic data to provide input to an AI algorithm number of jobs that will be lost to automation,
has so far proved difficult. We can therefore the majority of economic studies support a
expect final inspection of many finished very different narrative – that automation
surfaces to remain a manual operation for leads to a net increase in jobs within a country
now. and to an increase in gross domestic product
(GDP). It has also been recognized that very
Plant inspection few jobs can be fully automated in an
economically viable way. Rather, specific
tasks within them can be automated, altering
the nature of the job but not eliminating it in
the majority of cases. The changes in skills
requirements driven by automation apply at
all levels of the workforce.
AI expands the potential for robots to share
tasks or processes with workers, taking on
those parts of the task or process that are
unergonomic and repetitive, such as lifting,
fetching and carrying. These applications do
not necessarily require AI, but, as we have
discussed above, AI technologies enable the
Video 5: Mobile robot inspection at BASF. robot to work effectively in unpredictable or
rapidly changing environments.
A number of chemicals and oil & gas
companies such as BASF in Germany and See the IFR’s positioning papers on ‘The
Shell 12 are trialing applications for plant Impact of Robots on Productivity,
inspection using robots equipped with vision Employment and Jobs’ and ‘Robots and the
systems. A 3D camera is mounted on a mobile Workplace of the Future’ and ‘Next
robot (which can be legged, with tracks, Generation Skills: Enabling Today’s and
aquatic, or a drone), which moves through the Tomorrow’s Workforce to Benefit from
plant looking for anomalies such as leaking Automation’ for more information on how
pipes or gauges showing an abnormal robots affect workers and jobs.
reading. Subsea inspection using robots is
already well established, but the use of legged Future trends
or tracked mobile robots for plant inspection is
newer. ANYbotics 13 and Energy Robotics 14,
for example, offer AI-enabled plant inspection Market adoption
using legged robots. The Sprint Robotics
Collaborative 15, a consortium of chemical and As discussed, AI is currently most established
petrochemical companies and inspection in pick-and-place applications. While market


12 See https://www.shell.com/energy-and-innovation/digitalisation/digital-technologies/robotics.html
13 https://www.anybotics.com/anymal-autonomous-legged-robot/
14 https://www.energy-robotics.com/
15
See https://www.sprintrobotics.org/

Information Paper | February 2022 17

 Artificial Intelligence in Robotics

adoption is difficult to predict accurately, should complete as well as understanding

experts consulted for this paper foresee the voice commands.
following broad timeline for the adoption of AI Another important research area is semantic
in commercial robotic applications: intelligence aimed at enabling the robot to
Next 5 years understand the properties and characteristics
 Pick-and-place applications of the object or person it is interacting with
 Machine-tending and make appropriate decisions. For
example, a mobile assistance robot might in
 Welding
the future be able to respond differently to an
 Quality inspection
elderly person than it would to a child.
 Agriculture (precision application of
Semantic intelligence has most potential for
pesticide and fertilizer) applications in which mobile service robots
 Process optimization (e.g. fine-tuning path come into contact with objects and people in
planning based on external environment) public environments where it cannot be
 Medical screening: Robots with screening known in advance what and who the robot will
devices (for example ultrasound) can be encounter. However, it involves an extremely
used to scan the required body part, with complex series of steps. The robot must first
AI algorithms supporting the robot’s path identify the properties of the object - for
planning, as well as analyzing results of the example, person who is a child - then detect
scan the object or person’s intention – for example
 Recycling passing by or approaching with the intent to
interact – and then work out how to respond
appropriately. It will be decades before full-
5 to 10 years scale scenarios of this kind are commercially
 Assembly tasks in manufacturing viable. However, robot manufacturers are
 Clothing manufacturing already addressing first steps – for example,
 Agriculture (picking of crops) a mobile robot that can distinguish between
 Professional service assistant robots an object and a person therefore plan the
 Construction robots speed at which it has to slow down and
 Laboratory automation whether to issue a verbal warning if a person
is in its path. Semantic intelligence also has
potential in some industrial applications, for
Longer-term example plant inspection. A robot with
 Personal /household service assistant semantic intelligence would be able to identify
robots whether a pool of liquid was spilt oil or water,
 Quality inspection for functions requiring taking different actions, such as sending an
haptic feedback (e.g. smoothness) alert, if it identified oil.
Some research is focused on ‘swarm’
robotics, where robots coordinate their
Research directions movements among themselves, without input
There is a huge body of research directed to from a higher control system. Target
using AI to advance all of the functionality applications are ones in which humans are
discussed in this paper as well as enabling not in the area of operation, such as
new functionality. As discussed earlier, much excavation of disaster sites or agriculture,
research is being devoted to training AI rather than indoor industrial environments.
algorithms faster and enabling them to There are few if any commercial applications
generalize better so that the applications using to date that use swarm models with no central
them can be quickly and cost-effectively control.
adapted to new situations within the same
broad category. Research is also going into
enabling robots to generate their own program
by watching a demonstration of the task they

18 Information Paper | February 2022

 Artificial Intelligence in Robotics

Safety standards and deterministic layer responsible for actions

such as ensuring a robot stops if 10cm away
certification for AI in from any object, no matter what the AI
algorithm would otherwise determine to be
robotics the best course of action. This is almost
always the case in the robotic applications
There is currently much discussion on how to using AI in operation today. However, as
ensure AI applications do not lead to applications advance - especially in
undesirable outcomes, particularly bias in applications where the robot is physically
algorithms that power recommendations such interacting with a person whose position and
as who is a suitable job candidate. In most of movement are not pre-determined or in
these instances it is important to be able to unpredictable environments such as public
establish what data the algorithm was trained spaces - we can expect to see AI employed
on to ensure the data is representative as well in safety-critical components of the
as how the algorithm reaches its conclusions. application.
Much of the proposed regulation around AI –
such as the European Commission’s recent Enabling more variance in the robot’s reaction
‘Proposal for a Regulation on a European to safety-critical situations could also offer
approach for Artificial Intelligence 16’ – is productivity benefits in industrial
focused on ensuring fairness and environments. For example, if a mobile robot
‘explainability’ in AI algorithms. can identify that a worker is within a safety-
critical zone for the robot, but on the other
Robots are not currently used in situations that side of the robot arm, it could keep moving,
involve bias. The main concerns for robot whereas in a hard-coded application, it would
manufacturers and users when it comes to AI always stop. Research is being directed at
are safety and certification which we discuss how to ensure safety in these situations. For
below. Additionally, robot manufacturers using example, some research focuses on ‘learning
vision systems must be able to ensure that controllers’ which continuously process data
any data transferred from the vision system to determine the robot’s next move, followed
does not contravene data privacy laws, for by processing of the data generated as a
example, by enabling the identification of a result of that action, in order to determine the
specific individual. next move.
Currently, ISO safety standards exist for
Safety industrial robots 17 – including a technical
specification for industrial robots used in
Because the output of an AI algorithm is not collaborative applications 18 – certain service
known in advance, users may be robots 19 and mobile robots 20. These
understandably cautious about the impact on standards enable robot manufacturers to
safety. However, AI can never be used alone certify their robot as inherently safe. Since the
in a robot, as there are many other program safety of a robot application is dependent on
layers needed to control aspects of the many other factors – how well-lit is the
physical robot system – such as the production cell, is the robot manipulating a
movements of the robot’s axes. This means part with sharp edges – users must carry out
robot programmers have to determine which risk assessments of their applications and are
layer in the ‘stack’ has priority over others.
Priority can be given to the hard-coded,


16 See https://digital-strategy.ec.europa.eu/en/library/proposal-regulation-european-approach-artificial-intelligence
17 ISO 10218-1:2011 and ISO 10218-2:2011 - Safety requirements for industrial robots
18 ISO/TS 15066:2016 - Robots and robotic devices — Collaborative robots
19 ISO 13482:2014 - Robots and robotic devices — Safety requirements for personal care robots
20 ISO 3691-4:2020 - Industrial trucks — Safety requirements and verification — Part 4: Driverless industrial
trucks and their systems. Please also refer to ANSI/RIA R15.08 for safety of industrial mobile robots.

Information Paper | February 2022 19

 Artificial Intelligence in Robotics

responsible for the safety of their workers, the desired outcome will be met, which is
governed by national health and safety laws. difficult in the case of AI where the outcome
It is currently unclear how safety standards may not be known in advance, or possible to
and regulation will evolve for scenarios where explain after the fact.
AI is used in safety-critical components of the Beyond certification, understanding why an
robot application. algorithm reached a decision, and what data
was important in making the decision, could
also support better re-usability of algorithms
Certification for other purposes. Again, research is being
Certification is important for robot devoted to this. For example, it would be
manufacturers and systems integrators to possible – but computationally expensive – to
demonstrate that specific safety or other determine and monitor parameters for the
standards are met and, as a result, to neural network governing the robot’s motion,
differentiate themselves from non-certified ensuring the robot’s actions remain within a
competitors. pre-determined safety range, even when the
precise action that the robot will take is not
There is currently considerable debate and pre-determined or explainable. The
research on the possibility and benefits of Fraunhofer Institute for Manufacturing and
software certification or verification for Automation is devoting research to
programs partly or entirely composed of AI ‘dependable AI’ in robotics 21.
algorithms. Certification relies on certainty that

Figure 10: Image classification for supervised learning. Image credit: FANUC.

in environments with high variability. In

Conclusion manufacturing, AI is enabling the automation
of a number of tasks involving the picking,
Artificial intelligence opens up new placing and manipulating of objects, from
possibilities for robotic automation, particularly machine tending to assembly, that have


21 See https://www.ipa.fraunhofer.de/en/about-us/guiding-themes/ai/Dependable_AI.html

20 Information Paper | February 2022

 Artificial Intelligence in Robotics

previously only been possible to do manually. applications are important, particularly in

Employees are spared strenuous, heavy or enabling certification of applications using AI
unergonomic tasks. AI is also becoming well- algorithms.
established for robotic quality inspection. The notion of ‘general artificial intelligence’ in
AI-driven robotic applications bring greater which a robot would be able to apply learning
efficiency to logistics and retail, enabling from one task – such as opening a door – to
companies in these sectors to cope better with another - such as shutting a cupboard –
peaks in orders, high product variability and an without further input, as a human would be
often-unreliable labor supply. able to, is unrealistic in commercial
Robots are increasingly making their mark in environments. As we have noted in this
public domains, from hospitals to shopping paper, a robot application must be tailored not
malls. In future, AI will enable better only to a specific task, but also to the specific
interaction between robots and the people and environment. The AI algorithms used are
objects they encounter. AI will also help drive therefore ‘narrow AI’ and the application will
robotic automation in sectors such as always need to be programmed, even if it re-
agriculture which have not previously adopted uses portions of other applications or existing
robots. blocks of code.

AI will help reduce the resources and cost However, as we have discussed,
required to program and re-task a robot, programming is becoming easier, faster, and
opening up the possibility of automation to more intuitive. We can expect strides in
many companies for which automation has not reducing the overall time, and skills level,
previously been economically viable. needed to create a tailored robot application,
or re-task an existing one. AI will play an
However, it’s important to note that these increasingly important role in enabling faster
changes will take time. There are significant application development and re-tasking, and
advancements in enabling increasing will expand the range of tasks it is
generalizability in AI algorithms - such as in economically viable for a robot to perform.
the types of objects the robot can recognize This opens up the prospect of automation to
when performing a particular task. However, a new industry sectors, and to many small-to-
complete robot application involves many medium sized companies.
more program components and interfaces that
are specific not only to the task, but also to the
broader, company-specific automation Case studies and
architecture. For each application to be
commercially viable, the cost of automation
videos
has to be outweighed by productivity or other
Pick and place
gains. In many cases it will therefore be years
before developments in research labs gain  Bin picking solution with FANUC robot
widespread commercial adoption. and Soft Robotics gripper:
Artificial intelligence is attracting increasing https://youtu.be/N5fyTIv1WgQ
scrutiny from regulators and advocacy groups,  Order fulfilment at Obeta with Covariant
particularly regarding the question of ensuring AI: https://youtu.be/bQKgbWJrGFY
the avoidance of bias in AI software. In
robotics, the key issue is ensuring safety. In Palletizing and de-palletizing
most commercial robot applications using AI  Photoneo and ABB de-palletizing
today, the robot either does not come into
solultion: https://youtu.be/vYLGsphsG5E
contact with people, or uses deterministic
 Palletizing and de-palletizing with
algorithms that override any AI in order to
MechMind AI: https://youtu.be/aJhUUG4-
protect humans, for example to ensuring that
the robot stops if the distance from an object lr4
or person falls below a certain threshold.
Nevertheless, efforts to create frameworks
and models for ‘explainable AI’ in industrial

Information Paper | February 2022 21

 Artificial Intelligence in Robotics

Machine tending
 Machine tending at Allied Moulding with
Universal Robots: https://youtu.be/Fbne-
Hc9g2g

Recycling
 Waste recycling with AMP Robotics:
https://youtu.be/C1PEsXWl-ZM

Precision agriculture
 The Avo precision weeding robot from
Ecorobotix:
https://youtu.be/5vvQqqc1zHM
 Harvesting robot:
https://youtu.be/WPBnRcQ0NiU

Assistance robots
 The Lio assistance robot:
https://youtu.be/i47LFCMn_48

Predictive maintenance
 Mobile inspection at BASF:
https://youtu.be/NOcnTFEcISo

Robot programming
 Natural Language Processing from Neura
Robotics https://youtu.be/TorR-nV1EZw

Error-proofing
 FANUC error-proofing application:
https://youtu.be/RCvTyKwmk-g

22 Information Paper | February 2022