Module 5

Robotics
5.1.1. INTRO:
 Robots are physical agents that carry out tasks by manipulating the physical
world.
 They are outfitted with effectors, which include wheels, grippers, legs, joints,
and other parts of their body, all of which serve the singular function of
exerting physical pressures on their surroundings.
 Furthermore, robots are built with sensors that access them to feel their
surroundings.
 Numerous sensors are used in modern robotics, including gyroscopes and
accelerometers for motion detection and cameras and lasers for
environmental measurement.
 The majority of modern robots can be classified into one of three main
groups.
 Robot arms, also known as manipulators, are fixed to their work
environment, such as the International Space Station or a manufacturing
assembly line.
 Easily the most prevalent kind of industrial robots are manipulators.
 A few automobile manufactures couldn't function without robotic
manipulators, and some have even been utilised to create original artwork.
 Several mobile manipulators are utilised in hospitals to support surgeons.
Industrial robotic manipulator

 The second model is the mobile.

 This robots have been employed for similar tasks like carrying containers at loading
ports and distributing food in hospitals. They move around their structure using legs,v
or other systems.

 Also referred to as UGVs, autonomous ground vehicles (AGVs) go on roads, trails,

and through cities.
  The planetary rover shown in Figure studied Mars for three months in 1997.

Planetary rover
 Another type of mobile robot commonly used in military operations, crop-spraying,
and monitoring are unmanned aerial vehicles, or UAVs.
 Deep sea research uses autonomous underwater vehicles, or AUVs.

 Mobile robots that roam the office vacuum floors and distribute things.

Unmanned air vehicles

• Because it combines mobility and manipulation, the third class of robots is often referred to
as a mobile manipulator.
• Humanoid robots imitate the human torso.
5.1.2. ROBOT HARDWARE :

 Sensors :Sensors serve as the perceptual interface between a robot and its
surroundings.

 There are two kinds of sensors:

 passive sensors, Active sensors, such as , put energy into the environment and
rely on the energy being reflected back to the sensor; similar to cameras, these
sensors are real observers of their surrounds since they pick up signals produced
by others.
 Activate sensors tend to be more informative than passive ones, but they also
require more power and can cause interference when used in conjunction with
other active sensors.

 Sensors, whether active or passive, fall into one of three categories based on:
 if they are able to perceive their surroundings,
 the robot's position, or
 its internal architecture.
 sensors were widely used in early robotics systems to measure the distance
between two objects. These sensors are termed range finders.
 is the preferred technology for self-driving underwater vehicles.
 When directed sound waves are emitted by sensors and are reflected by objects,
some of the sound may return to the sensor.

➔ Several cameras are used in stereo vision to capture photos of the scene from
slightly different angles. The range of objects in the scene is calculated by
analysing the parallax in these images.
➔ Since and stereo vision are unreliable, they are rarely utilised for mobile
ground robots. Instead, optical range finders are currently present on the
majority of ground robots.
➔ Similar to sensors, optical range sensors discharge active signals, such light,
and measure the amount of time it takes for the signal's reflection to return to
the sensor.
➔ Other range sensor types use laser beams and specialist 1-pixel cameras that
can be orientated by complex mirror or rotating component arrangements.
➔ These sensors are called scanning lidars, an acronym for light detection and
ranging.
➔ Scanlidars frequently have wider ranges and perform better in direct sunshine
when compared to time-of-flight cameras.
 ➔ Other widely used range sensors are radar sensors, which are frequently used by
unmanned aerial vehicles (UAVs) to estimate distances up to several km.

➔ Touch-sensitive skin, bump panels, and whiskers are examples of tactile sensors that
are at the other end of the sensory spectrum.

➔ One important subclass of sensors are location sensors.

➔ Most location sensors identify their location primarily through range sensing.
➔ The Global Positioning System (GPS) is the most widely used remedy for the
localization problem outside. GPS determines the distance between pulsed-signaling
satellites.

➔ There are right now 31 satellites in orbit using various frequencies to send out
signals.

➔ GPS receivers can calculate the distance to these satellites by looking at phase shifts.

➔ By triangulating the signals from multiple satellites, GPS devices can pinpoint an
exact location on Earth to within a few metres.

➔ Under ideal circumstances, Differential GPS makes use of a second ground receiver
whose position is known to provide millimetre precision.

➔ Unfortunately, GPS is not functional underwater or inside.

 Proprioceptive sensors, that enable the robot to see its own motion, make up the third
significant type.
 Shaft decoders, which count motor revolutions in tiny s, are frequently installed on
motors in order to precisely measure the configuration of a robotic joint.
 Accurate information may be obtained at any time from shaft decoders on robot
arms.
 Wheel revolutions reported by shaft decoders on mobile robots can be utilised for
odometry, or measuring distance travelled.
 Gyroscopes and other inertial sensors depend on mass's resistance to velocity
changes. They could aid in lowering uncertainty.
  Force and torque sensors monitor other significant aspects of the robot's status.
These are essential when handling delicate goods or objects whose precise form and
placement are unclear by robots.

(i) Effectors :
• Robots use effectors to move and change the shape of their bodies.
• The concept of a degree of freedom (DOF) aids in the abstract explanation of motion
and shape.
•For every independent direction a robot or any of its effectors may travel, we count one
degree of freedom.

 This will help us comprehend the design of effectors.

 A rigid mobile robot with six degrees of freedom, for example, is an AUV. It has
three degrees of freedom for spatial position (x, y, and z) and three degrees of
freedom for angular orientation.
The sixdegrees determine the kinematic condition, or stance, of throbot.

 A robot's dynamicstate consists of thesesix dimensions plus six more.

 Additionally, each joint in a robot has a single, double, or triple degree of

freedom. An object, like a hand, requires six degrees of freedom to be positioned
at a precise angle and location.

.The limb possesses precisely sixdegrees of freedom, which are produced by fiverotating
revolutejoints and one sliding prismatic joint.

5.1.3. ROBOTIC PERCEPTION :

Robots employ a process known as perception to convert sensor readings into
internal representations of their environment. Because sensors are noisy and the
world is frequently dynamic, unexpected, and only partially observable, perception is
 challenging.
Stated differently, robots encounter all the difficulties related to state estimation (or
filtering).

• In general, three traits characterise effective internal representations for robots:

These comprise for the robot to make deft decisions,

They are set up in a way that makes updating easier.

They are considered natural since their internalvariables correspond to those that
exist in the world in its unaltered state.

(i) Localization and mapping :

 Finding the location of items, included the robot itself, is the challenge of
localization;

 Effective physical interaction with the environment begins with a knowledge of

its surroundings.

 For example, robot manipulators require location information in order to handle

items, and robots themselves require location information in order to travel.

To maintain simplicity, let us examine a movable, slowly moving two-

dimensional robot, flat environment.

 Assume for the moment that the robot has access to a precise map of its
surroundings.
 A mobile robot's twoCartesian coordinates, denoted by x and y, and its heading,
represented by θ, determine its pose.
 Whenever we align thos three values in a vector, the specific determined by

Xt = (xt, yt, θt) .

So far so good.
 Monte Carlo localization, or MCL, is a type of localization that uses particle filtering.

 An example of the particle-filtering algorithm is the MCL algorithm.

 All we have to do is provide the relevant sensor and motion models.

 The robot uses the algorithm to determine its location within an office building.

 Global uncertainty regarding the robot's position is indicated by the uniform

distribution of particles in the first image, which is based on the previous.
 When the initial set of measurements is received, as shown in the second image, the
particles group together in the regions with the highest posterior belief.
 In the third, there are sufficient measurements to force every particle into one spot.
Sometimes there isn't a map of the surroundings available, so the robot has to find one
on its own.
 • It's a bit of a chickenandegg scenario: the navigatin robot must determineits current
location in reference to an incomplete map while simultaneously constructing this
map without fully comprehending itsactual location.
 This challenge is significant for manyroboapplications,
 (ii) Alternative forms of perception:
 Robots are not limited to localization or mapping;
 they can also sense other stimuli including temperature, scents, auditory signals,
and more.
 Several of these parameters can be calculated with dynamic Bayes network
variants.
 A robot can also be programmed to act as a reactive agent,

(iii).The role of machine learning in robot perception: When the optimal internal
representation is unknown, machine learning is especially crucial to robot perception.

 Unsupervised machine learning techniques are frequently used to project high-

dimensional sensor streams into environments that are lower in dimension.

 We refer to this method as low-dimensional embedding.

 Learning sensor and motion models from data and simultaneously identifying
appropriate internal representations are made possible by machine learning.

 Robots can also be made to Constantly adjust to significant shifts in sensor

readings thanks to another machine learning technology.

 Robots can adapt to these kinds of changes thanks to approaches for adaptive
perception.

 Self-supervised methods are those that allow robots to gather their own training
data (complete with labels!). Here, the robot makes advantage of machine
learning.

5.1.4. PLANNING TO MOVE :

At the end of the day, all a robot's decisions revolve around Methods for moving effectors.
The aim of the point-to-point motion challenge is to get the robot or its end effector to
the specified location.
•The compliant motion challenge is more challenging since it calls for the robot to
 move while physically touching an object.
• A robot moving a box over a tabletop and a robot manipulator screwing in a
lightbulb are two examples of compliant motion.
•First, we choose a suitable symbol that enables us to identify and address motion-
planning problems.
• Configuration space—the space of robot states indicated by joint angles, location,
and orientation—is a more practical workplace than the original 3D space.
The path planning challenge is to find a path in configuration space between two
configurations.

We have previously run across a number of iterations of the path-planning issue.

There are two primary methods.

▪ celldecomposition
▪ skeleton ization.

(i) Configuration space :

▪ To begin, we will create a basic model for a straightforward robot motion issue.
▪ Take into consideration the robot arm, which has two independently moving joints.
▪ The elbow and gripper's (x, y) coordinates change when joints are moved. (The z-
direction cannot be traversed by the arm.)

This implies that the configuration of the robot can be expressed as a four-
dimensional coordinate: (xg, yg) for the gripper's location and (xe, ye) for the
elbow's location in reference to the surroundings.

These four coordinates clearly describe the robot's whole state. They make up what's referred
to be a workspace representation, and collision checking works well with them.

Because of linkage limits on the space of accessible workspace coordinates, even in the
absence of obstacles, not all workspace coordinates are truly reachable.
 This is the difficulty with the workspace representation.

 Regretfully, configuration spaces are not without issues.

  The dilemma of Which to set up configuration space and workspacecoordinates arises
because a robot's mission is typically defined in workspac coordinates rather than
configuration space coordinates.

 It is easy to convert configuration spacecoordinates into workspacecoordinates by

applying a few basic coordinate transformations, which are trigonometric for revolute
joints and linear for prismatic joints.

 Kinematics is the name given to this series of coordinate transformations.

 Inverse kinematics is the problem of configuring a robot with a known effector

position in workspace coordinates.

 Compute the inverse kinematics, particularly for robots with a large number of
degrees of freedom.

(ii) Cell decomposition methods :

 Using cell decomposition, the first path planning method divides the available divide
into afinite number of adjacent regions calledcells.
 One essential characteristic of these regions is that the path-planning problem
may be solved using straightforward methods (e.g., travelling in a straight
line) inside a single region.

 Then, similar to search problems, the path planning problem turns into a
discretegraph-search problem.

 most basic cell breakdown is represented by a grid with consistent spacing.

 value of eachgrid cell with empty spaces, or the cost of the lowest trip from
thatcell tothe objective, is indicated by grayscale shading.
 To discover the shortest path, of course, we can also apply the A∗ method.

 There are certain benefits to this kind of breakdown:

 Firstly, it can only be implemented in low-dimensional configuration spaces.

 The issue of what to do with "mixed" cells comes up second.

 Thirdly, there won't be a smooth path across a discretized state space.

(iii) Modified cost functions :
 Just as everyone driven a vehicle knows that a parkingspace withone millimetre of
clearance on eitherside is not genuinely a parkingspace at all, we would prefer
solution routes that are resilient with respect to tiny motion problems.
This issue can be resolved by introducing a potential field, which is a function
defined over state space whose value grows with the proximity to the nearest
obstacle.

 It is possible to include the potential field as an extra cost element in the shortest path
computation.
The robot aims to reduce the length of its journey to the objective, on the one hand.

 Nonetheless, it minimises the possible function in an attempt to avoid impediments.

(iv) Skeletonization methods :

 Based on the concept of skeletonization, the second major family of path-planning
algorithms reduces the robot's free space to a one-dimensional representation, making
the planning problem simpler.
 This lower-dimensional representation is referred to as a skeleton of the configuration
space.

 A of the open space, or the collection of all points equally spaced from two or more
obstacles, serves as an example of skeletonization.
 In order to plan a path using a , the robot first modifies its current configuration to a
point on the graph. It can be demonstrated with ease that a straight line motion in
configuration space can always accomplish this.
 Next, the robot moves along the until it reaches the point that is closest to the target
configuration.
  At last, the robot exits the and makes its way towards the destination. Once more,
straight-line mobility in configuration space is required for this last stage.

 The probabilisticroadmap is a skeletonization technique that gives more routes than

the s, making it more effective in handling large areas.

5.1.5. PLANNING UNCERTAIN MOVEMENTS :

 Uncertainty is a fundamental aspect of robotics problems that is not addressed by any
of the robot motion-planning methods that have been covered thus far.
 Uncertainty in robotics stems from the partially observable surroundings and the
random (or unmodeled) outcomes of robot movements.
 Even when a robot's environment is precisely modelled, errors can still occur via the
usage of approximation techniques, such as particlefiltering, do not provide the robot
with a precise belief state.
 The majority of modern robots rely on deterministic algorithms, like path-planning
algorithms, for decision-making.
 To achieve this, it is customary to extract the most likely state from the probability
distribution generated by the state estimation algorithm.
 This method has the benefit of being entirely computational arranging routes via
configuration space

▪ Robust methods :
 In addition to probabilistic methods, so-called robust control methods can
also be used to manage uncertainty.

 A robust solution is one that functions regardless of the actual values that
occur, as long as they fall within the assumed interval.
  A robust technique is one that assumes a bounded level of uncertainty in
every component of a problem, but does not assign probabilities to values
within the allowed interval.

 The conformant planning approach is an extreme example of a robust

method. In this article, we examine a different dependable technique

 A robot arm is moved in close proximity to a stationary environment object

during fine-motion planning.
The primary challenge with fine-motion planning is the tiny size of the
necessary motions and the environment's pertinent properties.

5.1.6. MOVING :
Up until now, we have discussed how to plan movements, but not how to carry them out.
Specifically, plans produced by Dynamics and control :

 The concept of the state of dynamic, which expands a robot's kinematic state based
on itsvelocity.

 rate of change, and possibly even the momentary acceleration of a robot joint are all
recorded by the dynamic [Link] impact of forces on this rate of change is included
in the transition model for a dynamic state representation.

 Typically, these models are expressed using differential equations. These formulas
connect a quantity (such as a kinematic state) to the manner in which a quantity varies
with time (such as velocity).

 To get beyond the limitations of kinematic plans, a common solution is to use a

controller—an independent mechanism—to keep the robot moving in the desired
direction.

(i) Potential-field control :

 Potential fields can be utilised directly to generate robot motion, skipping the
path planning stage entirely.

 We presented them as an extra robot motion planning using the cost function.

 In order to accomplish this, we must specify an attracting force that pushes the
robot away from impediments and a repellent potential field that pulls it
 towards its goal configuration.

(ii)Reactive control :
 Thus far, We've got examined control choices that necessitate an environment
model for creating either a potential field or a reference path.
Here are a few issues with this strategy.

 First, it is frequently challenging to obtain models that are sufficiently precise,

particularly for robots with limited sensors or in challenging or remote areas

like the surface of Mars.

 Secondly, even in situations where we are able to create a model that is

accurate enough, these methods may not be feasible due to computational
challenges and localization errors.

 Reactive control in a reflex agent design makes greater sense in certain

situations.
Consider a robot with legs that tries to lift one leg over a barrier.

 We could programme this robot to follow a rule that instructs it to elevate its
leg to a modest height (h) and move ahead; if it encounters an obstruction, it
should move back and begin again at a higher height.

(iii) Reinforcement learning control :

 One especially fascinating type of control is based on the policy search form
of reinforcement learning.
  This work has had a significant impact in recent years since it has provided
solutions for difficult robotics challenges that were previously unsolvable.

 Autonomous helicopter acrobatic flight is one instance.

5.1.7. ROBOTIC ARCHITECTURES :

 A architecture is framework for organising that consists of programming languages

and tools, as well as an overarching philosophy for how programmes might be
combined.

 The choice of how to integrate model-based deliberative planning and reactive control
is faced by contemporary software architectures for robotics.

 Reactive and deliberate tactics have strengths and disadvantages that are essentially
orthogonal to each other.

Reactive control relies on sensors and is suitable for making instantaneous low-level
decisions. Nevertheless, it seldom produces a workable solution at the global level since
judgements made for global control rely on information that cannot be felt at the moment of
decision making.
Deliberate planning is a better option for these kinds of issues. Hybrid architectures are those
that incorporate both purposeful and reactive design elements.
(i) Subsumption architecture :

 A framework for putting together reactive controllers from finite state machines is
called the subsumption architecture (Brooks, 1986).

 Arcs can be tagged with messages that will be generated when traversing them and
that are sent to the robot's motors or to other finite state machines.

 These machines' nodes might have tests for particular sensor variables; if so, the
  • The result of such a test determines the execution trace of a finite state machine.
Furthermore, internal timers, or clocks, in finite state machines regulate how
long it takes to complete an arc.

 The that are produced are known as AFSMs, because they employ clocks as part of
their augmentation.

(ii) Three-layer architecture :

 Reaction and deliberation are combined in hybrid designs.

 The robot is controlled at a low level via the reactive layer. Tight sensoraction
loop characterises it. Its decisioncycle is frequently milliseconds in duration.

 The executive layer, also known as the sequencing layer, functions as a bridge
between the deliberative and reactive layers.

 It takes instructions from the deliberative layer and arranges them in a

reactive layer order.

 The deliberative layer generates global solutions through planning for difficult
problems; its decision cycle is typically in the range of minutes due to the
computational cost of developing such solutions.

 Models are used in the deliberative layer of the decision making process.
These models can use state information acquired at the executive layer and can
be supplied or learnt from data.

 Most robot systems contain variations of the three-layer architecture; however,

the decomposition into these three layers is not very strict.

 Other robot software systems have additional layers, such as user interface
layers that regulate human-robot interaction or a multiagent level that
synchronises a robot's actions with those of other robots operating in the same
environment.

(iii) Pipeline architecture :

 The pipeline architecture is another type of robot architecture.
 The pipeline design uses numerous processes to be executed in parallel, just like the
subsumption architecture. Nonetheless, this architecture's particular modules are
similar to those of the three-layer architecture.

 An illustration of a pipeline design used to operate an autonomous vehicle. This

pipeline receives data at the sensor interface layer.
▪ Based on this information, the perception layer then modifies the robot’s internal
models of its surroundings.
▪These models are then given to the planning and control layer, which modifies the
internal plans of the robot and converts them into actual commands.

The vehicle interface layer is then used to relay those back to the car.

Ethics and Safety in AI:

What are the limitations for AI?
• Dependence on Data Quality
• The quality of AI algorithms, especially machine learning models, depends on the
quality of the training data. Results that are skewed, erroneous, or even discriminatory
can result from incomplete or biassed data. This is especially troubling for sensitive
applications such as financial risk assessment, law enforcement, and healthcare. When
we think about the moral ramifications of AI decision-making, the limitations of AI
become even more apparent.
• Lack of Emotional and Social Understanding
• Artificial intelligence (AI) is still unable to fully comprehend human emotions,
cultural norms, and social cues, despite advances in Natural Language Processing
(NLP) and machine learning. This presents difficulties in fields like mental health
diagnosis and customer service that demand a thorough grasp of human behaviour.
• Ethical and Privacy Concerns
• Important ethical issues are frequently brought up by AI applications, such as those
involving algorithmic bias, informed consent, and data protection. These drawbacks
can occasionally outweigh the advantages, particularly when the technology is used
carelessly.
• Complexity and Resource Intensity
 • Complex AI algorithms may need a lot of resources, including specialised gear and
powerful computers. Due to their high energy consumption, this renders them
unaffordable for several small and medium-sized businesses and raises environmental
issues.
• Lack of General Intelligence
• Artificial intelligence (AI) is good at certain tasks (narrow AI), but it is not as good as
humans at transferring knowledge from one domain to another. Artificial General
Intelligence (AGI) that is capable of carrying out any intellectual work that a human
can is still a long way off.
Ethics of AI – Surveillance
Artificial Intelligence is a flawed technology that can lead to unintended and intentional
harm.
We must be aware of the ethical issues that can arise when integrating AI into our monitoring
and surveillance systems, as well as how to guarantee appropriate use and results.
security and privacy

Concerns regarding data security and privacy have gained prominence as AI technologies
develop and spread. These technologies often rely on vast amounts of data, raising questions
about how personal information is collected, processed, and used.
[Link] privacy violations: AI systems frequently use enormous volumes of data for training
and decision-making. Sensitive personal data, including biometric data, financial
transactions, and medical records, may be included in this data. People's rights to privacy
may be violated if this data is handled improperly or if unauthorised parties gain access to it.
2. Vulnerabilities in Data Security: AI systems are vulnerable to adversarial assaults, model
poisoning, and data breaches, among other security flaws. Significant threats to privacy and
security could arise from malicious actors using these weaknesses to steal confidential
information, tamper with AI-driven choices, or jeopardise the integrity and dependability of
AI systems.
3. Algorithmic Bias and Discrimination: In sensitive fields like recruiting, lending, and law
enforcement, AI algorithms may unintentionally reinforce bias (Inconsistent labelling) and
discrimination, resulting in unfair or discriminatory outcomes.
4. Surveillance and Tracking: Concerns about widespread surveillance and invasions of
people's privacy are raised by AI-powered surveillance technology like facial recognition
software and location tracking instruments.
5. Lack of Transparency: A lot of AI systems function as "black boxes," which makes it
difficult to read their decision-making processes or hold them responsible for their deeds. AI
algorithms that are opaque and difficult to understand might make people lose faith in the
results they produce, especially in situations where fairness and privacy are crucial
considerations.
Overcoming Challenges and Safeguarding Privacy in AI
Privacy by Design: Incorporate privacy concerns from the beginning when designing and
developing AI systems. Adopt a privacy-by-design strategy that gives data security and
privacy top priority at every stage of the artificial intelligence lifecycle, from gathering and
processing data to training and deploying models.
Establish explicit standards and procedures for the ethical usage and governance of data to
make that AI systems abide by the values of accountability, openness, fairness, and
nondiscrimination.
Measures for Security and Compliance:To prevent unauthorised access, manipulation, and
exploitation of AI systems and data, implement strong security measures such as encryption,
access controls, and safe coding methods. To guarantee compliance and reduce security risks,
follow applicable privacy laws and standards, such as GDPR, CCPA, and HIPAA. You
should also carry out frequent security audits and assessments.
Algorithmic Fairness and Bias Mitigation: To find and lessen biases in AI algorithms, use
methods like algorithmic auditing, fairness testing, and bias detection. To ensure fair
outcomes and safeguard privacy rights, make sure training data sets are representative and
diverse, and apply algorithmic fairness techniques.
Transparency and Explainability: By implementing strategies like decision traceability,
algorithmic transparency, and model interpretability—which allow users to comprehend the
reasoning behind decisions made—AI systems can become more transparent and explainable.
Give people a comprehensive understanding of the decision-making process behind AI-
driven systems so they can recognise, question, and rectify biassed or incorrect results.
Data Minimization and Anonymization: Reduce the amount of personal data collected and
kept just to the level required to meet particular AI goals. Protect individual privacy while
maintaining data value for AI applications by implementing data anonymization and
pseudonymization procedures. Adopt technologies that protect privacy, such differential
privacy and federated learning, to facilitate collaborative data analysis without sacrificing
privacy.
Fairness and bias
• A systematic mistake in decision-making processes that produces unjust results is referred
to as bias.

Bias in AI can originate from a number of places, such as human interpretation, algorithm
design, and data gathering.

AI bias can originate from several phases of the machine learning pipeline, such as data
collecting, algorithm design, and user interactions. These stages include data bias,
algorithmic bias, and user bias.

When machine learning models are trained on incomplete or non-representative data, data
bias results in biassed outputs. This may occur if the material was gathered from biassed
sources, if it is inaccurate, lacking critical information, or incomplete.
However, when the algorithms used in machine learning models contain built-in biases that
show up in their outputs, this is known as algorithmic bias. This may occur when algorithms
employ biassed decision-making criteria or are predicated on biassed premises.
When users of AI systems unintentionally or intentionally incorporate their own biases or
preconceptions into the system, it is known as user bias.
ACTUAL INSTITUTIONS OF BIAS IN AI
It was discovered that an AI system used in healthcare to forecast patient mortality rates was
prejudiced against individuals of African American descent.

Law enforcement organisations' use of facial recognition technology is another instance of

bias in AI systems.

According to research conducted by the National Institute of Standards and Technology

(NIST), face recognition