Autonomous

Vehicles

- Chandra Karthik (IMT2020106)

- Vamsi Krishna (IMT2020111)
How do autonomous vehicles work ?
Tasks involved

● Localization
● Dynamic Scene Understanding
● Path planning
● Control
● User interaction
Localization

● Localization in AVs refers to the ability of the vehicle to determine its precise position in
the world. AVs require precise knowledge of their position (latitude and longitude),
heading (orientation), and lane-level accuracy, with a maximum error of only a few
centimeters it is crucial for safe autonomous vehicle (AV) navigation.
● High-definition maps, containing static and dynamic data, are essential references for
AVs.
● Sensor fusion, combining GNSS, IMUs, cameras, LiDAR, radar, ultrasonics, WiFi, and
odometers, is vital for accurate localization.
● Map-matching techniques compare real-time sensor data to accurate maps, with
data-driven approaches increasingly used for improved accuracy.
● Achieving precise localization is challenging but fundamental for AV safety and
effectiveness in real-world scenarios, and V2I communication enhances real-time map
data access
Dynamic Scene Understanding

● Accurate positioning within digital maps is the initial step for autonomous vehicles.
● Detecting and understanding dynamic elements, like vehicles and pedestrians, using
sensor data is crucial.
● Segmenting static elements from the vehicle's perspective provides real-time
information.
● Predicting future behaviors and trajectories of road agents is essential for safe
autonomous driving.
● Complex behavior modeling requires data-driven and deep learning-based solutions.
Path planning

● Local Motion Planning: This stage involves planning the vehicle's immediate movements
and trajectory. It's based on a globally defined route with centimeter-level accuracy,
typically specified as a set of waypoints.

● Consideration of Traffic Rules and Signals: Local planning takes into account traffic rules,
signal states, and predictions of dynamic agent trajectories to ensure safe and smooth
driving.

● Behavioral Decision Making: Motion planning includes making decisions such as lane
changes, overtaking, obstacle avoidance, emergency braking, and navigating
intersections.

● Different Planning Methods: Various planning methods are used, including graph-search,
variational or optimization-based, incremental or sample-based, and interpolation-based
approaches. Additionally, end-to-end data-driven methods are gaining prominence for
motion planning.
Lateral/longitudinal Control

● Lateral/Longitudinal Control: This involves using feedback controllers to guide the

vehicle along the planned path and maintain the desired speed profile.
● Controller Purpose: The controllers aim to adjust the vehicle's actuators to align its actual
state with the reference path and speed, even in the presence of modeling errors and
uncertainties.
● Key Considerations: Controllers should prioritize robustness, stability, safety, and
passenger comfort during the vehicle's motion.
● Controller Types: Various closed-loop controller types are employed, including path
stabilization, trajectory tracking, and more recently, predictive control approaches. These
methods ensure the vehicle follows its planned motions accurately.
User interaction

● Human-Vehicle Interfaces: Designing effective interfaces is crucial for autonomous

driving to facilitate interaction and communication with various in-vehicle users:
● In-Vehicle Users: These interfaces cater to both backup drivers (Level 3) and passengers
(Levels 4 and 5). Communication with them involves explicit methods like audio and video
signals and implicit cues from the vehicle's motion patterns (speed, distance, time gap).
● Communication Modalities: For drivers and passengers, interfaces include audio, tactile,
visual, vibro-tactile feedback, and, more recently, natural language processing methods.
● In-Vehicle Perception Systems: This layer also incorporates perception systems within
the vehicle to detect the status of users, enabling effective communication and
interaction.
Arising Questions ?

➔ Is localization accuracy enough? How close or far are we from exiting or entering a
pre-mapped region (e.g. ODD)?

➔ How certain is the system about the current and future status of dynamic objects?

➔ Did the model correctly learn and generalize to unseen or rarely encountered situations?

➔ Is it possible to reach a minimal risk condition?

➔ Is it feasible to achieve the planned local trajectory?

➔ How confident can the user be that the system has understood his/her questions or
commands?
Risk management

● New Risks with Connected AVs: The deployment of connected Autonomous Vehicles
(AVs) introduces new risks. These include potential issues with software and algorithms,
connectivity and network failures, and cybercrime threats like hacking.

● Unknown and Emerging Risks: Due to the high innovation in AVs and the complexity of
their operating environment, there's a higher potential for unknown or undetectable
risks, especially after market launch.

● Liability Frameworks: Current liability frameworks don't expect AV manufacturers to

address risks not common in the industry at the time of production. Balancing liability and
insurance costs between consumers, injured parties, and AV producers is essential.
Risk management

● Risk Mitigation Measures: Measures to reduce negative impacts include establishing an

ethics review board, providing risk training, creating processes for reporting
vulnerabilities by third parties, and incorporating "redress by design" mechanisms.

● Assessment for Trustworthy AI: There's a need for a specific process to assess adherence
to the Assessment List for Trustworthy AI (ALTAI) in the context of AVs, similar to other
AI-based technologies.
a) Localization Risk Management:

● Ensuring precise vehicle positioning within digital maps.

● Managing errors and uncertainties in localization data.
● Addressing challenges in sensor fusion for accurate location.
● Handling updates and corrections to digital maps as needed.

b) Dynamic Scene Understanding Risk Management:

● Detecting and understanding dynamic elements like vehicles, pedestrians, and cyclists.
● Managing the risk of misinterpretation or false positives/negatives in object detection.
● Ensuring real-time and accurate scene segmentation.
● Handling dynamic scenarios such as complex traffic situations.

c) Path Planning Risk Management:

● Planning safe and efficient routes based on global and local factors.
● Managing risks related to unpredictable behaviors of other road users.
● Addressing uncertainties in vehicle trajectories and environmental conditions.
● Handling complex maneuvers, including lane changes and obstacle avoidance.
d) Control Risk Management:

● Managing the control of the vehicle in terms of lateral (steering) and longitudinal
(acceleration/braking) movements.
● Ensuring stability, safety, and passenger comfort during control actions.
● Addressing uncertainties in sensor data and actuator response.
● Handling emergency situations and abrupt changes in the environment.

e) User Interaction Risk Management:

● Designing effective human-vehicle interfaces for drivers, passengers, and external road users.
● Ensuring secure communication channels to prevent cyberattacks and data breaches.
● Managing the risk of miscommunication or misunderstanding between users and the vehicle.
● Implementing safeguards, such as ethical guidelines and vulnerability reporting processes, to
address potential risks in user interactions.
Case study-1

Sometime during August 15th, a Chevrolet Bolt

operated by Cruise drove itself into wet concrete near
Fillmore and Steiner streets on Golden Gate Avenue
in San Francisco.

Reports suggest that it was passengerless at the time

of the incident.

"Autonomous vehicles can't recognize freshly poured

concrete as an obstacle because they lack the ability
to perceive and understand the road environment."
Case study-2

On 17 th august 2023 2 robotaxis crashed in San Francisco

Both involved Cruise driverless cars and one of them collided

with a fire engine responding to an emergency.

Investigators said that, even though the driverless taxi had

the green light, it was supposed to yield to an emergency
vehicle.
Accident Reports

https://www.dmv.ca.gov/portal/vehicle-industry-services/autonomous-vehicles/autonomous-v
ehicle-collision-reports/

https://www.dmv.ca.gov/portal/file/waymo_08192023-pdf/

https://www.dmv.ca.gov/portal/file/waymo_012223-pdf/

https://www.dmv.ca.gov/portal/file/waymo_060523-pdf/

https://www.dmv.ca.gov/portal/file/waymo_091022-pdf/
Risk Assessment and proposed solutions

● Develop a dynamic Bayesian network-based model for real-time risk assessment in

autonomous vehicles, allowing them to perceive surrounding risk factors and quantify
potential risks in visually occluded areas. Continuously monitor the autonomous driving
system to address safety concerns and maintain a human safety officer's presence.

● Prioritize risk assessment through simulation before deploying autonomous vehicles to

detect potential failures and vulnerabilities in the system.

● Leverage computer vision and neural networks to analyze the safety and risk aspects of
autonomous vehicles, enhancing their perception and decision-making capabilities.
● Ensure that the risk assessment process is grounded in scientific principles, while
recognizing the importance of addressing questions related to trust, acceptability,
uncertainty, and politics in risk management.

● Develop a comprehensive risk assessment approach that encompasses both safety and
security perspectives to address emerging threats and vulnerabilities.

● Focus on advancing higher levels of automation, known as automated driving systems, to

reduce human involvement in crash-prone situations while emphasizing that drivers will
continue to share driving responsibilities for the foreseeable future.

● Maintain continuous safety monitoring by a designated safety officer, recognizing that

full automation may not be achievable for safety reasons in the near term.
How safe is safe enough for AVs

The results of available studies show a clear trend on the part of consumers, who consider that
AVs must be much safer than the average driver
How safe is safe enough for AVs

Considering that benefits of AVs such as new mobility services for more people or the freeing up
of urban public spaces, it is worth asking whether it would be possible to accept a safety value
equal to that of the average driver, or just a slightly better. Even a 10 percent safer than the
average driver can involve hundreds of thousands of lives saved

Therefore, it may be highly advisable for both industry and policy makers to calibrate the
message about the improved safety that AVs can bring, to counterbalance public opinion biases.
 Human agency and autonomy

● Human agency in Autonomous Vehicles (AVs) is tied to the principle of human autonomy. It
affects both acceptance (e.g., whether users choose to engage with AVs) and safety (preventing
misuse). New Human Machine Interfaces (HMIs) and external HMIs (eHMIs) are essential to
maintain an appropriate level of human agency.
 Human agency and autonomy

● Human agency, in the form of varying levels of human oversight, addresses ethical concerns
and mitigates risks associated with AI systems. Human oversight ensures that autonomous
functions do not compromise human autonomy or cause adverse effects.

● Efficient methods to measure and calibrate the sense of agency are crucial for the effective
functioning of AVs. The level of human oversight in AVs differs based on the automation level.

● Additionally, external road users may also exercise oversight, which can sometimes lead to
misuse due to the expectation that AVs will always stop. Effective interaction between AVs and
other agents (both inside and outside the vehicle) requires mutual awareness. AVs should be
able to understand and respond to the actions and intentions of other road users.
Human agency and autonomy

● Developing methods to accurately represent and communicate the operational status of

the AV to users, including when human intervention is necessary, is a critical area for
future research.

● Proper oversight of AVs will demand new skills, both in terms of initial training and skills
that develop through exposure and usage
Respect for Human autonomy
● In self-driving cars (AVs), a fundamental ethical principle is unwavering respect for
human autonomy. This means ensuring that both vehicle occupants and external road
users have complete self-determination. AVs should not subjugate, coerce, deceive,
manipulate, or act against human wishes.

● AV design should aim to enhance human driving skills and mobility, especially for
vulnerable groups. Instead of dictating actions, AVs should collaborate with human
drivers and maintain mechanisms for human oversight when needed.

● This approach fosters cooperative relationships between humans and AVs, prioritizing
technology that empowers rather than restricts. It aligns with ethics and promotes safer
road interactions.
Categories of Human Oversight in Autonomous Systems:

● Human-in-the-Loop: In this category, the autonomous system performs certain tasks but
requires human commands to continue. The human operator remains continuously connected
to the system, and the autonomous capacity is limited to specific tasks.

● Human-on-the-Loop: Autonomous systems can execute tasks independently, but a human

plays a monitoring or supervisory role. The human can intervene if the system encounters
issues. This setup involves cooperation to achieve objectives, with critical decisions still made by
the human operator.
● Human-out-of-the-Loop: In this category, every decision is made autonomously by the system.
Humans are not considered "operators" and have no direct involvement in decision-making.
Human oversight

● The human-vehicle interaction loop, often

called human-vehicle cooperation, involves a
shared control strategy where the driver (or
passenger) and the vehicle work together as a
team in driving tasks.

● Various cooperative models for human-vehicle

interaction in automated driving have been
proposed, going beyond the traditional
automation levels. These models offer a more
flexible approach to human oversight,
emphasizing the need for both automated
systems and drivers to learn how to cooperate
safely and efficiently in dynamic driving
environments.
Human oversight
● For Drivers (Levels 1-3)
○ Autonomous Vehicles (AVs) must continuously monitor the state of the driver.
○ Monitoring includes assessing driver engagement and distraction, encompassing
manual, visual, and cognitive distractions.
○ Distractions occur when the driver's hands are off the steering wheel, their eyes
are off the road, or their attention diverts from driving.
○ If the driver's attention is insufficient, the AV's automation system may stop to
prevent misuse, potentially with fatal consequences.

● For Passengers (Levels 4 and 5)

○ Continuous monitoring of passenger state is not a strict requirement for proper
system use.
○ However, monitoring passenger state can enhance the overall user experience in
AVs.
Human oversight

● Understanding Autonomous System State: Drivers and passengers should understand

the state and driving capabilities of the autonomous system. This requires defining
efficient ways to communicate system status, information, and reliability without
overwhelming users.

● Continuous information transmission can include map data, global routes, local
trajectories, detected agents, traffic signs, and more, aiding in understanding vehicle
behavior.

● Discrete systems, such as numerical or color-coded indicators, can be used to convey

confidence or reliability levels for the automation system or specific tasks.

● At Levels 4 and 5, cooperation in driving tasks becomes irrelevant as the automation is

highly capable, and human intervention is not expected.
Importance of Take-Over Requests (TOR):

● In cooperative driving environments, Level 3 automation, or transitions between

automation levels (e.g., Level 4 to 3), take-over requests (TOR) or requests to intervene
are crucial.
● TOR notifies the user that they need to resume manual control of the vehicle, either for
driving tasks or to ensure minimal risk conditions.

Variables Studied for TOR:

Two main variables have been studied for TOR, often in simulated environments:

● Communication Modality: TOR notifications can be visual, vibrotactile, or auditory.

● Reaction Time: Factors affecting reaction time include driver distraction level, type of
non-driving task, communication modality, and previous experience with TOR.

Performance-Based TOR: Tailoring TOR notifications to the driver's behavior has proven
effective in addressing this critical task.
Training and Education on Oversight:

● Proper training on how to exercise oversight is essential.

● Driver experience and prior knowledge are crucial at all automation levels for building
acceptance, trust, and effective use of the automation system.

● Training standards for drivers of vehicles with automated systems should be considered.

● At lower levels of automation (1 to 3), control of the AV is similar to conventional vehicles,

encompassing strategic, tactical, and operational tasks.

● At higher levels (4 and 5), passengers may need to control strategic and tactical tasks,
such as selecting destinations, changing routes, or requesting stops. Modalities for
exercising control vary.

● Educational programs and skills for drivers may need updates if self-learning processes or
system updates substantially change how human oversight is exercised.
Explainability
● The huge advances in AI in recent years have come at a price: the increased complexity
of AI systems improves performance, but worsens the ability of humans to understand
how results have been generated from inputs.

● Most AI systems, especially Deep Neural Networks (DNNs), are often challenging to
interpret due to their vast and interconnected parameters, making them "black box"
models. In contrast, expert systems and rule-based models (white box models) are
inherently designed on human knowledge and are more interpretable. The need for
interpretability has led to the development of eXplainable AI (XAI) approaches to open
the black box of complex decision-making systems.

● XAI, or eXplainable AI, is defined as the capability of an artificial intelligence system to

provide specific details or reasons that clarify and simplify its operation, making it
understandable to a given audience
Explainability
● Explainability in the context of human-machine interactions involves the machine's
ability to enable the human user to comprehend its logic. This understanding is based on
how the human perceives the relationships between inputs and outputs within the
system

● An explanation should be able to provide human-interpretable information about the

factors used in a decision and their relative weight. "Weight" refers to how a specific
input change influences the output, or how two seemingly similar inputs can produce
varying outputs, and vice versa.
The need and importance of eXplainable AI (XAI) systems for Autonomous Vehicles (AVs) can be
analyzed from three distinct viewpoints. They are:
● Internal (backup drivers and passengers) and external road users
● Technical and scientific communities, producers, developers, etc
● Regulators, vehicle type approval authorities and insurers
Internal (backup drivers and passengers) and external road users:

● Internal users (backup drivers and passengers) of Autonomous Vehicles (AVs) may require
explanations in various situations such as sudden stops, unexpected route changes, or abrupt
lane changes.

● External road users like pedestrians and cyclists also benefit from some level of explanation.

● The communication of these explanations to users will be facilitated through

Human-Machine Interfaces (HMIs) and electronic HMIs (eHMIs), aligning with the essential
requirement of ensuring human agency and oversight.
Technical and scientific communities, producers, developers:

● Initially, it was thought that a model's performance and transparency had an inverse
relationship, with a performance-focused approach potentially reducing system
transparency.

● However, advancements in eXplainable AI (XAI) have shown that a deeper understanding of a

system can reveal its shortcomings, making explainability a valuable tool for improving
performance.

● Explanations can provide technical insights into a model's existing limitations and
weaknesses.
Regulators, vehicle type approval authorities and insurers:

● The inclusion of explainability as a requirement in future safety certification procedures,

whether for vehicle type approval (homologation) or self-certification frameworks, can
significantly enhance assessments of compliance with safety, human agency, oversight, and
transparency standards.

● This benefit extends to auditors, accident investigators, and insurers who can also gain
valuable insights from explainable systems.
Tradeoff:
● Safety vs. Other Requirements Trade-off: Balancing safety with factors like traffic
efficiency, environmental impact, and economic considerations is a key trade-off as
autonomous vehicles become more common.

● Responsibility for Accidents: Ethical dilemmas arise concerning accident responsibility in

fully automated scenarios. Questions surround whether manufacturers, operators, or
vehicle owners bear moral responsibility.

● Benefits vs. Risks: Autonomous vehicles present a trade-off akin to medical treatments,
with benefits like reduced accidents and increased mobility balanced against risks like
technical failures, cybersecurity threats, and potential job displacement.
Tradeoff:
● Cost, Technology, and Strategy Balancing: Developing autonomous vehicles involves
trade-offs in cost, technology, and strategy. Achieving higher autonomy levels requires
advanced, costly technology, which must be weighed against cost-effectiveness and
feasibility.

● Programming Trade-offs: Programming decisions in autonomous vehicles involve

trade-offs between minimizing costs (e.g., avoiding expensive accidents) and ensuring
safety. Government regulations may be necessary to establish ethical and safety-oriented
programming standards.

● Personal Autonomy Concerns: As automation increases, personal autonomy in

transportation becomes crucial. It involves individual control and self-determination,
raising questions about user acceptance and the need for manual override capabilities.
 References

● Images and News articles from Google

● https://web.stanford.edu/~mossr/pdf/Autonomous_Vehicle_Risk_Assessment.pdf
● https://www.techtarget.com/searchenterpriseai/definition/driverless-car
● https://www.wired.com/story/robots-infrastructure-transport/
● https://ai-watch.ec.europa.eu/document/download/201c8d02-2ffc-42bb-92a9-0d
2cebb25ed2_en
● https://www.cbsnews.com/sanfrancisco/news/robotaxi-crashes-san-francisco-fo
cus-autonomous-vehicle-safety/
● https://www.dmv.ca.gov/portal/vehicle-industry-services/autonomous-vehicles/a
utonomous-vehicle-collision-reports/
● https://www.wired.com/story/robots-infrastructure-transport/