Self-Driving Vehicles

Self-Driving Vehicles
Road vehicles equipped with systems that automate some or all driving tasks
Category	Vehicle automation and automated driving systems
Area	Automotive engineering, robotics, artificial intelligence and control systems
Related concepts	Autonomous Navigation, Computer Vision, AI Safety

Self-driving vehicles, also called autonomous vehicles or automated vehicles in various contexts, are road motor vehicles equipped with technologies that automate some or all aspects of the driving task. These systems combine sensing, computation and vehicle control to assist a human driver or to operate a vehicle without continuous human control under specified conditions.

The expression "self-driving" has no single, universally applied meaning in everyday usage. Technical classifications distinguish driver-assistance systems, which require continuing supervision, from automated driving systems (ADS), which perform the complete dynamic driving task when operating within their intended conditions. Under the classification developed by SAE International, Levels 3–5 constitute automated driving systems, whereas Levels 1–2 are driver-support automation. Level 0 denotes no sustained driving automation.

A vehicle's capabilities depend on the particular system installed, the feature engaged and the conditions for which that feature was designed. The ability to operate without human steering in a restricted environment does not establish that the vehicle can operate autonomously on every public road or in every weather condition.

Historical development

Research into automated road vehicles draws on developments in robotics, machine perception, navigation and control engineering. Modern research was influenced by experimental vehicles and organized competitions that required machines to navigate increasingly complicated environments.

A notable series of events was organized by the United States Defense Advanced Research Projects Agency (DARPA):

• In March 2004, DARPA conducted its first Grand Challenge, involving autonomous ground vehicles attempting a desert route between California and Nevada. No participating vehicle completed the course.
• In October 2005, five vehicles completed the second Grand Challenge. Stanford University's vehicle, Stanley, finished first.
• In November 2007, the DARPA Urban Challenge introduced a more complex setting in which autonomous vehicles had to negotiate intersections, moving traffic and other road situations while observing applicable traffic rules. Carnegie Mellon University's Tartan Racing team won the competition.

These competitions contributed to research on autonomous navigation and vehicle decision-making. Subsequent academic investigations examined the architectures, perception systems and control techniques required for operation on public roads.

Levels of driving automation

SAE International's J3016 taxonomy distinguishes six levels of driving automation, numbered 0 through 5. The classification concerns the responsibilities allocated to the human driver and the driving automation system, rather than the sophistication of an individual sensor or the marketing name of a feature.

Level	Classification	System capability and human responsibility
0	No driving automation	The human driver performs the driving task. Warning systems and momentary safety interventions may still be present.
1	Driver assistance	The system provides sustained support for either steering or speed control. The human driver continuously supervises and performs the remaining driving tasks.
2	Partial driving automation	The system provides sustained support for both steering and speed control. The human driver remains responsible for continuous supervision and intervention.
3	Conditional driving automation	The system performs the driving task under defined conditions. A human fallback-ready user must respond when intervention is required.
4	High driving automation	The system performs the driving task and manages fallback within its defined operating conditions without depending on human intervention.
5	Full driving automation	The system performs the driving task under all driver-manageable on-road conditions, without requiring a human driver.

The descriptions above summarize SAE J3016_202609 and NHTSA's explanatory material; they do not reproduce the complete normative definitions.

Distinguishing assistance from automated driving

At Levels 1 and 2, a human remains responsible for supervising the roadway and responding to hazards. Adaptive cruise control and systems that provide sustained lane-centering assistance are examples of functions that can form part of driver-support automation.

At Level 3, an automated driving system performs the complete driving task while engaged within its permitted operating conditions. The human user must remain available to respond to a request to resume driving.

At Level 4, the system is responsible for the driving task and the necessary fallback within its operating domain. Level 5 removes the domain restrictions that distinguish conditional operation from full driving automation.

Automatic emergency braking and collision-warning systems do not, by themselves, establish a higher level of driving automation. SAE distinguishes these momentary safety functions from systems that automate driving tasks on a sustained basis.

The SAE classification is a functional taxonomy, not a performance rating or certification that a vehicle is safe for every circumstance.

Technical architecture

Automated driving systems generally combine environmental sensing, localization, perception, prediction, planning and vehicle control. The arrangement of these functions varies between system architectures and research approaches.

Environmental sensing

Automated vehicles may obtain information about their surroundings through several types of sensor:

• Cameras capture visual information used to identify road markings, traffic signals, vehicles, pedestrians and other features of a driving scene.
• Radar measures properties of detected objects using radio waves and can support the estimation of their distance and relative motion.
• LiDAR uses laser measurements to obtain spatial information about surrounding objects and structures.
• Positioning and motion sensors provide information that can contribute to estimates of the vehicle's location, orientation and movement.

Sensor configurations differ among implementations. Some systems combine multiple sensing technologies, whereas others employ different combinations of hardware and software.

Perception and object tracking

Perception systems process incoming observations to construct a representation of the nearby driving environment. Relevant tasks include detecting vehicles, pedestrians, cyclists, obstacles, lanes and traffic-control devices.

Object tracking associates detections over time and estimates the movement of relevant road users. Sensor fusion can combine information from different sources to support these functions.

Localization and mapping

Localization determines the vehicle's position and orientation relative to its surroundings or an appropriate reference frame.

Automated driving research uses techniques involving sensor observations, vehicle motion estimates and digital maps. Some systems use detailed three-dimensional representations of roads and their surroundings to support localization and planning.

Maps can complement vehicle sensors, but they may need updating when road layouts, traffic controls or construction conditions change.

Prediction and decision-making

Automated driving in traffic requires estimates of how nearby road users may move. Prediction methods may consider the positions and motion of vehicles, pedestrians and other participants.

Decision-making systems use available observations and predictions to determine suitable maneuvers. These may include remaining in a lane, yielding, stopping at an intersection or changing lanes where permitted.

Motion planning and vehicle control

Motion planning produces a path or trajectory for the vehicle while considering road geometry, obstacles, vehicle constraints and surrounding traffic.

Control systems translate planned motion into commands for steering, acceleration and braking. Feedback control is used to account for differences between intended and observed vehicle motion.

Machine learning

Machine learning is used in research on automated driving, particularly for perception and the interpretation of complex driving scenes.

Published research examines both modular systems, in which perception, planning and control are implemented as distinguishable components, and end-to-end approaches that learn mappings from sensor inputs toward driving outputs.

These are different approaches to system design. The use of machine learning does not, by itself, determine a system's driving automation level or establish its safety.

Operational design domain

An operational design domain (ODD) describes the conditions under which a particular driving automation feature is designed to function.

Relevant restrictions may concern:

• Types of roads, such as controlled-access highways or urban streets.
• Geographic areas or designated service zones.
• Vehicle speed or traffic conditions.
• Weather, visibility and illumination.
• Roadway infrastructure and other environmental conditions.

A system designed for a restricted operating domain is not necessarily capable of operating safely outside that domain.

For example, an automated vehicle designed for operation within a mapped service area may be classified at Level 4 without being capable of Level 5 driving. The defining distinction is not whether a person is seated inside the vehicle, but whether the automated driving system can perform its assigned task and fallback responsibilities under the specified conditions.

Safety engineering

Safety engineering for automated vehicles includes the identification of hazards, analysis of system behavior, development of safeguards and verification that specified safety requirements are satisfied.

Several standards address distinct aspects of automotive safety.

Functional safety

The ISO 26262 series, Road vehicles — Functional safety, addresses safety-related electrical and electronic systems. Its parts cover activities such as hazard analysis, functional safety management, system development, hardware development and software development.

Safety of the intended functionality

ISO 21448:2022 addresses safety risks arising from insufficiencies in the specification or performance of intended functionality, including circumstances in which complex sensors and processing algorithms do not provide adequate situational awareness.

This subject is distinct from failures addressed by the ISO 26262 series. ISO 21448 provides guidance for identifying and evaluating potentially hazardous behavior associated with intended functions.

Cybersecurity

Automated vehicles incorporate electronic systems, software and interfaces that require cybersecurity risk management.

ISO/SAE 21434:2021 establishes an engineering framework for managing cybersecurity risks associated with road-vehicle electrical and electronic systems throughout their lifecycle. Cybersecurity engineering is related to vehicle safety but addresses a different category of threats from ordinary functional failures.

Human factors and supervision

The allocation of driving responsibility between a human and an automated system presents human-factors questions, especially when a person must supervise automation or resume manual control.

A research program reported by the National Highway Traffic Safety Administration in 2015 examined interactions with Level 2 and Level 3 driving concepts. Its experiments investigated driver engagement, requests to take over and the influence of human–machine interfaces on transitions between automated and manual operation.

The research identified differences between interface strategies and found that nonvisual components could contribute to effective takeover alerts in the experimental conditions studied.

Human supervision requirements therefore need to be described in relation to the actual automation level and system design rather than inferred from the presence of automated steering or speed control.

Testing and validation

Automated driving systems may be evaluated through simulation, controlled-track experiments and testing on public roads. These approaches examine different operating conditions and system behaviors.

A major challenge is determining how much evidence is required to support claims of safety and reliability.

Nidhi Kalra and Susan M. Paddock examined the statistical difficulty of demonstrating automated-vehicle reliability through road testing. Their 2016 analysis showed that extremely large distances could be required to establish statistically meaningful differences in rare outcomes such as serious crashes and fatalities.

The study concluded that ordinary road testing alone is unlikely to provide an adequate means of demonstrating safety for all relevant outcomes, motivating complementary methods of evaluation.

Interpreting crash statistics

Crash reports and public testing summaries require careful interpretation.

The U.S. National Highway Traffic Safety Administration publishes incident information under its Standing General Order on Crash Reporting. The agency explains that reported incident counts are not necessarily normalized for factors such as the number of vehicles operated, distance traveled or operating conditions.

Consequently, simple comparisons of reported crash totals between different automated-driving operators cannot independently establish which system is safer.

Regulation and oversight

Legal requirements for the testing and operation of automated vehicles vary by jurisdiction. Technical standards, government approval processes, traffic laws and operational permissions serve different functions and should not be treated as interchangeable.

Within the United Nations vehicle-regulation framework, UN Regulation No. 157 establishes technical requirements for Automated Lane Keeping Systems (ALKS). The regulation entered into force in January 2021 and concerns a defined class of automated driving functionality rather than unrestricted vehicle autonomy.

In the United States, the National Highway Traffic Safety Administration oversees vehicle-safety matters and operates reporting arrangements for certain incidents involving automated driving systems and Level 2 driver-assistance systems.

A system's classification under SAE J3016 does not establish that it has obtained permission to operate on public roads in a particular jurisdiction. SAE J3016 supplies terminology and definitions rather than approval requirements or a vehicle-safety certification process.

Potential societal effects

Research on automated vehicles has examined possible effects on road safety, personal mobility, transportation costs, traffic congestion, energy consumption and land use.

A RAND Corporation research report published in 2016 identified potential benefits including greater mobility for people unable to drive and changes in the use of travel time. It also identified possible disadvantages, including increased traffic demand if automation reduces the perceived or actual cost of travelling by road.

These outcomes were considered potential consequences rather than established effects applicable to every automated-vehicle deployment.

The extent of such effects depends on vehicle capabilities, deployment conditions, transportation policies and how automated vehicles are used.

Limitations and research challenges

Research and safety assessments identify several continuing technical challenges.

Automated driving systems must interpret changing road environments, interact with other road users and respond to situations that may differ from those encountered during development.

Research literature examines limitations affecting sensing, scene interpretation, behavior prediction, motion planning, system reliability and the evaluation of learned driving functions.

Automated driving also depends on accurately interpreting traffic rules and road-control information, which may differ between jurisdictions or change because of temporary road conditions.

The presence of automated driving capability must therefore be distinguished from evidence that a particular system performs safely within its claimed operational domain. Neither a successful demonstration nor an automation-level designation independently establishes unrestricted driving capability or comparative safety superiority.

See also

• Autonomous Navigation
• Autonomous Robots
• Computer Vision
• AI Safety