Self-driving car

This article is about the road vehicle type. For the general concept, see Vehicular automation. For wider applications, see Unmanned ground vehicle.

A self-driving car, also known as a autonomous car (AC), driverless car, robotaxi, robotic car or robo-car, is a car that is capable of operating with reduced or no human input. Self-driving cars are responsible for all driving activities, such as perceiving the environment, monitoring important systems, and controlling the vehicle, which includes navigating from origin to destination.

As of late 2024, no system has achieved full autonomy (SAE Level 5). In December 2020, Waymo was the first to offer rides in self-driving taxis to the public in limited geographic areas (SAE Level 4), and as of April 2024 offers services in Arizona (Phoenix) and California (San Francisco and Los Angeles). In June 2024, after a Waymo self-driving taxi crashed into a utility pole in Phoenix, Arizona, all 672 of its Jaguar I-Pace were recalled after they were found to have susceptibility to crashing into pole like items and had their software updated. In July 2021, DeepRoute.ai started offering self-driving taxi rides in Shenzhen, China. Starting in February 2022, Cruise offered self-driving taxi service in San Francisco, but suspended service in 2023. In 2021, Honda was the first manufacturer to sell an SAE Level 3 car, followed by Mercedes-Benz in 2023.

A Waymo-operated Jaguar I-Pace in San Francisco, 2023. This model was recalled after a crash into a pole.

Roborace autonomous racing car on display at the 2017 New York City ePrix

Baidu self-driving car

History

Main article: History of self-driving cars

Experiments have been conducted on advanced driver assistance systems (ADAS) since at least the 1920s. The first ADAS system was cruise control, which was invented in 1948 by Ralph Teetor.

Trials began in the 1950s. The first semi-autonomous car was developed in 1977, by Japan's Tsukuba Mechanical Engineering Laboratory. It required specially marked streets that were interpreted by two cameras on the vehicle and an analog computer. The vehicle reached speeds of 30 km/h (19 mph) with the support of an elevated rail.

Carnegie Mellon University's Navlab and ALV semi-autonomous projects launched in the 1980s, funded by the United States' Defense Advanced Research Projects Agency (DARPA) starting in 1984 and Mercedes-Benz and Bundeswehr University Munich's EUREKA Prometheus Project in 1987. By 1985, ALV had reached 31 km/h (19 mph), on two-lane roads. Obstacle avoidance came in 1986, and day and night off-road driving by 1987. In 1995 Navlab 5 completed the first autonomous US coast-to-coast journey. Traveling from Pittsburgh, Pennsylvania and San Diego, California, 98.2% of the trip was autonomous. It completed the trip at an average speed of 63.8 mph (102.7 km/h). Until the second DARPA Grand Challenge in 2005, automated vehicle research in the United States was primarily funded by DARPA, the US Army, and the US Navy, yielding incremental advances in speeds, driving competence, controls, and sensor systems.

The US allocated US$650 million in 1991 for research on the National Automated Highway System, which demonstrated automated driving, combining highway-embedded automation with vehicle technology, and cooperative networking between the vehicles and highway infrastructure. The programme concluded with a successful demonstration in 1997. Partly funded by the National Automated Highway System and DARPA, Navlab drove 4,584 km (2,848 mi) across the US in 1995, 4,501 km (2,797 mi) or 98% autonomously. In 2015, Delphi piloted a Delphi technology-based Audi, over 5,472 km (3,400 mi) through 15 states, 99% autonomously. In 2015, Nevada, Florida, California, Virginia, Michigan, and Washington DC allowed autonomous car testing on public roads.

From 2016 to 2018, the European Commission funded development for connected and automated driving through Coordination Actions CARTRE and SCOUT programs. The Strategic Transport Research and Innovation Agenda (STRIA) Roadmap for Connected and Automated Transport was published in 2019.

In November 2017, Waymo announced testing of autonomous cars without a safety driver. However, an employee was in the car to handle emergencies.

In March 2018, Elaine Herzberg became the first reported pedestrian killed by a self-driving car, an Uber test vehicle with a human backup driver; prosecutors did not charge Uber, while the human driver was sentenced to probation.

In December 2018, Waymo was the first to commercialize a robotaxi service, in Phoenix, Arizona. In October 2020, Waymo launched a robotaxi service in a (geofenced) part of the area. The cars were monitored in real-time, and remote engineers intervened to handle exceptional conditions.

In March 2019, ahead of Roborace, Robocar set the Guinness World Record as the world's fastest autonomous car. Robocar reached 282.42 km/h (175.49 mph).

In March 2021, Honda began leasing in Japan a limited edition of 100 Legend Hybrid EX sedans equipped with Level 3 "Traffic Jam Pilot" driving technology, which legally allowed drivers to take their eyes off the road when the car was travelling under 30 kilometres per hour (19 mph).

In December 2020, Waymo became the first service provider to offer driverless taxi rides to the general public, in a part of Phoenix, Arizona. Nuro began autonomous commercial delivery operations in California in 2021. DeepRoute.ai launched robotaxi service in Shenzhen in July 2021. In December 2021, Mercedes-Benz received approval for a Level 3 car. In February 2022, Cruise became the second service provider to offer driverless taxi rides to the general public, in San Francisco. In December 2022, several manufacturers scaled back plans for self-driving technology, including Ford and Volkswagen. In 2023, Cruise suspended its robotaxi service. Nuro was approved for Level 4 in Palo Alto in August, 2023.

As of August 2023, vehicles operating at Level 3 and above were an insignificant market factor; as of early 2024, Honda leases a Level 3 car in Japan, and Mercedes sells two Level 3 cars in Germany, California and Nevada.

Definitions

Organizations such as SAE have proposed terminology standards. However, most terms have no standard definition and are employed variously by vendors and others. Proposals to adopt aviation automation terminology for cars have not prevailed.

Names such as AutonoDrive, PilotAssist, Full-Self Driving or DrivePilot are used even though the products offer an assortment of features that may not match the names. Despite offering a system it called Full Self-Driving, Tesla stated that its system did not autonomously handle all driving tasks. In the United Kingdom, a fully self-driving car is defined as a car so registered, rather than one that supports a specific feature set. The Association of British Insurers claimed that the usage of the word autonomous in marketing was dangerous because car ads make motorists think "autonomous" and "autopilot" imply that the driver can rely on the car to control itself, even though they do not.

Automated driving system

SAE identified 6 levels for driving automation from level 0 to level 5. An ADS is an SAE J3016 level 3 or higher system.

Advanced driver assistance system

Main article: Advanced driver-assistance system

An ADAS is a system that automates specific driving features, such as Forward Collision Warning (FCW), Automatic Emergency Braking (AEB), Lane Departure Warning (LDW), Lane Keeping Assistance (LKA) or Blind Spot Warning (BSW). An ADAS requires a human driver to handle tasks that the ADAS does not support.

Autonomy versus automation

Autonomy implies that an automation system is under the control of the vehicle rather than a driver. Automation is function-specific, handling issues such as speed control, but leaves broader decision-making to the driver.

Euro NCAP defined autonomous as "the system acts independently of the driver to avoid or mitigate the accident".

In Europe, the words automated and autonomous can be used together. For instance, Regulation (EU) 2019/2144 supplied:

• "automated vehicle" means a vehicle that can move without continuous driver supervision, but that driver intervention is still expected or required in the operational design domains (ODD);
• "fully automated vehicle" means a vehicle that can move entirely without driver supervision;

Cooperative system

A remote driver is a driver that operates a vehicle at a distance, using a video and data connection.

According to SAE J3016,

Some driving automation systems may indeed be autonomous if they perform all of their functions independently and self-sufficiently, but if they depend on communication and/or cooperation with outside entities, they should be considered cooperative rather than autonomous.

Operational design domain

This section is an excerpt from Operational design domain.[edit]

Operational design domain (ODD) is a term for a particular operating context for an automated system, often used in the field of autonomous vehicles. The context is defined by a set of conditions, including environmental, geographical, time of day, and other conditions. For vehicles, traffic and roadway characteristics are included. Manufacturers use ODD to indicate where/how their product operates safely. A given system may operate differently according to the immediate ODD.

The concept presumes that automated systems have limitations. Relating system function to the ODD it supports is important for developers and regulators to establish and communicate safe operating conditions. Systems should operate within those limitations. Some systems recognize the ODD and modify their behavior accordingly. For example, an autonomous car might recognize that traffic is heavy and disable its automated lane change feature.

Vendors have taken a variety of approaches to the self-driving problem. Tesla's approach is to allow their "full self-driving" (FSD) system to be used in all ODDs as a Level 2 (hands/on, eyes/on) ADAS. Waymo picked specific ODDs (city streets in Phoenix and San Francisco) for their Level 5 robotaxi service. Mercedes Benz offers Level 3 service in Las Vegas in highway traffic jams at speeds up to 40 miles per hour (64 km/h). Mobileye's SuperVision system offers hands-off/eyes-on driving on all road types at speeds up to 130 kilometres per hour (81 mph). GM's hands-free Super Cruise operates on specific roads in specific conditions, stopping or returning control to the driver when ODD changes. In 2024 the company announced plans to expand road coverage from 400,000 miles to 750,000 miles. Ford's BlueCruise hands-off system operates on 130,000 miles of US divided highways.

Self-driving

The Union of Concerned Scientists defined self-driving as "cars or trucks in which human drivers are never required to take control to safely operate the vehicle. Also known as autonomous or 'driverless' cars, they combine sensors and software to control, navigate, and drive the vehicle."

The British Automated and Electric Vehicles Act 2018 law defines a vehicle as "driving itself" if the vehicle is "not being controlled, and does not need to be monitored, by an individual".

Another British government definition stated, "Self-driving vehicles are vehicles that can safely and lawfully drive themselves".

British definitions

In British English, the word automated alone has several meanings, such as in the sentence: "Thatcham also found that the automated lane keeping systems could only meet two out of the twelve principles required to guarantee safety, going on to say they cannot, therefore, be classed as 'automated driving', preferring 'assisted driving'". The first occurrence of the "automated" word refers to an Unece automated system, while the second refers to the British legal definition of an automated vehicle. British law interprets the meaning of "automated vehicle" based on the interpretation section related to a vehicle "driving itself" and an insured vehicle.

In November 2023 the British Government introduced the Automated Vehicles Bill. It proposed definitions for related terms:

• Self-driving: "A vehicle “satisfies the self-driving test” if it is designed or adapted with the intention that a feature of the vehicle will allow it to travel autonomously, and it is capable of doing so, by means of that feature, safely and legally."
• Autonomy: A vehicle travels “autonomously” if it is controlled by the vehicle, and neither the vehicle nor its surroundings are monitored by a person who can intervene.
• Control: control of vehicle motion.
• Safe: a vehicle that conforms to an acceptably safe standard.
• Legal: a vehicle that offers an acceptably low risk of committing a traffic infraction.

SAE classification

Tesla Autopilot is classified as an SAE Level 2 system.

A six-level classification system – ranging from fully manual to fully automated – was published in 2014 by SAE International as J3016, Taxonomy and Definitions for Terms Related to On-Road Motor Vehicle Automated Driving Systems; the details are revised occasionally. This classification is based on the role of the driver, rather than the vehicle's capabilities, although these are related. After SAE updated its classification in 2016, (J3016_201609), the National Highway Traffic Safety Administration (NHTSA) adopted the SAE standard. The classification is a topic of debate, with various revisions proposed.

Classifications

A "driving mode", aka driving scenario, combines an ODD with matched driving requirements (e.g., expressway merging, traffic jam). Cars may switch levels in accord with the driving mode.

Above Level 1, level differences are related to how responsibility for safe movement is divided/shared between ADAS and driver rather than specific driving features.

J3016 Automation Levels

Level	Name	Narrative		Direction and speed control	Monitoring	Fallback responsibility	Mode coverage
0	No Automation	Full-time performance by the driver of all aspects of driving, even when "enhanced by warning or intervention systems"		Driver	Driver	Driver	n/a
1	Driver Assistance	Driving mode-specific control by an ADAS of either steering or speed	Uses information about the driving environment and with the expectation that the driver performs all other driving tasks.	Driver and system			Some
2	Partial Automation	Driving mode-specific execution by one or more driver assistance systems of both steering and speed		System
3	Conditional Automation	Driving mode-specific control by an ADAS of all aspects of driving	Driver must appropriately respond to a request to intervene.		System
4	High Automation		If a driver does not respond appropriately to a request to intervene, the car can stop safely.			System	Many
5	Full Automation		System controls the vehicle under all conditions.				All

SAE Automation Levels have been criticized for their technological focus. It has been argued that the structure of the levels suggests that automation increases linearly and that more automation is better, which may not be the case. SAE Levels also do not account for changes that may be required to infrastructure and road user behavior.

Mobileye System

Mobileye CEO Amnon Shashua and CTO Shai Shalev-Shwartz proposed an alternative taxonomy for autonomous driving systems, claiming that a more consumer-friendly approach was needed. Its categories reflect the amount of driver engagement that is required. Some vehicle makers have informally adopted some of the terminology involved, while not formally committing to it.

Eyes-on/hands-on

The first level, hands-on/eyes-on, implies that the driver is fully engaged in operating the vehicle, but is supervised by the system, which intervenes according to the features it supports (e.g., adaptive cruise control, automatic emergency braking). The driver is entirely responsible, with hands on the wheel, and eyes on the road.

Eyes-on/hands-off

Eyes-on/hands-off allows the driver to let go of the wheel. The system drives, the driver monitors and remains prepared to resume control as needed.

Eyes-off/hands-off

Eyes-off/hands-off means that the driver can stop monitoring the system, leaving the system in full control. Eyes-off requires that no errors be reproducible (not triggered by exotic transitory conditions) or frequent, that speeds are contextually appropriate (e.g., 80 mph on limited-access roads), and that the system handle typical maneuvers (e.g., getting cut off by another vehicle). The automation level could vary according to the road (e.g., eyes-off on freeways, eyes-on on side streets).

No driver

The highest level does not require a human driver in the car: monitoring is done either remotely (telepresence) or not at all.

Safety

A critical requirement for the higher two levels is that the vehicle be able to conduct a Minimum Risk Maneuver and stop safely out of traffic without driver intervention.

Technology

Main article: Vehicular automation

Architecture

The perception system processes visual and audio data from outside and inside the car to create a local model of the vehicle, the road, traffic, traffic controls and other observable objects, and their relative motion. The control system then takes actions to move the vehicle, considering the local model, road map, and driving regulations.

Several classifications have been proposed to describe ADAS technology. One proposal is to adopt these categories: navigation, path planning, perception, and car control.

Navigation

Main article: Hybrid navigation

Navigation involves the use of maps to define a path between origin and destination. Hybrid navigation is the use of multiple navigation systems. Some systems use basic maps, relying on perception to deal with anomalies. Such a map understands which roads lead to which others, whether a road is a freeway, a highway, are one-way, etc. Other systems require highly detailed maps, including lane maps, obstacles, traffic controls, etc.

Perception

ACs need to be able to perceive the world around them. Supporting technologies include combinations of cameras, LiDAR, radar, audio, and ultrasound, GPS, and inertial measurement. Deep neural networks are used to analyse inputs from these sensors to detect and identify objects and their trajectories. Some systems use Bayesian simultaneous localization and mapping (SLAM) algorithms. Another technique is detection and tracking of other moving objects (DATMO), used to handle potential obstacles. Other systems use roadside real-time locating system (RTLS) technologies to aid localization. Tesla's "vision only" system uses eight cameras, without LIDAR or radar, to create its bird's-eye view of the environment.

Path planning

Path planning finds a sequence of segments that a vehicle can use to move from origin to destination. Techniques used for path planning include graph-based search and variational-based optimization techniques. Graph-based techniques can make harder decisions such as how to pass another vehicle/obstacle. Variational-based optimization techniques require more stringent restrictions on the vehicle's path to prevent collisions. The large scale path of the vehicle can be determined by using a voronoi diagram, an occupancy grid mapping, or a driving corridor algorithm. The latter allows the vehicle to locate and drive within open space that is bounded by lanes or barriers.

Maps

Maps are necessary for navigation. Map sophistication varies from simple graphs that show which roads connect to each other, with details such as one-way vs two-way, to those that are highly detailed, with information about lanes, traffic controls, roadworks, and more. Researchers at the MITComputer Science and Artificial Intelligence Laboratory (CSAIL) developed a system called MapLite, which allows self-driving cars to drive with simple maps. The system combines the GPS position of the vehicle, a "sparse topological map" such as OpenStreetMap (which has only 2D road features), with sensors that observe road conditions. One issue with highly-detailed maps is updating them as the world changes. Vehicles that can operate with less-detailed maps do not require frequent updates or geo-fencing.

Sensors

Sensors are necessary for the vehicle to properly respond to the driving environment. Sensor types include cameras, LiDAR, ultrasound, and radar. Control systems typically combine data from multiple sensors. Multiple sensors can provide a more complete view of the surroundings and can be used to cross-check each other to correct errors. For example, radar can image a scene in, e.g., a nighttime snowstorm, that defeats cameras and LiDAR, albeit at reduced precision. After experimenting with radar and ultrasound, Tesla adopted a vision-only approach, asserting that humans drive using only vision, and that cars should be able to do the same, while citing the lower cost of cameras versus other sensor types. By contrast, Waymo makes use of the higher resolution of LiDAR sensors and cites the declining cost of that technology.

Drive by wire

Main article: Drive by wire

Drive by wire is the use of electrical or electro-mechanical systems for performing vehicle functions such as steering or speed control that are traditionally achieved by mechanical linkages.

Driver monitoring

Main article: Driver monitoring system

Driver monitoring is used to assess the driver's attention and alertness. Techniques in use include eye monitoring, and requiring the driver to maintain torque on the steering wheel. It attempts to understand driver status and identify dangerous driving behaviors.

Vehicle communication

Main article: Vehicular communication systems

Vehicles can potentially benefit from communicating with others to share information about traffic, road obstacles, to receive map and software updates, etc.

ISO/TC 22 specifies in-vehicle transport information and control systems, while ISO/TC 204 specifies information, communication and control systems in surface transport. International standards have been developed for ADAS functions, connectivity, human interaction, in-vehicle systems, management/engineering, dynamic map and positioning, privacy and security.

Rather than communicating among vehicles, they can communicate with road-based systems to receive similar information.

Software update

See also: Over-the-air programming

Software controls the vehicle, and can provide entertainment and other services. Over-the-air updates can deliver bug fixes and additional features over the internet. Software updates are one way to accomplish recalls that in the past required a visit to a service center. In March 2021, the UNECE regulation on software update and software update management systems was published.

Safety model

A safety model is software that attempts to formalize rules that ensure that ACs operate safely.

IEEE is attempting to forge a standard for safety models as "IEEE P2846: A Formal Model for Safety Considerations in Automated Vehicle Decision Making". In 2022, a research group at National Institute of Informatics (NII, Japan) enhanced Mobileye's Reliable Safety System as "Goal-Aware RSS" to enable RSS rules to deal with complex scenarios via program logic.

Notification

The US has standardized the use of turquoise lights to inform other drivers that a vehicle is driving autonomously. It will be used in the 2026 Mercedes-Benz EQS and S-Class sedans with Drive Pilot, an SAE Level 3 driving system.

As of 2023, the Turquoise light had not been standardized by the P.R.C or the UN-ECE.

Artificial Intelligence

Artificial intelligence (AI) plays a pivotal role in the development and operation of autonomous vehicles (AVs), enabling them to perceive their surroundings, make decisions, and navigate safely without human intervention. AI algorithms empower AVs to interpret sensory data from various onboard sensors, such as cameras, LiDAR, radar, and GPS, to understand their environment and improve its technological ability and overall safety over time.

Challenges

Autonomous delivery vehicles stuck in one place by attempting to avoid one another

Obstacles

The primary obstacle to ACs is the advanced software and mapping required to make them work safely across the wide variety of conditions that drivers experience. In addition to handling day/night driving in good and bad weather on roads of arbitrary quality, ACs must cope with other vehicles, road obstacles, poor/missing traffic controls, flawed maps, and handle endless edge cases, such as following the instructions of a police officer managing traffic at a crash site.

Other obstacles include cost, liability, consumer reluctance, ethical dilemmas, security, privacy, and legal/regulatory framework. Further, AVs could automate the work of professional drivers, eliminating many jobs, which could slow acceptance.

Concerns

Deceptive marketing

Tesla calls its Level 2 ADAS "Full Self-Driving (FSD) Beta". US Senators Richard Blumenthal and Edward Markey called on the Federal Trade Commission (FTC) to investigate this marketing in 2021. In December 2021 in Japan, Mercedes-Benz was punished by the Consumer Affairs Agency for misleading product descriptions.

Mercedes-Benz was criticized for a misleading US commercial advertising E-Class models. At that time, Mercedes-Benz rejected the claims and stopped its "self-driving car" ad campaign that had been running. In August 2022, the California Department of Motor Vehicles (DMV) accused Tesla of deceptive marketing practices.

With the Automated Vehicles Bill (AVB) self-driving car-makers could face prison for misleading adverts in the United-Kingdom.

Security

In the 2020s, concerns over ACs' vulnerability to cyberattacks and data theft emerged.

Espionage

In 2018 and 2019 former Apple engineers were charged with stealing information related to Apple's self-driving car project. In 2021 the United States Department of Justice (DOJ) accused Chinese security officials of coordinating a hacking campaign to steal information from government entities, including research related to autonomous vehicles. China has prepared "the Provisions on Management of Automotive Data Security (Trial) to protect its own data".

Cellular Vehicle-to-Everything technologies are based on 5G wireless networks. As of November 2022, the US Congress was considering the possibility that imported Chinese AC technology could facilitate espionage.

Testing of Chinese automated cars in the US has raised concern over which US data are collected by Chinese vehicles to be stored in Chinese country and concern with any link with the Chinese communist party.

Driver communications

ACs complicate the need for drivers to communicate with each other, e.g., to decide which car enters an intersection first. In an AC without a driver, traditional means such as hand signals do not work (no driver, no hands).

Behavior prediction

ACs must be able to predict the behavior of possibly moving vehicles, pedestrians, etc in real time in order to proceed safely. The task becomes more challenging the further into the future the prediction extends, requiring rapid revisions to the estimate to cope with unpredicted behavior. One approach is to wholly recompute the position and trajectory of each object many times per second. Another is to cache the results of an earlier prediction for use in the next one to reduce computational complexity.

Handover

The ADAS has to be able to safely accept control from and return control to the driver.

Trust

Consumers will avoid ACs unless they trust them as safe. Robotaxis operating in San Francisco received pushback over perceived safety risks. Automatic elevators were invented in 1900, but did not become common until operator strikes and trust was built with advertising and features such as an emergency stop button. However, with repeated use of autonomous driving functions, drivers' behavior and trust in autonomous vehicles gradually improved and both entered a more stable state. At the same time this also improved the performance and reliability of the vehicle in complex conditions, thereby increasing public trust.

Economics

Autonomous also present various political and economic implications. The transportation sector holds significant sway in many the political and economic landscapes. For instance, many US states generates much annual revenue from transportation fees and taxes. The advent of self-driving cars could profoundly affect the economy by potentially altering state tax revenue streams. Furthermore, the transition to autonomous vehicles might disrupt employment patterns and labor markets, particularly in industries heavily reliant on driving professions. Data from the U.S. Bureau of Labor Statistics indicates that in 2019, the sector employed over two million individuals as tractor-trailer truck drivers. Additionally, taxi and delivery drivers represented approximately 370,400 positions, and bus drivers constituted a workforce of over 680,000. Collectively, this amounts to a conceivable displacement of nearly 2.9 million jobs, surpassing the job losses experienced in the 2008 Great Recession.

Equity and Inclusion

The prominence of certain demographic groups within the tech industry inevitably shapes the trajectory of autonomous vehicle (AV) development, potentially perpetuating existing inequalities. There are others in society without a political agenda who believe that the advancement of technology has nothing to do with promoting inequalities in certain groups and see this as a ridiculous presumption.

Ethical issues

See also: Machine ethics

Pedestrian Detection

Research from Georgia Tech revealed that autonomous vehicle detection systems were generally five percent less effective at recognizing darker-skinned individuals. This accuracy gap persisted despite adjustments for environmental variables like lighting and visual obstructions.

Rationale for liability

Standards for liability have yet to be adopted to address crashes and other incidents. Liability could rest with the vehicle occupant, its owner, the vehicle manufacturer, or even the ADAS technology supplier, possibly depending on the circumstances of the crash. Additionally, the infusion of ArtificiaI Intelligence technology in autonomous vehicles adds layers of complexity to ownership and ethical dynamics. Given that AI systems are inherently self-learning, a question arises of whether accountability should rest with the vehicle owner, the manufacturer, or the AI developer?

Trolley problem

The trolley problem is a thought experiment in ethics. Adapted for ACs, it considers an AC carrying one passenger confronts a pedestrian who steps in its way. The ADAS notionally has to choose between killing the pedestrian or swerving into a wall, killing the passenger. Possible frameworks include deontology (formal rules) and utilitarianism (harm reduction).

One public opinion survey reported that harm reduction was preferred, except that passengers wanted the vehicle to prefer them, while pedestrians took the opposite view. Utilitarian regulations were unpopular. Additionally, cultural viewpoints exert substantial influence on shaping responses to these ethical quandaries. Another study found that cultural biases impact preferences in prioritizing the rescue of certain individuals over others in car accident scenarios.

Privacy

Some ACs require an internet connection to function, opening the possibility that a hacker might gain access to private information such as destinations, routes, camera recordings, media preferences, and/or behavioral patterns, although this is true of an internet-connected device.

Road infrastructure

ACs make use of road infrastructure (e.g., traffic signs, turn lanes) and may require modifications to that infrastructure to fully achieve their safety and other goals. In March 2023, the Japanese government unveiled a plan to set up a dedicated highway lane for ACs. In April 2023, JR East announced their challenge to raise their self-driving level of Kesennuma Line bus rapid transit (BRT) in rural area from the current Level 2 to Level 4 at 60 km/h.

Testing

Approaches

ACs can be tested via digital simulations, in a controlled test environment, and/or on public roads. Road testing typically requires some form of permit or a commitment to adhere to acceptable operating principles. For example, New York requires a test driver to be in the vehicle, prepared to override the ADAS as necessary.

2010s and disengagements

A prototype of Waymo's self-driving car, navigating public streets in Mountain View, California in 2017

In California, self-driving car manufacturers are required to submit annual reports describing how often their vehicles autonomously disengaged from autonomous mode. This is one measure of system robustness (ideally, the system should never disengage).

In 2017, Waymo reported 63 disengagements over 352,545 mi (567,366 km) of testing, an average distance of 5,596 mi (9,006 km) between disengagements, the highest (best) among companies reporting such figures. Waymo also logged more autonomous miles than other companies. Their 2017 rate of 0.18 disengagements per 1,000 mi (1,600 km) was an improvement over the 0.2 disengagements per 1,000 mi (1,600 km) in 2016, and 0.8 in 2015. In March 2017, Uber reported an average of 0.67 mi (1.08 km) per disengagement. In the final three months of 2017, Cruise (owned by GM) averaged 5,224 mi (8,407 km) per disengagement over 62,689 mi (100,888 km).

Disengagement data

Car maker	California, 2016		California, 2018		California, 2019
	Distance between disengagements	Total distance traveled	Distance between disengagements	Total distance traveled	Distance between disengagements	Total distance traveled
Waymo	5,128 mi (8,253 km)	635,868 mi (1,023,330 km)	11,154 mi (17,951 km)	1,271,587 mi (2,046,421 km)	11,017 mi (17,730 km)	1,450,000 mi (2,330,000 km)
BMW	638 mi (1,027 km)	638 mi (1,027 km)
Nissan	263 mi (423 km)	6,056 mi (9,746 km)	210 mi (340 km)	5,473 mi (8,808 km)
Ford	197 mi (317 km)	590 mi (950 km)
General Motors	55 mi (89 km)	8,156 mi (13,126 km)	5,205 mi (8,377 km)	447,621 mi (720,376 km)	12,221 mi (19,668 km)	831,040 mi (1,337,430 km)
Aptiv	15 mi (24 km)	2,658 mi (4,278 km)
Tesla	3 mi (4.8 km)	550 mi (890 km)
Mercedes-Benz	2 mi (3.2 km)	673 mi (1,083 km)	1.5 mi (2.4 km)	1,749 mi (2,815 km)
Bosch	7 mi (11 km)	983 mi (1,582 km)
Zoox			1,923 mi (3,095 km)	30,764 mi (49,510 km)	1,595 mi (2,567 km)	67,015 mi (107,850 km)
Nuro			1,028 mi (1,654 km)	24,680 mi (39,720 km)	2,022 mi (3,254 km)	68,762 mi (110,662 km)
Pony.ai			1,022 mi (1,645 km)	16,356 mi (26,322 km)	6,476 mi (10,422 km)	174,845 mi (281,386 km)
Baidu (Apolong)			206 mi (332 km)	18,093 mi (29,118 km)	18,050 mi (29,050 km)	108,300 mi (174,300 km)
Aurora			100 mi (160 km)	32,858 mi (52,880 km)	280 mi (450 km)	39,729 mi (63,938 km)
Apple			1.1 mi (1.8 km)	79,745 mi (128,337 km)	118 mi (190 km)	7,544 mi (12,141 km)
Uber			0.4 mi (0.64 km)	26,899 mi (43,290 km)		0 mi (0 km)

2020s

Disengagement definitions

Reporting companies use varying definitions of what qualifies as a disengagement, and such definitions can change over time. Executives of self-driving car companies have criticized disengagements as a deceptive metric, because it does not consider varying road conditions.

Standards

In April 2021, WP.29 GRVA proposed a "Test Method for Automated Driving (NATM)".

In October 2021, Europe's pilot test, L3Pilot, demonstrated ADAS for cars in Hamburg, Germany, in conjunction with ITS World Congress 2021. SAE Level 3 and 4 functions were tested on ordinary roads.

In November 2022, an International Standard ISO 34502 on "Scenario based safety evaluation framework" was published.

Collision avoidance

In April 2022, collision avoidance testing was demonstrated by Nissan. Waymo published a document about collision avoidance testing in December 2022.

Simulation and validation

In September 2022, Biprogy released Driving Intelligence Validation Platform (DIVP) as part of Japanese national project "SIP-adus", which is interoperable with Open Simulation Interface (OSI) of ASAM.

Toyota

In November 2022, Toyota demonstrated one of its GR Yaris test cars, which had been trained using professional rally drivers. Toyota used its collaboration with Microsoft in FIA World Rally Championship since the 2017 season.

Pedestrian reactions

In 2023 David R. Large, senior research fellow with the Human Factors Research Group at the University of Nottingham, disguised himself as a car seat in a study to test people's reactions to driverless cars. He said, "We wanted to explore how pedestrians would interact with a driverless car and developed this unique methodology to explore their reactions." The study found that, in the absence of someone in the driving seat, pedestrians trust certain visual prompts more than others when deciding whether to cross the road.

Incidents

Tesla

See also: List of Tesla Autopilot crashes

As of 2023, Tesla's ADAS Autopilot/Full Self Driving (beta) was classified as Level 2 ADAS.

On 20 January 2016, the first of five known fatal crashes of a Tesla with Autopilot occurred, in China's Hubei province. Initially, Tesla stated that the vehicle was so badly damaged from the impact that their recorder was not able to determine whether the car had been on Autopilot at the time. However, the car failed to take evasive action.

Another fatal Autopilot crash occurred in May in Florida in a Tesla Model S that crashed into a tractor-trailer. In a civil suit between the father of the driver killed and Tesla, Tesla documented that the car had been on Autopilot. According to Tesla, "neither Autopilot nor the driver noticed the white side of the tractor-trailer against a brightly lit sky, so the brake was not applied." Tesla claimed that this was Tesla's first known Autopilot death in over 130 million miles (210 million kilometers) with Autopilot engaged. Tesla claimed that on average one fatality occurs every 94 million miles (151 million kilometers) across all vehicle types in the US. However, this number also includes motorcycle/pedestrian fatalities. The ultimate National Transportation Safety Board (NTSB) report concluded Tesla was not at fault; the investigation revealed that for Tesla cars, the crash rate dropped by 40 percent after Autopilot was installed.

Google Waymo

See also: Waymo § Crashes

Google's in-house automated car

In June 2015, Google confirmed that 12 vehicles had suffered collisions as of that date. Eight involved rear-end collisions at a stop sign or traffic light, in two of which the vehicle was side-swiped by another driver, one in which another driver rolled a stop sign, and one where a driver was controlling the car manually. In July 2015, three employees suffered minor injuries when their vehicle was rear-ended by a car whose driver failed to brake. This was the first collision that resulted in injuries.

According to Google Waymo's accident reports as of early 2016, their test cars had been involved in 14 collisions, of which other drivers were at fault 13 times, although in 2016 the car's software caused a crash. On 14 February 2016 a Google vehicle attempted to avoid sandbags blocking its path. During the maneuver it struck a bus. Google stated, "In this case, we clearly bear some responsibility, because if our car hadn't moved, there wouldn't have been a collision." Google characterized the crash as a misunderstanding and a learning experience. No injuries were reported.

Uber's Advanced Technologies Group (ATG)

In March 2018, Elaine Herzberg died after she was hit by an AC tested by Uber's Advanced Technologies Group (ATG) in Arizona. A safety driver was in the car. Herzberg was crossing the road about 400 feet from an intersection. Some experts said a human driver could have avoided the crash. Arizona governor Doug Ducey suspended the company's ability to test its ACs citing an "unquestionable failure" of Uber to protect public safety. Uber also stopped testing in California until receiving a new permit in 2020.

NTSB's final report determined that the immediate cause of the accident was that safety driver Rafaela Vasquez failed to monitor the road, because she was distracted by her phone, but that Uber's "inadequate safety culture" contributed. The report noted that the victim had "a very high level" of methamphetamine in her body. The board called on federal regulators to carry out a review before allowing automated test vehicles to operate on public roads.

In September 2020, Vasquez pled guilty to endangerment and was sentenced to three years' probation.

NIO Navigate on Pilot

On 12 August 2021, a 31-year-old Chinese man was killed after his NIO ES8 collided with a construction vehicle. NIO's self-driving feature was in beta and could not deal with static obstacles. The vehicle's manual clearly stated that the driver must take over near construction sites. Lawyers of the deceased's family questioned NIO's private access to the vehicle, which they argued did not guarantee the integrity of the data.

Pony.ai

In November 2021, the California Department of Motor Vehicles (DMV) notified Pony.ai that it was suspending its testing permit following a reported collision in Fremont on 28 October. In May 2022, DMV revoked Pony.ai's permit for failing to monitor the driving records of its safety drivers.

Cruise

See also: Cruise (autonomous vehicle) § Incidents

In April 2022, Cruise's testing vehicle was reported to have blocked a fire engine on emergency call, and sparked questions about its ability to handle unexpected circumstances.

Ford

In February 2024, a driver using the Ford BlueCruise hands-free driving feature struck and killed the driver of a stationary car with no lights on in the middle lane of a freeway in Texas.

In March 2024, a drunk driver who was speeding, holding her cell phone, and using BlueCruise on a Pennsylvania freeway struck and killed two people who had been driving two cars. The first car had become disabled and was on the left shoulder with part of the car in the left driving lane. The second driver had parked his car behind the first car presumably to help the first driver.

The NTSB is investigating both incidents.

Total incidents

The NHTSA began mandating incident reports from autonomous vehicle companies in June 2021. Some reports cite incidents from as early as August 2019, with current data available through June 17, 2024.

There have been a total of 3,979 autonomous vehicle incidents (both ADS and ADAS) reported during this timeframe. 2,146 of those incidents (53.9%) involved Tesla vehicles.

Public opinion surveys

2010s

In a 2011 online survey of 2,006 US and UK consumers, 49% said they would be comfortable using a "driverless car".

A 2012 survey of 17,400 vehicle owners found 37% who initially said they would be interested in purchasing a "fully autonomous car". However, that figure dropped to 20% if told the technology would cost US$3,000 more.

In a 2012 survey of about 1,000 German drivers, 22% had a positive attitude, 10% were undecided, 44% were skeptical and 24% were hostile.

A 2013 survey of 1,500 consumers across 10 countries found 57% "stated they would be likely to ride in a car controlled entirely by technology that does not require a human driver", with Brazil, India and China the most willing to trust automated technology.

In a 2014 US telephone survey, over three-quarters of licensed drivers said they would consider buying a self-driving car, rising to 86% if car insurance were cheaper. 31.7% said they would not continue to drive once an automated car was available.

In 2015, a survey of 5,000 people from 109 countries reported that average respondents found manual driving the most enjoyable. 22% did not want to pay more money for autonomy. Respondents were found to be most concerned about hacking/misuse, and were also concerned about legal issues and safety. Finally, respondents from more developed countries were less comfortable with their vehicle sharing data. The survey reported consumer interest in purchasing an AC, stating that 37% of surveyed current owners were either "definitely" or "probably" interested.

In 2016, a survey of 1,603 people in Germany that controlled for age, gender, and education reported that men felt less anxiety and more enthusiasm, whereas women showed the opposite. The difference was pronounced between young men and women and decreased with age.

In a 2016 US survey of 1,584 people, "66 percent of respondents said they think autonomous cars are probably smarter than the average human driver". People were worried about safety and hacking risk. Nevertheless, only 13% of the interviewees saw no advantages in this new kind of cars.

In a 2017 survey of 4,135 US adults found that many Americans anticipated significant impacts from various automation technologies including the widespread adoption of automated vehicles.

In 2019, results from two opinion surveys of 54 and 187 US adults respectively were published. The questionnaire was termed the autonomous vehicle acceptance model (AVAM), including additional description to help respondents better understand the implications of various automation levels. Users were less accepting of high autonomy levels and displayed significantly lower intention to use autonomous vehicles. Additionally, partial autonomy (regardless of level) was perceived as requiring uniformly higher driver engagement (usage of hands, feet and eyes) than full autonomy.

In the 2020s

In 2022, a survey reported that only a quarter (27%) of the world's population would feel safe in self-driving cars.

In 2024, a study by Saravanos et al. at New York University reported that 87% of their respondents (from a sample of 358) believed that conditionally automated cars (at Level 3) would be easy to use.

Opinion surveys may have little salience given that few respondents had any personal experience with ACs.

Regulation

Main article: Regulation of self-driving cars

See also: Regulation of algorithms

The regulation of autonomous cars concerns liability, approvals, and international conventions.

In the 2010s, researchers openly worried that delayed regulations could delay deployment. In 2020, UNECE WP.29 GRVA was issued to address regulation of Level 3 automated driving.

Commercialization

See also: History of self-driving cars § The 2020s

Vehicles operating below Level 5 still offer many advantages.

As of 2023 most commercially available ADAS vehicles are SAE Level 2. A couple of companies reached higher levels, but only in restricted (geofenced) locations.

Level 2 – Partial Automation

See also: Lane centering § Sample of level 2 automated cars, and List of self-driving system suppliers § Date of first public road driverless operation

SAE Level 2 features are available as part of the ADAS systems in many vehicles. In the US, 50% of new cars provide driver assistance for both steering and speed.

Ford started offering BlueCruise service on certain vehicles in 2022; the system is named ActiveGlide in Lincoln vehicles. The system provided features such as lane centering, street sign recognition, and hands-free highway driving on more than 130,000 miles of divided highways. The 2022 1.2 version added features including hands-free lane changing, in-lane repositioning, and predictive speed assist. In April 2023 BlueCruise was approved in the UK for use on certain motorways, starting with 2023 models of Ford's electric Mustang Mach-E SUV.

Tesla's Autopilot and its Full Self-Driving (FSD) ADAS suites are available on all Tesla cars since 2016. FSD offers highway and street driving (without geofencing), navigation/turn management, steering, and dynamic cruise control, collision avoidance, lane-keeping/switching, emergency braking, obstacle avoidance, but still requires the driver to remain ready to control the vehicle at any moment. Its driver management system combines eye tracking with monitoring pressure on the steering wheel to ensure that drives are both eyes on and hands on.

Tesla's FSD rewrite V12 (released in March 2024) uses a single deep learning transformer model for all aspects of perception, monitoring, and control. It relies on its eight cameras for its vision-only perception system, without use of LiDAR, radar, or ultrasound. As of April 2024, FSD has been deployed on two million Tesla cars. As of January 2024, Tesla has not initiated requests for Level 3 status for its systems and has not disclosed its reason for not doing so.

Development

General Motors is developing the "Ultra Cruise" ADAS system, that will be a dramatic improvement over their current "Super Cruise" system. Ultra Cruise will cover "95 percent" of driving scenarios on 2 million miles of roads in the US, according to the company. The system hardware in and around the car includes multiple cameras, short- and long-range radar, and a LiDAR sensor, and will be powered by the Qualcomm Snapdragon Ride Platform. The luxury Cadillac Celestiq electric vehicle will be one of the first vehicles to feature Ultra Cruise.

Europe is developing a new "Driver Control Assistance Systems" (DCAS) level 2 regulation to no longer limit the use of lane changing systems to roads with 2 lanes and a physical separation from traffic in the opposite direction.

Level 3 – Conditional Automation

As of April 2024, two car manufacturers have sold or leased Level 3 cars: Honda in Japan, and Mercedes in Germany, Nevada and California.

Mercedes Drive Pilot has been available on the EQS and S-class sedan in Germany since 2022, and in California and Nevada since 2023. A subscription costs between €5,000 and €7,000 for three years in Germany and $2,500 for one year in the United States. Drive Pilot can only be used when the vehicle is traveling under 40 miles per hour (64 km/h), there is a vehicle in front, readable line markings, during the day, clear weather, and on freeways mapped by Mercedes down to the centimeter (100,000 miles in California). As of April 2024, one Mercedes vehicle with this capability has been sold in California.

Development

Honda continued to enhance its Level 3 technology. As of 2023, 80 vehicles with Level 3 support had been sold.

Mercedes-Benz received authorization in early 2023 to pilot its Level 3 software in Las Vegas. California also authorized Drive Pilot in 2023.

BMW commercialized its AC in 2021. In 2023 BMW stated that its Level-3 technology was nearing release. It would be the second manufacturer to deliver Level-3 technology, but the only one with a Level 3 technology which works in the dark.

In 2023, in China, IM Motors, Mercedes, and BMW obtained authorization to test vehicles with Level 3 systems on motorways.

In September 2021, Stellantis presented its findings from its Level 3 pilot testing on Italian highways. Stellantis's Highway Chauffeur claimed Level 3 capabilities, as tested on the Maserati Ghibli and Fiat 500X prototypes.

Polestar, a Volvo Cars' brand, announced in January 2022 its plan to offer Level 3 autonomous driving system in the Polestar 3 SUV, a Volvo XC90 successor, with technologies from Luminar Technologies, Nvidia, and Zenseact.

In January 2022, Bosch and the Volkswagen Group subsidiary CARIAD released a collaboration for autonomous driving up to Level 3. This joint development _targets Level 4 capabilities.

Hyundai Motor Company is enhancing cybersecurity of connected cars to offer a Level 3 self-driving Genesis G90. Kia and Hyundai Korean car makers delayed their Level 3 plans, and will not deliver Level 3 vehicles in 2023.

Level 4 – High Automation

Waymo offers robotaxi services in parts of Arizona (Phoenix) and California (San Francisco and Los Angeles), as fully autonomous vehicles without safety drivers.

In April 2023 in Japan, a Level 4 protocol became part of the amended Road Traffic Act. ZEN drive Pilot Level 4 made by AIST operates there.

Development

In July 2020, Toyota started public demonstration rides on Lexus LS (fifth generation) based TRI-P4 with Level 4 capability. In August 2021, Toyota operated a potentially Level 4 service using e-Palette around the Tokyo 2020 Olympic Village.

In September 2020, Mercedes-Benz introduced world's first commercial Level 4 Automated Valet Parking (AVP) system named Intelligent Park Pilot for its new S-Class. In November 2022, Germany’s Federal Motor Transport Authority (KBA) approved the system for use at Stuttgart Airport.

In September 2021, Cruise, General Motors, and Honda started a joint testing programme, using Cruise AV. In 2023, the Origin was put on indefinite hold following Cruise's loss of its operating permit.

In January 2023, Holon announced an autonomous shuttle during the 2023 Consumer Electronics Show (CES). The company claimed the vehicle is the world's first Level 4 shuttle built to automotive standard.

See also

• Autopilot
• Driving

Further reading

Media related to Autonomous automobiles at Wikimedia Commons

• O'Toole, Randal (18 January 2010). Gridlock: Why We're Stuck in Traffic and What To Do About It. Cato Institute. ISBN 978-1-935308-24-9.
• Macdonald, Iain David Graham (2011). A Simulated Autonomous Car (PDF) (thesis). The University of Edinburgh. Retrieved 17 April 2013.
• Knight, Will (22 October 2013). "The Future of Self-driving Cars". MIT Technology Review. Retrieved 22 July 2016.
• Taiebat, Morteza; Brown, Austin; Safford, Hannah; Qu, Shen; Xu, Ming (2019). "A Review on Energy, Environmental, and Sustainability Implications of Connected and Automated Vehicles". Environmental Science & Technology. 52 (20): 11449–11465. arXiv:1901.10581. Bibcode:2019arXiv190110581T. doi:10.1021/acs.est.8b00127. PMID 30192527. S2CID 52174043.
• Glancy, Dorothy (2016). A Look at the Legal Environment for Driverless Vehicles (PDF) (Report). National Cooperative Highway Research Program Legal Research Digest. Vol. 69. Washington, DC: Transportation Research Board. ISBN 978-0-309-37501-6. Retrieved 22 July 2016.
• Newbold, Richard (17 June 2015). "The driving forces behind what would be the next revolution in the haulage sector". The Loadstar. Retrieved 22 July 2016.
• Bergen, Mark (27 October 2015). "Meet the Companies Building Self-Driving Cars for Google and Tesla (And Maybe Apple)". re/code.
• John A. Volpe National Transportation Systems Center (March 2016). "Review of Federal Motor Vehicle Safety Standards (FMVSS) for Automated Vehicles: Identifying potential barriers and challenges for the certification of automated vehicles using existing FMVSS" (PDF). National Transportation Library. US Department of Transportation. Archived from the original (PDF) on 16 June 2017. Retrieved 6 April 2016.
• Slone, Sean (August 2016). "State Laws on Autonomous Vehicles" (PDF). Capitol Research – Transportation Policy. Council of State Governments. Archived from the original (PDF) on 28 February 2021. Retrieved 28 September 2016.
• Henn, Steve (31 July 2015). "Remembering When Driverless Elevators Drew Skepticism".
• Anderson, James M.; et al. (2016). "Autonomous Vehicle Technology: A Guide for Policymakers" (PDF). RAND Corporation.
• Meyer, Gereon; Beiker, Sven, eds. (2014). "Road Vehicle Automation". Lecture Notes in Mobility. doi:10.1007/978-3-319-05990-7. ISBN 978-3-319-05989-1. ISSN 2196-5544.
  • Meyer, Gereon; Beiker, Sven, eds. (2015). "Road Vehicle Automation 2". Lecture Notes in Mobility. doi:10.1007/978-3-319-19078-5. ISBN 978-3-319-19077-8. ISSN 2196-5544.
  • Meyer, Gereon; Beiker, Sven, eds. (2016). "Road Vehicle Automation 3". Lecture Notes in Mobility. doi:10.1007/978-3-319-40503-2. ISBN 978-3-319-40502-5. ISSN 2196-5544.
  • Meyer, Gereon; Beiker, Sven, eds. (2018). "Road Vehicle Automation 4". Lecture Notes in Mobility. doi:10.1007/978-3-319-60934-8. ISBN 978-3-319-60933-1. ISSN 2196-5544.
  • Meyer, Gereon; Beiker, Sven, eds. (2019). "Road Vehicle Automation 5". Lecture Notes in Mobility. doi:10.1007/978-3-319-94896-6. ISBN 978-3-319-94895-9. ISSN 2196-5544. S2CID 168659939.
  • Meyer, Gereon; Beiker, Sven, eds. (2019). "Road Vehicle Automation 6". Lecture Notes in Mobility. doi:10.1007/978-3-030-22933-7. ISBN 978-3-030-22932-0. ISSN 2196-5544.

These books are based on presentations and discussions at the Automated Vehicles Symposium organized annually by TRB and AUVSI.

• Kemp, Roger (2018). "Autonomous vehicles – who will be liable for accidents?". [15 Digital Evidence and Electronic Signature Law Review (2018) 33 – 47].