A self-driving car, also called an autonomous vehicle or driverless car, is a vehicle that uses sensors, software and artificial intelligence to perceive its surroundings and drive with little or no human input. The car identifies lanes, vehicles, pedestrians and signs, predicts how others will move, plans a safe path and controls the steering, brakes and throttle. Autonomous driving is one of the most ambitious applications of AI robotics, and it draws on computer vision, sensor fusion, motion planning and machine learning.
Levels of automation
SAE International’s J3016 standard describes six levels of driving automation, which are widely used in regulation and industry [1].
- Level 0. No automation. The driver does everything, though warning systems may exist.
- Level 1. Driver assistance. The system helps with either steering or speed, as in adaptive cruise control.
- Level 2. Partial automation. The system controls both steering and speed, but the driver must supervise constantly and be ready to take over.
- Level 3. Conditional automation. The system drives in certain conditions and the human must take over when asked.
- Level 4. High automation. The system drives itself within a defined operating domain, such as a mapped city area, and does not need a human fallback inside it.
- Level 5. Full automation. The system drives anywhere in all conditions that a human could.
Most cars sold to consumers offer Level 1 or 2 driver assistance. Driverless ride-hailing services operate at Level 4 in limited areas. Level 5 has not been achieved. The terms “self-driving” and “autopilot” in marketing can mislead drivers into thinking that a Level 2 system is more capable than it is, which has been a safety concern.
History
Experiments with automated driving began in the 1920s and 1930s with radio-controlled cars, and in the 1980s, Ernst Dickmanns’ team in Germany and Carnegie Mellon University’s Navlab projects built vehicles that drove on roads using cameras. The DARPA Grand Challenge competitions, in 2004, 2005 and the 2007 Urban Challenge, spurred rapid progress, with Stanford’s Stanley completing the 2005 desert course [2]. Google began a self-driving project in 2009 that later became Waymo, and traditional carmakers and start-ups followed. In 2020, Waymo started offering fully driverless rides to the public in the Phoenix area [3]. Since then, robotaxi services have expanded to more cities, and driver assistance has spread across car brands, while trucking, delivery and shuttle applications continue to develop.
How self-driving cars work
### Sensors
A typical vehicle combines several sensor types, because each has strengths and weaknesses.
- Cameras capture color and texture and read signs and lights, but they struggle in darkness and glare.
- Radar measures distance and speed and works in rain and fog, but has low detail.
- Lidar emits laser pulses to create a precise 3D point cloud, but costs more and can be affected by weather.
- Ultrasonic sensors detect nearby objects at low speed.
- GPS and inertial units provide position and motion, often with high-definition maps.
The data are merged through sensor fusion. Some companies pursue camera-centered designs, while others combine many sensor types, and the choice is debated.
### Perception
Neural networks detect and classify vehicles, pedestrians, cyclists, lanes and signs; see object detection, image segmentation and object tracking. The vehicle builds a representation of the scene, including drivable space; see scene understanding and robotic perception.
### Localization and mapping
The car estimates its precise position in the lane by matching sensor data to maps, using methods related to simultaneous localization and mapping.
### Prediction
The system forecasts what other road users will do, such as whether a pedestrian will step out, using learned models of behavior.
### Planning and control
Behavior planning chooses maneuvers such as lane changes and yielding, while trajectory planning generates a smooth, safe path; see motion planning. Controllers then command steering, throttle and braking.
### End-to-end learning
Some approaches train a single large neural network to map sensor input directly to driving actions, with imitation of human drivers and reinforcement learning in simulation. These can handle subtle behavior but are harder to interpret and verify, so many systems combine learned components with rule-based safety layers.
Safety
Safety is the central issue. Advocates argue that automated systems could reduce crashes, since human error contributes to most road accidents, and that machines do not get tired or distracted. Critics point out that systems can fail in unusual situations, and that fatal crashes involving automated vehicles have occurred, including incidents with partial automation where drivers over-relied on the system. Evidence on safety compared with human drivers is still developing, and fair comparison requires matching roads, conditions and miles driven.
Developers use extensive simulation, closed-course testing and supervised road testing, collect data on interventions and incidents, and design redundancy in sensors, computing and braking. Regulators require reporting of crashes in some places. Key technical questions include handling rare “edge cases,” bad weather and unpredictable human behavior.
Regulation, ethics and society
Rules differ across countries and states, covering testing permits, driver monitoring, liability and data reporting. Legal questions include who is responsible when an automated system crashes, the driver, the manufacturer or the operator. Ethical discussions include how vehicles should prioritize safety in unavoidable harm situations, though engineers stress that this rarely resembles philosophical thought experiments and that avoiding such situations is the main goal. Other issues are privacy from cameras and location data, cybersecurity, effects on drivers’ jobs, accessibility for people who cannot drive, and impacts on cities and public transport.
Operational design domain
A key idea in autonomous driving is the operational design domain, the set of conditions in which a system is designed to work: certain roads, speed ranges, weather, times of day and geographic areas. A Level 4 robotaxi may drive confidently in a mapped part of a city in good weather and decline to operate in heavy snow or on unmapped roads. Describing the domain clearly lets operators, regulators and passengers know what to expect, and expanding it gradually, as the system proves itself, is how most services grow. Systems also need a minimal risk maneuver, such as pulling over safely, for when they reach the edge of their domain or detect a fault.
Remote assistance and fleet operations
Even driverless vehicles are not entirely alone. Many operators maintain remote assistance teams that can answer a car’s request for guidance in unusual situations, such as a blocked lane or an unclear construction zone, by suggesting a path that the car then evaluates and follows by itself. Fleet operations also handle charging, cleaning, maintenance and software updates, and they monitor vehicles for faults. Safe operation therefore depends on a whole organization and not only on the technology in the car, in keeping with the idea of supervised autonomy in human-in-the-loop AI.
Impact and outlook
If automated driving fulfils its promise, it could reduce road deaths, extend mobility to older people and those with disabilities, and change how cities use space. Risks include job losses for professional drivers, more total travel and congestion if rides become cheap, and uneven access. Progress has been slower than many early forecasts, as the last few percent of rare situations proved hard. Most observers expect gradual growth, with driver assistance becoming universal and driverless services expanding area by area.
FAQs
What is a self-driving car? It is a vehicle that uses sensors, software and AI to drive with little or no human input.
What are the levels of self-driving? SAE defines six levels, from 0 with no automation to 5 with full automation in all conditions. Most consumer cars today are at Level 1 or 2.
How do self-driving cars see? They combine cameras, radar, lidar and other sensors, and use neural networks to detect and track objects.
Are self-driving cars safe? It depends on the system and setting. Evidence is still developing, and partial automation still needs an attentive driver.
Are there fully driverless cars today? Yes, in limited areas. Robotaxi services operate without a driver at Level 4 in selected cities, but fully general Level 5 driving does not exist.
More in Robotics and Autonomous Technology
Key facts
- The SAE J3016 standard defines six levels of driving automation, from Level 0, no automation, to Level 5, full automation in all conditions1.
- Stanford's vehicle Stanley won the 2005 DARPA Grand Challenge, and in the 2007 Urban Challenge vehicles had to drive autonomously in a mock city2.
- Waymo began offering fully driverless ride-hailing to the public in Phoenix in 20203.
Related concepts
Quick checkWhat is a self-driving car?Show answer
A self-driving car is a vehicle that uses sensors such as cameras, radar and lidar together with software and artificial intelligence to perceive its surroundings, plan a route and control steering, braking and acceleration with little or no human input.
Sources
- SAE International, J3016 Taxonomy and Definitions for Terms Related to Driving Automation Systems, https://www.sae.org/standards/content/j3016_202104/
- Defense Advanced Research Projects Agency, The DARPA Grand Challenge and Urban Challenge, https://www.darpa.mil/about/history/archives
- Waymo, Waymo One launches public fully autonomous service in Phoenix, https://waymo.com/blog/2020/10/waymo-is-opening-its-fully-driverless
Editorially reviewed by Specialty Digest Editorial TeamLast reviewed November 17, 2026Researched and drafted with AI assistanceReport an issue
