A self-driving car, also called an autonomous vehicle, is a road vehicle equipped with sensors, computing, and software that allow it to perceive its surroundings and control steering, acceleration, and braking with reduced or eliminated human input. The Society of Automotive Engineers defines six levels of driving automation, from Level 0, no automation, through Level 2, partial automation requiring constant driver supervision, to Level 5, full automation under all conditions without a human driver; as of the mid-2020s, no commercially deployed system had reliably reached Level 5, and most publicly available systems operated at Level 2 or, in limited geofenced areas, Level 4.
From DARPA to commercial deployment
Modern self-driving research traces to the DARPA Grand Challenge, a series of races for autonomous vehicles across desert terrain that DARPA staged in 2004 and 2005 after no team completed the first course; the 2005 winner, Stanford's Stanley, led by Sebastian Thrun, and the subsequent 2007 Urban Challenge helped seed the talent and technology base for the modern industry. Google started its self-driving car project in 2009 under Thrun, which was spun out as Google DeepMind sibling company Waymo in 2016. Waymo became the first company to operate a driverless commercial robotaxi service without a human safety driver, launching in Phoenix, Arizona in 2020 and later expanding to San Francisco, Los Angeles, and additional U.S. cities. Cruise, backed by General Motors, ran a competing robotaxi service until a 2023 pedestrian-dragging incident in San Francisco led to a nationwide suspension of its fleet and major leadership and strategy changes.
Tesla's approach
Tesla, under Elon Musk, pursued a different technical path, relying on camera-based Computer vision rather than the lidar sensors used by Waymo and most other developers, and shipping its driver-assistance features, marketed as Autopilot and Full Self-Driving, incrementally to consumer vehicles rather than restricting the technology to a small geofenced robotaxi fleet. Despite the "Full Self-Driving" branding, the system has required active driver supervision throughout its public rollout and remains classified as Level 2 automation; the approach and its marketing have drawn regulatory scrutiny, including investigations by the U.S. National Highway Traffic Safety Administration into crashes involving the feature.
Technical approach
Self-driving systems generally combine several kinds of sensors, cameras, radar, and, in most approaches other than Tesla's, lidar, fused through perception software built on computer vision and, increasingly, learned Neural network models rather than purely hand-coded rules. Path planning and control layers then translate perceived surroundings into driving decisions, with recent systems increasingly using end-to-end learned models that map sensor input more directly to driving actions, an approach conceptually related to the broader push toward learned, rather than hand-engineered, control across Robotics.
The long tail problem
The central technical obstacle to full autonomy is often described as the long tail: the vast majority of driving situations are routine and already handled reliably, but a small, effectively unbounded set of rare edge cases, unusual road markings, erratic other drivers, unpredictable pedestrian behavior, or unfamiliar weather, continue to cause failures, and covering enough of that tail to match or exceed human safety across all conditions has proven far more difficult than early industry predictions from the 2010s anticipated. This gap between demonstrated capability in common conditions and robustness in rare ones parallels similar reliability challenges observed in other AI domains, including Hallucination (AI) in Large language model systems, and has repeatedly pushed back timelines for full Level 5 deployment across the industry.