Autonomous driving

What it is

Autonomous driving means a computer system, rather than a person, handles some or all of the task of driving a road vehicle. How much of the task is handed over is conventionally described on a six-level scale published by the US National Highway Traffic Safety Administration (NHTSA), running from Level 0 to Level 5^[5]. At Levels 0 to 2 the driver drives and monitors the vehicle, and the system only assists with warnings, braking or steering^[5]. At Level 3 the system performs the driving task, but the driver must be available to take over when asked^[5]. At Level 4 the system handles all driving tasks within limited service areas, and occupants act only as passengers^[5]. At Level 5 the system can drive everywhere, under all conditions and on all roadways, without a human driver^[5]. NHTSA says it follows industry practice in avoiding the term self-driving for higher levels of automation, because that term describes a state of operation rather than a capability^[5].

The scale matters because it splits one broad phrase into very different engineering problems. A system that keeps a car in lane on a motorway and a system that drives through a crowded city without anyone aboard are both called automation, but they require different sensors, different software and different assumptions about who is responsible when something goes wrong.

How it works

Two technical traditions have shaped how such systems are built. One is a modular pipeline that separates perception, prediction and planning into distinct stages: the vehicle first works out what is around it, then estimates what other road users are likely to do, then chooses a path^[2][4]. The other is end-to-end learned driving, in which a model is trained to map what the sensors take in directly to driving actions, without hand-separated stages^[2][4].

The difference becomes concrete once other traffic is involved. The 2007 Urban Challenge added traffic laws and other moving vehicles to the course, so a vehicle had to make real-time decisions about what other drivers were doing rather than simply following a clear path^[4]. Prediction, in other words, is not a detail bolted onto perception; it is part of the driving problem itself.

Where it came from

The research lineage begins with DARPA's Strategic Computing Initiative, starting in 1984, which produced the first autonomous ground vehicles able to drive on and off roads, leading to the Autonomous Land Vehicle and Navlab^[1]. Parallel work was pursued in Germany by Ernst Dickmanns and in Italy through the ARGO project, alongside the EUREKA Prometheus programme^[1].

These vehicles then left the test track. In 1994 Dickmanns' VaMP and its twin VITA-2 drove more than 1,000 km in normal highway traffic at up to 130 km/h, with a safety driver aboard^[2]. A year later the same line of work covered more than 2,000 km from Munich to Copenhagen and back at up to 180 km/h^[2]. A 2024 history of European vehicle automation research by Fabian Kröger covers the EUREKA PROMETHEUS programme and the performance of the VITA-II and VaMP research vehicles, and places the long-distance drives by VaMP and by the Italian ARGO vehicle in 1995 and 1998^[6].

The American programme then shifted to competition as a driver. DARPA's 2005 Grand Challenge was the first time multiple vehicles completed a long off-road driverless course, with five of 23 finalists finishing 212 km^[3]. The 2007 Urban Challenge added traffic laws and other moving vehicles^[4]. Many of the teams that took part in these events later carried the same people and methods into commercial work.

What limits it

The field's open problem is the long tail: rare situations that no training set covers, which is why safety drivers and remote operators persist in commercial services^[2][4]. A system can handle the ordinary mile after mile and still meet the one junction, the one fallen sign or the one unusual vehicle that its data never described. Until that tail is addressed, the six-level scale remains a description of operating conditions rather than a promise that a vehicle can go anywhere.