Autonomous vehicles

An autonomous vehicle is a vehicle capable of sensing its environment and operating without continuous human input. Such vehicles are also called driverless cars or self-driving cars. They perceive their surroundings using cameras, radar, lidar and ultrasonic sensors, interpret that data in software to plan a path, and control steering, acceleration and braking on their own.

Autonomy is being applied not only to passenger cars but also to trucks, buses, shuttles, delivery robots, agricultural machinery, mining equipment and military vehicles. Forecasts for full commercialisation still vary widely, but driverless robotaxi services operating within limited areas and conditions have already entered commercial operation in several cities.

History

Early experiments

The idea of a vehicle that moves by itself dates back to the early twentieth century. At the 1939 New York World's Fair, General Motors' Futurama exhibit imagined cars guided by electronic devices embedded in the roadway. During the 1950s and 1960s, experiments in the United States and the United Kingdom tested vehicles that followed cables buried in the road surface.

In the 1980s, vehicles began to perceive their environment for themselves. A team led by Ernst Dickmanns at Bundeswehr University Munich developed VaMoRs, an experimental vehicle using computer vision, which in 1987 drove autonomously on a motorway at more than 90 km/h. In the United States, Carnegie Mellon University's NavLab project and DARPA's Autonomous Land Vehicle programme pursued similar goals.

The DARPA challenges

The turning point for modern research was a series of driverless races organised by DARPA. No vehicle finished the first Grand Challenge in 2004, but in 2005 five vehicles completed the desert course, including Stanford University's Stanley. The 2007 Urban Challenge introduced traffic rules and other moving vehicles, establishing much of the technical foundation for later commercial development. Many participants went on to lead autonomous driving teams at companies such as Google and Uber.

Commercialisation

Google began its self-driving car project in 2009; it was spun out as Waymo in 2016. From the late 2010s, Waymo, GM Cruise, Baidu Apollo Go and Zoox began pilot and commercial operations in selected cities in the United States and China. Tesla took a different route, shipping Autopilot and Full Self-Driving features on consumer vehicles and accumulating large volumes of real-world data.

Levels of automation

The J3016 standard published by SAE International defines six levels of driving automation, from 0 to 5, and is widely used internationally.

• Level 0 (no automation): the driver performs all tasks. Warning systems such as forward collision warning and lane departure warning fall in this category.
• Level 1 (driver assistance): the system assists with either steering or acceleration and braking. Adaptive cruise control and lane keeping assist are examples.
• Level 2 (partial automation): the system assists with steering and speed simultaneously, but the driver must monitor the environment at all times and be ready to intervene. Most advanced driver assistance systems on sale today belong here.
• Level 3 (conditional automation): within specific conditions the system handles driving, and the driver intervenes only when requested. It has been offered on a few production cars in limited situations such as congested motorways.
• Level 4 (high automation): within a defined operational design domain (ODD), the vehicle drives without human intervention and brings itself to a safe stop if a problem occurs. Robotaxis and autonomous shuttles are typical examples.
• Level 5 (full automation): the vehicle can drive anywhere a human could, without restriction. This has not yet been achieved.

The distinction between Level 2 and Level 3 is one of responsibility. At Level 2 the human continues to supervise the drive; from Level 3 onwards, responsibility for the driving task transfers to the system while it is engaged.

Technology

Sensors

• Cameras provide rich visual information about traffic lights, signs, lane markings and pedestrians, but are affected by lighting and weather.
• Radar uses radio waves to measure range and relative speed. It performs well in bad weather but has low resolution.
• Lidar scans the surroundings as a three-dimensional point cloud using laser pulses. It offers precise range measurement; once expensive, its cost has fallen sharply with solid-state designs.
• Ultrasonic sensors detect nearby obstacles and assist with parking.
• Satellite navigation (GNSS) and inertial measurement units (IMU) estimate the vehicle's absolute position and attitude.

Software pipeline

A typical autonomous driving stack comprises the following modules.

1. Perception: fusing sensor data to detect and classify vehicles, pedestrians, cyclists, lane markings and signals.
2. Localisation: matching sensor data against a high-definition map to determine position to lane level, and sometimes to centimetres.
3. Prediction: estimating how surrounding objects will move over the next few seconds.
4. Planning: deciding the route to the destination and behaviours such as lane changes, stopping and yielding.
5. Control: producing steering, throttle and brake commands to follow the planned trajectory.

Recent work has focused on end-to-end learning, which merges these stages into a single neural network, and on large-scale simulation for training on rare situations.

Competing approaches

Two broad strategies exist in industry. One relies heavily on lidar and high-definition maps to achieve high reliability in a limited area before expanding it (Waymo, Baidu and others); the other uses a lower-cost, camera-centric configuration and gathers data from a large fleet of production cars to pursue generality (Tesla). Which will reach full autonomy first remains debated.

Applications

• Robotaxis: driverless ride-hailing services operate in Phoenix, San Francisco and Los Angeles in the United States, and in Wuhan and Beijing in China.
• Autonomous freight: long-haul trucking on motorways involves a simpler environment and is expected to commercialise relatively early.
• Last-mile delivery: small delivery robots and driverless vans are being trialled on campuses and in residential areas.
• Public transport shuttles: low-speed shuttles running fixed routes are used at airports and industrial parks.
• Industrial sites: driverless haul trucks in mines, autonomous tractors and combines, and unmanned yard tractors at ports operate in controlled environments and are already substantially commercialised.

Potential benefits

• Improved safety: since most road crashes stem from human error, systems that do not tire, drink or become distracted are expected to reduce collisions.
• Greater mobility: older people, people with disabilities and those without a licence gain freedom of movement.
• Efficiency: cooperative driving, congestion relief, reduced parking demand and lower fuel consumption are frequently cited.
• Productivity: travel time can be spent on other activities.
• Lower logistics costs: driver wages and mandatory rest periods become less constraining.

Challenges and limitations

Technical difficulties

The hardest problem is the long tail. Ninety-nine per cent of ordinary driving is comparatively easy, but rare cases are effectively unlimited: unexpected obstacles, a worker directing traffic by hand at a construction site, lane markings covered by snow, or telling a plastic bag apart from a real obstacle. Heavy rain, snow, dense fog and low sun degrade sensor performance.

Validation and safety assurance

Some analyses suggest that proving statistically that a system is safer than a human driver would require hundreds of millions of kilometres of driving. This has driven research into validation regimes combining simulation, scenario-based testing and formal verification.

Security

Connected vehicles are exposed to remote intrusion, sensor spoofing and software supply chain attacks, which has led to automotive cybersecurity standards such as UNECE R155 and ISO/SAE 21434.

Ethics and liability

The so-called trolley problem — what a system should choose when a collision is unavoidable — is often discussed. In practice, however, the more consequential issue is the allocation of legal liability among manufacturer, software supplier, owner and occupant when a crash occurs.

Regulation and law

The 1968 Vienna Convention on Road Traffic required every vehicle to have a driver, but was later amended to accommodate automated systems. The United Nations Economic Commission for Europe adopted Regulation R157 covering Level 3 Automated Lane Keeping Systems.

In the United States, testing and deployment are permitted state by state, with California, Arizona and Texas among the most active. Germany passed legislation in 2021 allowing Level 4 operation in defined areas; Japan amended its road traffic law to permit Level 3 and limited Level 4 operation. China has approved commercial robotaxi services within designated demonstration zones in several cities. South Korea designates pilot operation districts and issues temporary operating permits under its autonomous vehicle act.

Safety record and notable incidents

Crashes during development have strongly influenced regulation and public opinion. In 2018 an Uber test vehicle in Tempe, Arizona, struck and killed a pedestrian crossing the road, the first pedestrian death involving an autonomous test vehicle; the National Transportation Safety Board criticised the operator's safety culture and inadequate monitoring of its safety drivers. Several fatal crashes have also involved drivers who over-relied on partial automation, underlining the need for driver monitoring systems. In 2023 one robotaxi operator had its permits suspended over its handling of a collision.

Social and economic effects

Widespread autonomy is expected to have far-reaching effects on cities and labour markets. Reduced parking demand could reshape urban space, and greater use of shared fleets could lower private car ownership. Others argue that cheaper travel would instead increase trips and empty running, worsening congestion. Job transitions for professional drivers, structural change in the insurance industry, and a reordering of the automotive and software value chain are further points of contention.

Public perception

Surveys consistently show that trust in autonomous vehicles remains low, with large shares of respondents saying they would feel uneasy riding in a fully driverless car, and reported crashes markedly reducing confidence. Satisfaction among people who have actually used robotaxis is reported to be comparatively high, suggesting that direct experience strongly shapes acceptance.