An autonomous vehicle (AV), often called a self-driving car, is a vehicle whose automated driving system performs the entire dynamic driving task within a defined set of conditions, without a human being required to monitor the road or act as a fallback. In autonomous driving terminology, this corresponds to SAE Level 4 or Level 5; Level 3 systems are usually described as conditionally automated rather than fully autonomous.
The defining feature is transfer of responsibility. A car with advanced driver assistance still needs an attentive driver, while an autonomous vehicle detects hazards, decides and executes maneuvers, and reaches a safe stop on its own if something fails. Most current AVs are electric, because battery drivetrains offer the power, control precision and drive-by-wire interfaces the computing stack needs.
How an autonomous vehicle perceives, plans and acts
An autonomous vehicle runs a continuous loop of perception, prediction, planning and control, repeated many times per second. Each stage can be a separate software module or part of a single learned model, but the functional tasks are the same.
- Perception fuses camera, radar and often lidar data into a model of lanes, vehicles, pedestrians, signs and free space, frequently in a bird’s-eye view representation or an occupancy network.
- Localization determines the vehicle’s position to within centimeters, using satellite positioning, inertial sensors, odometry and matching against an HD map where one is used.
- Prediction estimates the likely future trajectories of other road users over the next few seconds.
- Planning applies the driving policy to choose a safe, legal and comfortable trajectory.
- Control turns that trajectory into steering, braking and propulsion commands through redundant actuators.
Recent systems increasingly use end-to-end autonomous driving models that learn several of these stages jointly from large volumes of driving data. Modular stacks remain common where engineers need explicit, verifiable safety rules between stages.
Sensors and computing behind the vehicle
No single sensor sees everything reliably, so autonomous vehicles combine modalities with complementary strengths and weaknesses. The sensor suite is one of the main design choices separating developers.
| Sensor | Strength | Weakness |
|---|---|---|
| Camera | Color, text, lane markings, high resolution | No direct depth, sensitive to glare and darkness |
| Radar | Direct velocity, works in rain and fog | Low angular resolution |
| Lidar | Precise 3D distance measurement | Cost, reduced range in heavy precipitation |
| Ultrasonic | Cheap close-range detection | Only a few meters of range |
| GNSS and IMU | Absolute position and motion | Drift and loss in tunnels and urban canyons |
Processing these streams requires an onboard computer delivering hundreds to thousands of trillions of operations per second, drawing from several hundred watts up to more than a kilowatt. In an electric vehicle this load is visible as a measurable reduction in range, which makes compute efficiency a practical engineering concern, not only a cost issue.
Levels of automation and the operational design domain
The widely used SAE J3016 taxonomy describes levels of driving automation from 0 to 5. Only the top two levels describe a true autonomous vehicle, because only there is no human fallback required within the system’s domain.
| Level | System capability | Human role |
|---|---|---|
| Level 2 | Steering and speed support | Driver monitors at all times |
| Level 3 | Drives itself in limited conditions | Driver must take over on request |
| Level 4 | Drives itself fully within its ODD | No fallback driver needed inside ODD |
| Level 5 | Drives anywhere a human could | Occupants are passengers only |
The operational design domain (ODD) sets the boundaries: road types, geographic area, speed range, weather, lighting and traffic conditions. A robotaxi may be Level 4 in one city at speeds up to about 70 km/h, yet unable to operate outside that geofence. No Level 5 system exists in commercial service, and many engineers doubt it is a meaningful near-term target.
Deployment models and how they differ
Autonomous vehicles are reaching the road through several business models, each with a different ODD and a different relationship with a human operator. The table compares the main ones.
| Deployment model | Typical ODD | Human involvement |
|---|---|---|
| Robotaxi service | Geofenced urban areas | Remote assistance, no driver on board |
| Autonomous trucking | Highway corridors between hubs | Human drivers for first and last mile |
| Shuttles and delivery robots | Fixed routes, low speeds | Remote monitoring of fleets |
| Private vehicles | Mostly highways at present | Driver available as fallback at lower levels |
Fleet operators rely on remote assistance to resolve rare situations, such as a blocked street, by giving high-level guidance rather than steering. Remote driving, where a human controls the car directly over a network, is a different concept with stricter latency requirements. During testing, a safety driver sits behind the wheel and each manual takeover is logged as a disengagement.
Safety validation and open challenges
Proving that an autonomous vehicle is safe enough is harder than building one that usually works. Serious crashes are rare per kilometer, so demonstrating statistical superiority over human drivers through road testing alone would require an impractically large distance. Developers therefore combine simulation, scenario-based testing, closed-course trials and staged public deployment.
Standards help structure this work. ISO 26262 covers functional safety of electrical and electronic systems, and ISO 21448, known as SOTIF, addresses hazards from performance limitations such as a misclassified object. Remaining challenges include the long tail of rare events, unusual construction zones, interaction with human gestures, adverse weather, cybersecurity and the legal question of liability after a crash.
A frequent misconception is that a high disengagement-free distance proves safety. Disengagement figures depend heavily on the ODD and reporting rules, so they are poor comparisons between companies and say little about rare, severe failures.
In plain words
An autonomous vehicle is a car that can drive itself without anyone needing to watch the road, at least in the places and conditions it was designed for. It uses cameras and other sensors as its eyes and a powerful computer as its brain.
The key difference from ordinary driver assistance is who is in charge. With assistance, you are still the driver. In a true autonomous vehicle, the car is the driver, and the people inside are simply passengers, much like riders in a taxi.
A real-world example
In several large cities, electric robotaxis operate within a mapped service area of a few hundred square kilometers. A passenger opens an app, requests a ride, and an empty car arrives at the curb. There is no one in the driver’s seat; the vehicle navigates intersections, yields to pedestrians and handles unprotected left turns on its own.
When the car meets a lane closed by emergency vehicles, it slows, stops safely and asks a remote assistance operator to confirm an alternative path. The operator approves a detour, and the vehicle completes the trip itself. Outside the geofence, or in conditions outside its ODD, the service simply does not offer rides, which illustrates how autonomy today is defined by its domain.