How Self-Driving Cars Work and Where Autonomous Vehicle Technology Stands in 2026

Autonomous vehicles do not drive the way people imagine from science fiction. Every system on the road today, from a Tesla on Autopilot to a driverless Waymo robotaxi, relies on the same basic pipeline: sensors collect raw data, software turns that data into a model of the surrounding world, and a planning system decides how to steer, accelerate or brake. The difference between systems is how much of that work the car does on its own, and how much a human still has to supervise.

How the technology actually works

Three sensor types do most of the perception work. Cameras capture visual detail and are needed to read traffic-light colors, road signs and lane markings, but they struggle to judge exact distance and perform worse in fog, glare or heavy rain, according to Nvidia's technical explainer on self-driving perception. Radar sends out radio waves that bounce back off objects, revealing their speed and position even at night or in poor weather, though it cannot tell what kind of object it detected. Lidar fires laser pulses to build a continuous 3D map of the environment and works well in low light, but the sensor units cost more than cameras or radar.

Because no single sensor is reliable in every condition, autonomous vehicles use what engineers call sensor fusion: a central computer combines the camera, radar and (where fitted) lidar data streams into one unified picture, then runs that picture through software that detects objects, predicts how they will move, and plans a path through them, Nvidia's explainer states. This redundancy is meant to let the car keep functioning if one sensor is blocked or degraded.

The industry has split over whether lidar is necessary at all. Waymo and most robotaxi operators keep lidar in their sensor suite for its higher-resolution 3D view, while Tesla has built its Full Self-Driving system on cameras alone, arguing that if humans can drive using vision only, cars should eventually be able to as well, according to background in the Wikipedia overview of self-driving cars.

The SAE levels: who is actually driving

The automotive and tech industries classify automation using the SAE J3016 standard, which defines six levels from 0 to 5. The dividing line that matters most to drivers falls between Level 2 and Level 3.

No production car on sale anywhere has reached Level 5. Most driver-assistance systems sold today, including Tesla's Autopilot and Full Self-Driving, are Level 2: the driver is legally required to stay attentive even when the car is steering and braking itself.

Level 3 just lost its only US example

Mercedes-Benz's Drive Pilot was, until January 2026, the only certified Level 3 system available to US consumers, letting drivers legally look away from the road on approved California and Nevada highway stretches below about 40 mph, in daylight, in good weather and clear lane markings. In January 2026 Mercedes-Benz paused the system and will not offer it in the facelifted S-Class, according to InsideEVs. The company cited the high cost of the hardware, a roughly $2,500 annual subscription that drew limited customer interest, and the collapse of its lidar supply deal after partner Luminar failed to meet contract terms. Mercedes is replacing it with MB.Drive Assist Pro, a Level 2++ system using ten cameras, five radar sensors and twelve ultrasonic sensors that still requires the driver's hands and attention.

Where driverless robotaxis operate today

Level 4 driverless taxi service, where no one sits in the driver's seat at all, exists only within geofenced areas. Waymo, which launched the first driverless taxi service in Phoenix in 2017, operated in 14 US cities as of September 2026, including Phoenix, San Francisco, Los Angeles, Austin, Atlanta, Dallas, Houston, Miami, Nashville, Orlando, San Antonio, Denver, San Diego and Tampa, according to TechCrunch. Its fleet passed 4,000 vehicles, combining Jaguar I-Pace models with a newer Zeekr-built minivan called the Ojai. In China, DeepRoute.ai runs a comparable driverless service in Shenzhen, according to the Wikipedia overview of self-driving cars.

The road to driverless commercialization has also included high-profile retreats. General Motors stopped funding robotaxi development at its Cruise unit in December 2024, folding the division into GM's personal-vehicle driver-assistance work instead, TechCrunch reported; the move left minority investor Microsoft absorbing an $800 million impairment charge on its Cruise stake.

How regulators are responding

US regulation has been moving to accommodate Level 4 deployment rather than Level 3. On July 30, 2026, the National Highway Traffic Safety Administration issued its first commercial exemption letting robotaxi maker Zoox manufacture vehicles that do not meet several Federal Motor Vehicle Safety Standards built around human controls, according to a summary from the National Law Review. The same week NHTSA opened a rulemaking process and a three-year, $5 million partnership with the SAE Industry Technologies Consortia, called ASCEND, to gather data toward the first dedicated performance standards for automated driving systems. On July 31, 2026, NHTSA also issued an interim rule letting vehicles built before an exemption is granted still qualify for commercial-deployment exemptions, a change from the prior approach that required the exemption first.

Practical steps for a reader evaluating these systems

1. Check the SAE level stated by the manufacturer, not just marketing names like "Full Self-Driving" or "Autopilot" — most consumer systems sold today are Level 2 and legally require constant driver attention.
2. Look up the system's operational design domain (ODD): the specific roads, speeds, weather and lighting conditions it is certified for, since no current system works everywhere.
3. For Level 2 systems, keep hands and attention on the task regardless of how the system behaves, since liability and legal responsibility remain with the driver.
4. For Level 4 robotaxi rides, confirm the service area on the operator's own app or site, since coverage is geofenced and changes as cities are added.
5. Treat any company's full-autonomy timeline as a forecast, not a guarantee — Level 3 has already been withdrawn from the one country where it was sold to individual drivers.