Autonomous driving
Last reviewed
Sources
22 citations
Review status
Source-backed
Revision
v7 ยท 7,393 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Sources
22 citations
Review status
Source-backed
Revision
v7 ยท 7,393 words
Add missing citations, update stale details, or suggest a clearer explanation.
Autonomous driving is the capability of a vehicle to perceive its environment, make decisions, and control its own motion without human input, using a combination of artificial intelligence, sensors, and software. Also known as self-driving technology, it is one of the most ambitious applications of AI, integrating computer vision, sensor fusion, machine learning, and robotics into systems that must operate safely in complex, unpredictable real-world conditions. Capabilities are classified on the six-level SAE J3016 scale, from Level 0 (no automation) to Level 5 (full automation). [1]
As of early 2026, autonomous driving has moved from research into commercial deployment at scale: Waymo provides roughly 500,000 paid robotaxi rides per week across about 10 US cities and has driven more than 170 million rider-only miles without a human driver. [9][18] Tesla's Full Self-Driving (Supervised) fleet crossed 10 billion cumulative miles in May 2026, and Baidu's Apollo Go in China has surpassed 20 million cumulative rides. [11][19][12] No system has achieved Level 5 (unrestricted) autonomy in commercial use.
The development of autonomous vehicles (AVs) has progressed from early academic research and military-funded competitions in the 2000s to commercial robotaxi services operating in multiple cities around the world as of 2026. Companies such as Waymo, Tesla, Zoox, Baidu, and Aurora are deploying autonomous systems at increasing scale, while regulatory frameworks continue to evolve to keep pace with the technology.
The SAE International standard J3016, first published in 2014 and most recently updated in April 2021, defines six levels of driving automation ranging from Level 0 (no automation) to Level 5 (full automation). [1] This taxonomy has become the globally accepted framework for classifying autonomous driving capabilities and is referenced by regulators, automakers, and researchers worldwide.
The six levels are divided into two broad groups. At Levels 0 through 2, the human driver performs all or part of the dynamic driving task (DDT). At Levels 3 through 5, an automated driving system (ADS) performs the entire DDT while engaged. [1] As SAE J3016 states, the levels "describe the full range of driving automation features" and were "developed to provide a common taxonomy and definitions" so that the industry, regulators, and the public share a consistent vocabulary. [1]
| SAE Level | Name | Description | Driver Role | Example Systems |
|---|---|---|---|---|
| Level 0 | No Driving Automation | The human driver performs all driving tasks. The vehicle may have warning systems but no sustained control. | Full control at all times | Automatic emergency braking, blind spot warning |
| Level 1 | Driver Assistance | The system can assist with either steering or acceleration/braking, but not both simultaneously. | Monitors environment and performs remaining tasks | Adaptive cruise control, lane keeping assist |
| Level 2 | Partial Driving Automation | The system can control both steering and acceleration/braking simultaneously under certain conditions. The human must monitor the environment at all times. | Must supervise and be ready to intervene | Tesla Autopilot, GM Super Cruise, Ford BlueCruise |
| Level 3 | Conditional Driving Automation | The system performs the full DDT within its operational design domain (ODD). The human must be ready to resume control when the system requests. | Fallback-ready; does not need to monitor constantly | Mercedes-Benz Drive Pilot (approved in Germany and select US states) |
| Level 4 | High Driving Automation | The system performs the full DDT and can handle fallback situations within a defined ODD. No human intervention is required within that domain. | Not required within ODD | Waymo Driver, Zoox robotaxi, Baidu Apollo |
| Level 5 | Full Driving Automation | The system can perform the full DDT under all conditions that a human driver could handle. No operational design domain restrictions. | Not required under any conditions | No commercially deployed system as of 2026 |
As of early 2026, the most widely deployed autonomous systems operate at Level 2 (Tesla Autopilot, GM Super Cruise) and Level 4 (Waymo, Baidu Apollo Go) within geofenced operational areas. No system has achieved Level 5 autonomy in commercial deployment.
The concept of self-driving vehicles dates back decades. In the 1980s, Ernst Dickmanns at the Bundeswehr University Munich developed VaMoRs, a Mercedes van equipped with cameras and computer vision that could drive on highways. Carnegie Mellon University's Navlab project, running from 1984 through the 1990s, produced a series of autonomous vehicles that demonstrated increasingly capable self-driving on roads. In 1995, the Navlab 5 completed a cross-country trip from Pittsburgh to San Diego, with the vehicle autonomously steering for 98.2% of the 2,849-mile journey, though a human controlled throttle and brakes.
The modern era of autonomous vehicle development was catalyzed by the DARPA Grand Challenge, funded by the United States Department of Defense. [2] The 2001 National Defense Authorization Act set a goal that one-third of operational ground combat vehicles be unmanned by 2015, and DARPA created the Grand Challenge to spur innovation. [2]
The first competition took place on March 13, 2004, in the Mojave Desert. Teams were challenged to build vehicles capable of autonomously navigating a 150-mile off-road course from Barstow, California, to Primm, Nevada. None of the 15 competing vehicles finished the route. Carnegie Mellon University's Sandstorm, a converted Humvee, traveled the farthest at 7.32 miles (11.78 km) before getting stuck. The $1 million prize went unclaimed. [2]
The second competition on October 8, 2005, demonstrated dramatic progress. Five vehicles completed a 132-mile desert course in southern Nevada. Stanford University's "Stanley", a modified Volkswagen Touareg led by Sebastian Thrun, finished first with a time of 6 hours and 53 minutes, winning the $2 million prize. [3] Carnegie Mellon's two entries, Sandstorm and Highlander, finished second and third. Stanley's victory demonstrated the effectiveness of probabilistic reasoning, sensor fusion, and machine learning for autonomous navigation. [3]
The 2007 Urban Challenge raised the bar significantly by requiring vehicles to navigate a simulated urban environment with traffic, intersections, and parking. The competition took place on November 3, 2007, at the former George Air Force Base in Victorville, California. Eleven teams competed on a 60-mile urban course. [4]
Carnegie Mellon's "Boss", a modified Chevrolet Tahoe developed by the Tartan Racing team in collaboration with General Motors, won the $2 million first prize, finishing 20 minutes ahead of the runner-up. [4] Stanford's "Junior" (a Volkswagen Passat) took second place ($1 million), and Virginia Tech's "Odin" (a Ford Escape Hybrid) placed third ($500,000). [4]
In January 2009, Google launched its self-driving car project at the secretive Google X lab. The project was led by Sebastian Thrun, who had led Stanford's Stanley team, along with Anthony Levandowski. Google co-founders Sergey Brin and Larry Page challenged the team to complete ten 100-mile autonomous routes across California without human intervention. By December 2009, the team completed the first route, and by mid-2010 all ten had been finished.
The project was publicly revealed in October 2010. Google spent approximately $1.1 billion on the project between 2009 and 2015. In the fall of 2015, the project achieved a milestone when a fully autonomous vehicle carried a non-employee passenger on a public road for the first time in Austin, Texas. The passenger was a blind man, highlighting the technology's potential to provide mobility to those unable to drive.
In December 2016, the project was spun out of Google as Waymo, a standalone subsidiary of Alphabet.
The perception stack is the collection of sensors and software that enables an autonomous vehicle to understand its surroundings. Modern autonomous vehicles rely on multiple complementary sensor types, each with distinct strengths and limitations.
Cameras provide rich visual information including color, texture, and fine-grained detail. They are essential for tasks such as reading traffic signs, detecting traffic lights, recognizing lane markings, and classifying objects by type. Modern AV platforms use between 8 and 30 cameras positioned around the vehicle to provide 360-degree coverage. Waymo's sixth-generation system, for example, uses 13 cameras with 17-megapixel sensors. [9]
The primary limitations of cameras include sensitivity to lighting conditions (glare, darkness, low contrast) and the difficulty of directly measuring depth from a single image, though stereo camera setups and monocular depth estimation using deep learning can partially address this.
LiDAR (Light Detection and Ranging) sensors emit pulsed laser beams and measure the time each pulse takes to return after reflecting off objects. This produces a dense 3D point cloud representing the geometry of the surrounding environment with centimeter-level accuracy. LiDAR operates effectively in both daylight and darkness and provides precise distance measurements.
Modern LiDAR sensors can detect objects hundreds of meters away. Aurora's next-generation FirstLight LiDAR, for instance, can detect objects at distances up to 1,000 meters. [10] LiDAR point clouds are widely used for 3D object detection, localization against HD maps, and obstacle avoidance.
The main drawbacks of LiDAR include higher cost compared to cameras and radar, reduced performance in heavy rain, fog, or snow (which can scatter laser pulses), and lower resolution compared to cameras for texture and color information.
Radar (Radio Detection and Ranging) uses radio waves to detect the position, distance, and velocity of objects. Radar is highly robust to adverse weather conditions, including rain, fog, snow, and dust, making it a reliable complement to cameras and LiDAR. Radar excels at measuring the relative speed of objects through the Doppler effect.
Autonomous vehicles typically use multiple radar units, including long-range forward-facing radar for highway driving and short-range radar for parking and close-proximity detection. Waymo's sixth-generation system uses six radar units. [9]
Ultrasonic sensors emit high-frequency sound waves and measure the time for echoes to return. They are effective at very short ranges (typically under 5 meters) and are commonly used for parking assistance and detecting nearby obstacles at low speeds. While less critical for highway autonomy, they provide useful supplementary data in urban stop-and-go scenarios.
No single sensor type is sufficient for safe autonomous driving. Sensor fusion combines data from multiple sensor modalities to create a more complete and reliable representation of the environment. This process compensates for the weaknesses of individual sensors: cameras provide semantic detail that LiDAR lacks, LiDAR provides accurate depth that cameras struggle with, and radar provides velocity data and weather robustness that both cameras and LiDAR may lack.
Sensor fusion can be performed at different levels:
Modern approaches increasingly use Bird's Eye View (BEV) representations, where data from cameras, LiDAR, and radar is projected into a unified top-down spatial view. Transformer-based architectures have proven particularly effective for multi-modal BEV fusion, enabling structured alignment across different sensor inputs.
Typical Level 4 platforms deploy 15 to 30 cameras, 5 to 20 radar units, and 4 to 7 LiDAR sensors per vehicle.
Computer vision forms the backbone of autonomous vehicle perception. Key tasks include:
Object Detection: Identifying and localizing vehicles, pedestrians, cyclists, traffic signs, traffic lights, and other relevant objects in sensor data. Modern systems use convolutional neural networks and transformer-based architectures. The YOLO (You Only Look Once) family of detectors has been influential for real-time object detection, with recent variants such as YOLO26 supporting unified detection, segmentation, classification, and pose estimation. Transformer-based detectors such as DETR and its successors have also gained traction for their ability to model global context.
Semantic Segmentation: Assigning a class label to every pixel in an image, enabling the vehicle to understand the full scene layout, including drivable surfaces, sidewalks, vegetation, and buildings.
Instance Segmentation: Distinguishing individual instances of the same class (for example, separating two adjacent cars) by combining detection and segmentation.
Lane Detection: Identifying lane boundaries and road markings, which is critical for path planning and lane keeping. Lane detection must handle curved lanes, merging lanes, faded markings, and construction zones.
After perceiving the environment, an autonomous vehicle must predict the future behavior of surrounding agents, including other vehicles, pedestrians, and cyclists. Prediction modules estimate the likely trajectories of each detected agent over a time horizon of several seconds.
Modern prediction systems use neural networks, particularly graph neural networks (GNNs) and transformers, trained on large-scale driving datasets. These models capture interactions between agents (for example, a car yielding to a pedestrian) and contextual cues from the road layout. Trajectory prediction is inherently multimodal, meaning the system must consider multiple possible future paths for each agent and assign probabilities to each.
The planning module determines the autonomous vehicle's trajectory, including speed, lane changes, turns, and stops. Planning must balance multiple objectives: following the intended route, obeying traffic rules, maintaining safety margins, and providing a comfortable ride.
Motion planning approaches include:
Recent systems increasingly integrate prediction and planning into unified frameworks, where predicted trajectories of other agents directly inform the ego vehicle's planning decisions through differentiable optimization.
Localization is the process of determining a vehicle's precise position in the world, typically to within centimeters. Standard GPS provides accuracy of only a few meters, which is insufficient for lane-level positioning. Autonomous vehicles achieve precise localization by combining GPS with inertial measurement units (IMUs), wheel odometry, and matching live sensor data against pre-built high-definition (HD) maps.
HD maps are specialized digital representations that capture road geometry, lane boundaries, traffic sign positions, signal locations, and other features at centimeter-level accuracy. They serve as a reference framework against which the vehicle continuously compares its live sensor readings.
SLAM (Simultaneous Localization and Mapping) techniques allow vehicles to build and update maps while simultaneously determining their position within those maps. HD maps are typically built offline by driving mapping vehicles through the road network multiple times, collecting data from LiDAR, cameras, and GPS, and then processing this data using optimization-based SLAM to produce accurate maps.
Not all autonomous driving approaches depend on HD maps. Tesla's system, for example, does not use pre-built HD maps, instead relying on real-time perception and navigation integrated into its neural network. [11] The trade-off is that HD maps provide a reliable prior for localization and planning but require ongoing maintenance to stay current with road changes.
Two fundamental architectural approaches compete in autonomous driving system design.
The traditional approach decomposes the driving task into a series of specialized modules, each handling a specific subtask: perception (detecting objects and understanding the scene), prediction (forecasting the behavior of other agents), planning (deciding what the vehicle should do), and control (executing the planned trajectory through steering, throttle, and braking commands).
The modular pipeline offers significant advantages in interpretability and debugging. When something goes wrong, engineers can identify which specific module failed. Each module can be developed, tested, and improved independently. However, errors can accumulate and propagate through the pipeline; a perception error may cause a cascade of incorrect predictions and dangerous planning decisions.
The end-to-end approach uses a single neural network (or a tightly integrated set of networks) to map directly from raw sensor inputs to driving outputs (steering, acceleration, braking). This approach avoids the information loss that occurs at module boundaries and can potentially learn more holistic driving strategies from data.
End-to-end systems have shown promising results, particularly as large-scale driving datasets and compute resources have grown. Tesla's FSD system has moved increasingly toward an end-to-end architecture. [11] However, end-to-end systems are harder to interpret and debug due to the black-box nature of deep learning, and they can fail unpredictably when encountering scenarios not well represented in their training data.
As of 2025-2026, the industry is converging on hybrid architectures that blend modular structure with end-to-end learning. For instance, Li Auto's Vision-Language-Action (VLA) architecture integrates end-to-end neural networks with visual language models to enhance spatial understanding. Many systems use learned components within a modular framework, benefiting from both the interpretability of modular design and the flexibility of neural networks.
One of the most debated topics in autonomous driving is the choice of sensor modality.
Tesla has committed to a vision-only approach, arguing that cameras alone can provide sufficient information for autonomous driving, similar to how humans rely primarily on vision. In 2021, Tesla removed radar sensors from new vehicles and later removed ultrasonic sensors, relying entirely on cameras processed by its neural network. [11]
Tesla CEO Elon Musk has argued that LiDAR is an unnecessary "crutch" and that a vision-based system, if sufficiently advanced, should be able to match or exceed human driving performance using cameras alone. Tesla's approach has the advantage of lower hardware cost (cameras are far cheaper than LiDAR) and the ability to leverage the massive fleet of existing Tesla vehicles for data collection.
As of May 2026, Tesla's Full Self-Driving (Supervised) system has accumulated more than 10 billion cumulative miles driven with FSD, a milestone reached on May 3, 2026, with the fleet adding roughly 1 million FSD miles per day. [11][19] Tesla transitioned to a subscription-only model for FSD access in January 2026, removing the one-time purchase option. However, on March 18, 2026, NHTSA escalated its review (campaign EA26002) to a formal engineering analysis covering approximately 3.2 million FSD-equipped vehicles, investigating whether FSD is safe in reduced-visibility conditions such as fog, heavy rain, and sun glare, after documenting crashes in which the system reportedly failed to detect or respond to those conditions. [20] An engineering analysis is the step that typically precedes a recall, highlighting a fundamental concern with purely vision-based systems.
Most other leading autonomous driving companies, including Waymo, Zoox, Aurora, Baidu, Mobileye, and Pony.ai, use LiDAR as a core component of their sensor suites alongside cameras and radar. Proponents argue that LiDAR provides a critical safety layer through its precise 3D measurements, independent of lighting conditions and without the ambiguities inherent in interpreting 2D images.
Waymo's sixth-generation system uses four LiDAR sensors, 13 cameras, and six radar units, representing a reduction from earlier generations while maintaining redundancy across sensor types. [9] Aurora's next-generation FirstLight LiDAR can detect objects at up to 1,000 meters. [10]
The debate remains unresolved. Tesla's approach bets on the long-term trajectory of neural network capabilities, while LiDAR-based approaches prioritize redundancy and geometric precision as an immediate safety measure.
Waymo, a subsidiary of Alphabet, is the most established commercial robotaxi operator. Originating from Google's self-driving car project in 2009, Waymo launched its first fully driverless public service in Phoenix in October 2020. Commercial service expanded to San Francisco in June 2024, Los Angeles in November 2024, Austin in March 2025 (via Uber partnership), and Atlanta in June 2025.
Waymo's weekly paid ridership grew roughly tenfold in under two years, from about 50,000 rides per week in May 2024 to roughly 500,000 paid rides per week across about 10 US cities by late March 2026. [18] Through December 2025, Waymo had driven 170.7 million rider-only miles without a human driver, distributed across Phoenix (68.6M), the San Francisco Bay Area (53.5M), Los Angeles (37.9M), and Austin (10.7M). [9][18] The company announced expansion to Miami, Dallas, Houston, San Antonio, and Orlando, with commercial service in those cities launching during 2026. Additional cities planned for 2026 include Detroit, Las Vegas, Nashville, San Diego, and Washington, D.C. International expansion to London is also planned for 2026.
Waymo's sixth-generation vehicle, the Ojai, is built on a Zeekr platform manufactured in China and outfitted with Waymo's sensor suite at a facility in Arizona. [9] The Ojai features a boxier frame with a lower step and higher ceiling than the previous Jaguar I-PACE fleet, with an estimated per-vehicle cost below $100,000 (down from $150,000 to $200,000 for the I-PACE). Waymo aims to reach 1 million trips per week by the end of 2026.
Tesla's approach to autonomous driving differs fundamentally from robotaxi-focused companies. Rather than purpose-built vehicles operating in geofenced areas, Tesla aims to deliver autonomy through over-the-air software updates to its existing consumer fleet of millions of vehicles.
Tesla's Autopilot (Level 2) provides lane centering and adaptive cruise control. Full Self-Driving (Supervised) adds the ability to navigate city streets, handle intersections, and make turns, but requires a human driver to remain attentive and ready to take over at all times. [11] Tesla's FSD uses an end-to-end neural network architecture processing data from eight cameras, with no LiDAR, radar (removed in 2021-2022), or HD maps. [11]
Tesla unveiled the Cybercab in October 2024, a purpose-built two-passenger robotaxi designed without a steering wheel or pedals. Production is set to begin at Gigafactory Texas in April 2026, with a pilot service launched in Austin, Texas in June 2025 and unsupervised robotaxi rides reported in Dallas and Houston by April 2026. [19] The Cybercab has a planned range of 200 miles (320 km), a 35 kWh battery, and will support inductive charging. However, questions remain about whether Tesla's AI hardware (the AI4 chip currently available) can achieve fully unsupervised autonomy, and Tesla's chairwoman has indicated the company may add a steering wheel and pedals if unsupervised driving is not ready by launch.
Cruise, founded in 2013 and acquired by General Motors in 2016, operated robotaxis in San Francisco, Phoenix, Austin, and Houston before a major incident in October 2023 halted operations. A pedestrian who had been struck by a hit-and-run driver was knocked into the path of a Cruise robotaxi, which ran over her, failed to detect her under the vehicle, and dragged her approximately 20 feet while attempting to pull over. [6]
The aftermath was compounded by Cruise's handling of the incident. The company filed a report with NHTSA that omitted the dragging, and it took 10 days to correct the record. [6] The California DMV suspended Cruise's autonomous vehicle permits. Cruise admitted to filing a false report and agreed to pay a $500,000 criminal fine. NHTSA separately imposed a $1.5 million penalty. [6]
In December 2024, GM announced it would stop funding Cruise as a standalone robotaxi company, having invested over $10 billion in the venture. Cruise was merged back into GM in February 2025, with the technology redirected toward advanced driver assistance systems for personal vehicles rather than robotaxis.
Aurora, founded in 2017 by Chris Urmson (former leader of Google's self-driving car project), Sterling Anderson (former Tesla Autopilot director), and Drew Bagnell (a Carnegie Mellon robotics professor), focuses on autonomous trucking. Aurora launched the first commercial driverless trucking service on public roads in April 2025, hauling freight between Dallas and Houston in Texas. [10]
By April 30, 2026, Aurora had logged more than 5.3 million cumulative commercial miles and added a second driverless route between Fort Worth and El Paso. [21] The company plans to deploy hundreds of trucks during 2026 using its next-generation Aurora Driver hardware, with operations expanding across the Sun Belt region. Aurora's next-generation FirstLight LiDAR doubles the detection range to 1,000 meters while cutting overall hardware cost by half, with the second-generation commercial hardware kit expected in mid-2026. [10][21]
Zoox, acquired by Amazon in 2020 for approximately $1.2 billion, has developed a purpose-built robotaxi with a distinctive boxy design, bidirectional driving capability, and no traditional driver controls. The vehicle seats four passengers and uses a sensor suite with cameras, LiDAR, and radar providing 360-degree coverage.
Zoox launched a public driverless service in Las Vegas in September 2025 and San Francisco in November 2025 through its Zoox Explorers program. NHTSA granted Zoox an exemption in August 2025 to demonstrate its purpose-built vehicles on public roads. As of March 2026, Zoox had carried more than 350,000 riders and driven nearly 2 million autonomous miles, was expanding testing toward Austin and Miami, and planned to begin charging for rides in 2026 pending regulatory approvals. [22]
Mobileye, an Israeli company acquired by Intel in 2017 for $15.3 billion (and subsequently taken public again in 2022), is a leading supplier of ADAS and autonomous driving technology. Mobileye's EyeQ chips power driver assistance systems in vehicles from numerous automakers.
The latest EyeQ6 system-on-chip supports applications ranging from basic ADAS to semi-autonomous and autonomous driving. Mobileye's SuperVision system provides hands-free highway driving, and Chauffeur targets higher levels of autonomy. The company estimates future delivery of more than 19 million EyeQ6H-based systems across automakers including Volkswagen Group and a major US automaker announced in January 2026.
Mobileye has begun development of EyeQ7 and EyeQ8 chips, with EyeQ8 targeting "mind-off" driving (where no human intervention is required) by 2029-2030.
Baidu's Apollo platform is the leading autonomous driving service in China. Apollo Go, the commercial robotaxi service, has expanded to approximately 22 cities worldwide, including Beijing, Shanghai, Wuhan, Shenzhen, Hong Kong, and others. [12]
In Q3 2025, Apollo Go delivered 3.1 million fully driverless ride-hailing trips, and its global driverless count reached about 3.4 million in Q4 2025, up more than 200% year over year. [12] As of October 31, 2025, weekly order volume exceeded 250,000 (100% fully driverless), with the weekly peak surpassing 300,000 during Q4 2025, putting Baidu's volume on par with Waymo. By February 2026, Apollo Go had completed more than 20 million cumulative rides worldwide. [12] Baidu has allocated $1.2 billion over three years for fleet expansion and vehicle procurement, with deployment in 20 cities expected by Q4 2026. Baidu also expanded internationally, launching fully driverless commercial service in Dubai in partnership with Dubai Taxi Company in April 2026. [12]
Pony.ai, founded in 2016 and headquartered in Guangzhou, China, operates a fleet of over 726 robotaxis, with 202 in commercial operation. The company completed dual listings, trading on NASDAQ and the Hong Kong Stock Exchange (November 2025), raising approximately $860 million in its Hong Kong IPO. Pony.ai is expanding internationally into Luxembourg, South Korea, and Dubai.
WeRide, also based in China, operates a fleet of over 1,500 autonomous vehicles. The company went public on the Hong Kong Stock Exchange in November 2025. WeRide has partnered with Uber to launch robotaxi services in Abu Dhabi and plans trial operations in Dubai.
| Company | Parent/Backer | SAE Level | Sensor Approach | Primary Focus | Deployment Status (Early 2026) |
|---|---|---|---|---|---|
| Waymo | Alphabet | Level 4 | LiDAR + cameras + radar | Robotaxi | ~500,000 paid rides/week across ~10 US cities |
| Tesla | Independent | Level 2 (FSD Supervised) | Vision-only (cameras) | Consumer vehicles + robotaxi | 10B+ FSD miles; robotaxi pilots in Austin, Dallas, Houston |
| Cruise | General Motors | Level 4 (discontinued) | LiDAR + cameras + radar | Robotaxi (now ADAS) | Merged into GM; robotaxi operations ceased |
| Aurora | Independent (public) | Level 4 | LiDAR + cameras + radar | Autonomous trucking | 5.3M+ commercial miles; Dallas-Houston and Fort Worth-El Paso |
| Zoox | Amazon | Level 4 | LiDAR + cameras + radar | Robotaxi | Public service in Las Vegas and San Francisco; ~2M miles |
| Mobileye | Intel | Level 2-4 | Camera-first + LiDAR + radar | ADAS supplier + AV | EyeQ6 shipping to automakers; AV testing ongoing |
| Baidu Apollo | Baidu | Level 4 | LiDAR + cameras + radar | Robotaxi | 20M+ cumulative rides; 20+ cities; launched in Dubai |
| Pony.ai | Independent (public) | Level 4 | LiDAR + cameras + radar | Robotaxi + trucking | Operating in China; expanding to Middle East, Europe |
| WeRide | Independent (public) | Level 4 | LiDAR + cameras + radar | Robotaxi + logistics | Operating in China; partnered with Uber for Middle East |
In the United States, autonomous vehicle regulation involves both federal and state authorities. At the federal level, the National Highway Traffic Safety Administration (NHTSA) oversees vehicle safety standards. In April 2025, US Transportation Secretary Sean Duffy announced a new NHTSA Automated Vehicle Framework based on three principles: prioritizing safety of AV operations on public roads, removing unnecessary regulatory barriers to innovation, and enabling commercial deployment. [7]
NHTSA proposed the AV STEP (ADS-equipped Vehicle Safety, Transparency, and Evaluation Program), a voluntary national framework for evaluating vehicles with automated driving systems. The agency also launched three rulemakings to modernize Federal Motor Vehicle Safety Standards (FMVSS) for vehicles with ADS, aiming to reduce the patchwork of state regulations. [7]
In the absence of comprehensive federal legislation, US states have taken varying approaches to autonomous vehicle regulation. As of 2025-2026:
The United Nations Economic Commission for Europe (UNECE) provides the global regulatory platform for vehicle regulations through its World Forum for Harmonization of Vehicle Regulations (WP.29). In 2018, WP.29 established the Working Party on Automated/Autonomous and Connected Vehicles (GRVA) to develop international standards. [8]
Key international developments include:
The 1968 Vienna Convention on Road Traffic, which originally required a human driver to control every vehicle, was amended in 2016 to permit automated driving features. The amendment entered into force on July 14, 2022.
Germany became the first country to establish a legal framework for Level 4 autonomous driving in 2021, and Mercedes-Benz received approval for its Level 3 Drive Pilot system. China has established regulations allowing robotaxi operations in designated zones across multiple cities. Japan legalized Level 4 autonomous driving in April 2023.
Proponents of autonomous driving point to data suggesting that AV technology can be, on average, safer than human drivers, while critics caution that AVs still operate largely in favorable conditions. The most detailed evidence comes from Waymo, which reports its crash performance under NHTSA's Standing General Order and in peer-reviewed research.
Through December 2025, across 170.7 million rider-only miles, Waymo reported the following reductions in crashes compared with human-driver benchmarks for the same roads: 82% fewer injury-causing crashes (544 fewer), 92% fewer crashes resulting in serious injury or worse (35 fewer), 83% fewer crashes with airbag deployment (230 fewer), 92% fewer pedestrian-injury crashes (62 fewer), 85% fewer cyclist-injury crashes (39 fewer), and 81% fewer motorcyclist-injury crashes (25 fewer). [9] A study of 56.7 million rider-only miles, accepted for publication in the journal Traffic Injury Prevention in 2025, found statistically significant reductions in any-injury-reported, airbag-deployment, and suspected-serious-injury-or-worse crash rates, concluding that "the Waymo ADS significantly outperformed the overall human benchmarks" across multiple crash types. [16][18] A separate 2024 study with reinsurer Swiss Re found that, compared with the most advanced human-driven vehicles, the Waymo Driver was associated with an 88% reduction in property-damage claims and a 92% reduction in bodily-injury claims (the company has subsequently cited figures of roughly 86% and 90% at larger mileage). [18]
Critics note that autonomous vehicles still operate primarily in favorable conditions (good weather, well-mapped urban areas) and that comparing AV safety to overall human-driver averages can be misleading without controlling for driving conditions, road type, and severity. High-profile failures, including the 2018 Uber and 2023 Cruise incidents detailed below, also illustrate the consequences when these systems do fail.
On March 18, 2018, at approximately 9:58 PM MST, a prototype Uber self-driving car (a modified Volvo XC90) struck and killed Elaine Herzberg, 49, as she was pushing a bicycle across Mill Avenue in Tempe, Arizona, outside a designated crosswalk. This was the first known fatality involving a fully autonomous vehicle striking a pedestrian. [5]
The National Transportation Safety Board (NTSB) investigation revealed multiple failures. The vehicle's self-driving system detected Herzberg 5.6 seconds before impact but cycled between classifying her as a vehicle, a bicycle, and an unknown object. [5] The system could not classify objects as pedestrians unless they were near a crosswalk. Furthermore, Uber had disabled the vehicle's automatic emergency braking system to prevent erratic behavior during testing. [5]
The human safety backup driver, Rafaela Vasquez, was watching a television show on her smartphone at the time of the crash and did not intervene until less than a second before impact. Vasquez was charged with negligent homicide, pled guilty to a lesser charge of endangerment in July 2023, and was sentenced to three years of probation.
The NTSB cited contributing factors including Uber's inadequate safety procedures, ineffective driver oversight, the decision to disable emergency braking, Herzberg's choice to cross outside a crosswalk, and Arizona's insufficient oversight of AV testing. [5] Uber sold its ATG division to Aurora in December 2020.
On October 2, 2023, in San Francisco, a pedestrian was struck by a hit-and-run driver at Market and Fifth streets and knocked into the path of a Cruise robotaxi. The Cruise vehicle ran over the pedestrian, stopped briefly, failed to detect the person beneath the car, and then attempted to pull over to the side of the road, dragging the victim approximately 20 feet. [6]
The incident was made significantly worse by Cruise's response. The company filed a report with NHTSA that omitted any mention of the dragging. In a videoconference with NHTSA the next morning, Cruise employees described the accident without mentioning the dragging. The record was not corrected for 10 days. Cruise later admitted to filing a false report and agreed to a deferred prosecution agreement with a $500,000 criminal fine. NHTSA separately imposed a $1.5 million civil penalty. [6]
The California DMV suspended Cruise's autonomous vehicle permits. Several executives were dismissed. The incident contributed to GM's December 2024 decision to shut down Cruise as a robotaxi company and absorb the technology into GM's personal vehicle ADAS programs.
Despite high-profile incidents, proponents of autonomous driving point to data suggesting that AV technology is, on average, safer than human drivers. Waymo has published peer-reviewed research indicating that its vehicles have significantly lower crash rates than human-driven vehicles in comparable conditions, as summarized in the safety section above. [9][16] However, critics note that autonomous vehicles still operate primarily in favorable conditions (good weather, well-mapped urban areas) and that comparing AV safety to human driver averages is misleading without controlling for driving conditions.
Simulation plays a critical role in the development and validation of autonomous driving systems. Testing in simulation allows developers to safely expose their systems to rare and dangerous scenarios that would be impractical or unethical to replicate on public roads.
CARLA (Car Learning to Act) is an open-source simulator for autonomous driving research, built on Unreal Engine. [15] CARLA provides a flexible simulation environment with configurable sensor suites (LiDAR, cameras, depth sensors, GPS), environmental conditions (weather, time of day), and dynamic actors (vehicles, pedestrians). The simulator supports development, training, and validation of perception, planning, and control algorithms. CARLA has become one of the most widely used platforms in academic autonomous driving research. [15]
The Waymo Open Dataset is a large-scale collection of sensor data collected by Waymo's autonomous vehicles in diverse conditions. [16] It includes the Perception dataset (high-resolution LiDAR and camera data with labels), the Motion dataset (object trajectories and 3D maps for over 103,000 scenes), and the End-to-End Driving dataset (camera data with high-level navigation commands). [16] The dataset is widely used for benchmarking perception, prediction, and planning algorithms.
nuScenes, developed by Motional (formerly nuTonomy), is a public large-scale dataset for autonomous driving. [17] It contains 1,000 driving scenes collected in Boston and Singapore, with full 360-degree LiDAR, camera, and radar coverage. [17] The nuScenes format has become a de facto standard for multi-sensor 3D object detection and tracking benchmarks.
Additional important datasets include the KITTI dataset (from the Karlsruhe Institute of Technology), Argoverse (from Argo AI, now used by academic community), and the Lyft Level5 dataset. Commercial simulation platforms include NVIDIA DRIVE Sim, Applied Intuition, and Foretellix, which are used by AV companies to run billions of simulated miles.
Waymo alone has reported running billions of miles in simulation, complementing its real-world testing. Simulation enables scenario-based testing where developers can systematically vary conditions (pedestrian behavior, weather, road geometry) to evaluate system robustness.
As of early 2026, autonomous driving technology is in a period of rapid commercial expansion, particularly for robotaxi and autonomous trucking applications.
Waymo is the clear market leader in the United States, operating paid robotaxi service in Phoenix, San Francisco, Los Angeles, Austin, and Atlanta, with expansion to additional US cities and London planned for 2026. Waymo provides roughly 500,000 paid trips per week across about 10 US cities and targets 1 million weekly trips by the end of 2026. [9][18]
Zoox has launched public driverless service in Las Vegas and San Francisco, has carried more than 350,000 riders across nearly 2 million autonomous miles, and plans to begin charging for rides in 2026. [22]
Baidu Apollo Go leads in China with service in over 20 cities, more than 20 million cumulative rides, and weekly trip volumes that peaked above 300,000 in Q4 2025, comparable to Waymo. [12]
Tesla launched a pilot robotaxi service in Austin using its existing vehicle fleet in June 2025 and reported unsupervised rides in Dallas and Houston by April 2026, with the purpose-built Cybercab entering production in April 2026. [19]
Pony.ai and WeRide operate commercially in China and are expanding to the Middle East.
Aurora is the leader in autonomous trucking, operating commercial driverless freight service across Texas (Dallas-Houston and Fort Worth-El Paso) with more than 5.3 million cumulative commercial miles and plans to scale to hundreds of trucks by the end of 2026. [10][21]
Other companies in the autonomous trucking space include Kodiak Robotics, TuSimple (which pivoted away from US operations), and Gatik (focused on middle-mile logistics).
Several technical trends define the 2025-2026 period:
The economic implications of autonomous driving are substantial. Market forecasts vary but consistently project significant growth:
| Forecast Source | Market Segment | Projected Size by 2030 | CAGR |
|---|---|---|---|
| Grand View Research | Global autonomous vehicle market | $214 billion | 19.9% |
| Morgan Stanley | Self-driving technology market | $200 billion | N/A |
| Goldman Sachs | Robotaxi market | $25-45 billion | N/A |
| Multiple analysts | Autonomous trucking | $400+ billion | N/A |
Adoption of vehicles with partial to full automation is projected to jump from 8% in 2024 to 28% by 2030 in developed markets. [13] By 2035, some analysts forecast 76 million self-driving cars on the road globally.
The economic impact extends beyond the vehicle market itself:
Timeline predictions for widespread autonomous driving have historically been overly optimistic. Elon Musk first predicted full self-driving capability by 2018. Waymo initially expected broad commercial deployment by the early 2020s. Industry consensus as of 2026 suggests that Level 4 robotaxis will operate in dozens of cities by 2028-2030, while true Level 5 autonomy (unrestricted by geography or conditions) remains at least a decade away.