Autonomous vehicle
Last reviewed
Jun 9, 2026
Sources
17 citations
Review status
Source-backed
Revision
v4 · 5,804 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Jun 9, 2026
Sources
17 citations
Review status
Source-backed
Revision
v4 · 5,804 words
Add missing citations, update stale details, or suggest a clearer explanation.
An autonomous vehicle, also known as a self-driving car, driverless car, or robotic vehicle, is a vehicle that uses artificial intelligence, sensors, and software to navigate and operate without human input. Autonomous vehicles perceive their surroundings using a combination of cameras, lidar, radar, and ultrasonic sensors, then use AI-based planning and control systems to make driving decisions. The development of autonomous vehicles draws on advances in computer vision, deep learning, sensor fusion, and robotics, and represents one of the most complex real-world applications of artificial intelligence.
As of early 2026, fully driverless robotaxi services are commercially operating in multiple cities in the United States and China. Waymo, a subsidiary of Alphabet, provides over 250,000 paid rides per week across five U.S. cities, while Baidu's Apollo Go operates approximately 1,000 autonomous vehicles in China [1]. Tesla launched a limited robotaxi service in Austin, Texas in late January 2026 [2]. However, no vehicle system has yet achieved full autonomy in all conditions and environments (SAE Level 5).
The Society of Automotive Engineers (SAE) International established the J3016 standard, which defines six levels of driving automation ranging from no automation to full automation. This classification system, first published in 2014 and updated in subsequent revisions, has become the globally accepted framework for describing vehicle automation capabilities.
| SAE Level | Name | Description | Driver attention required | Example systems |
|---|---|---|---|---|
| Level 0 | No Driving Automation | The human driver performs all driving tasks. The vehicle may have warning systems (collision alerts, lane departure warnings) but does not take control. | Full attention at all times | Standard vehicles with basic warning systems |
| Level 1 | Driver Assistance | The vehicle can assist with either steering or acceleration/braking, but not both simultaneously. The driver remains fully responsible. | Full attention at all times | Adaptive cruise control, lane-keeping assist |
| Level 2 | Partial Driving Automation | The vehicle can control both steering and acceleration/braking simultaneously under specific conditions. The driver must remain attentive and ready to intervene at any time. | Hands may be off wheel, but eyes must remain on road | Tesla Autopilot, GM Super Cruise, Ford BlueCruise |
| Level 3 | Conditional Driving Automation | The vehicle handles all aspects of driving in specific conditions (such as highway driving or traffic jams). The driver can disengage from driving but must be ready to take over when the system requests it. | Can disengage attention in defined conditions; must respond to takeover requests | Mercedes-Benz Drive Pilot (certified in Germany, Nevada, and California) |
| Level 4 | High Driving Automation | The vehicle can operate fully autonomously within defined operational design domains (geographic areas, weather conditions, road types). No human intervention is required within those domains. If the system cannot continue, it performs a safe stop. | No attention required within operational domain | Waymo One robotaxi, Baidu Apollo Go, Zoox |
| Level 5 | Full Driving Automation | The vehicle can drive itself under all conditions that a human driver could handle, with no geographic or environmental limitations. No steering wheel or pedals are required. | No driver needed | No production vehicle exists at this level as of 2026 |
As of 2025, Level 2 systems dominate the consumer market in terms of adoption. Level 3 remains limited to a small number of certified systems in specific jurisdictions. Level 4 is operational in commercial robotaxi and autonomous trucking services within defined geographic areas [3].
The concept of self-driving vehicles predates the computer age. In 1925, the Houdina Radio Control company demonstrated a radio-controlled car called "American Wonder" that drove through the streets of New York City, though it was remotely operated rather than autonomous. In the 1950s and 1960s, General Motors and RCA showcased concept vehicles guided by embedded highway infrastructure at exhibitions like the 1939 World's Fair (GM's Futurama) and the 1960 follow-up.
The first truly autonomous vehicles emerged from academic research. In 1986, Ernst Dickmanns at the Bundeswehr University Munich equipped a Mercedes-Benz van with cameras and custom processors, creating VaMoRs (Versuchsfahrzeug fur autonome Mobilitat und Rechnersehen), which could drive on empty highways at speeds up to 96 km/h (60 mph) using computer vision. By 1994, Dickmanns' upgraded VaMP and VITA-2 vehicles completed a drive of over 1,000 km on a Paris highway in heavy traffic with minimal human intervention [4].
In the United States, Carnegie Mellon University's Navlab project produced a series of autonomous vehicles starting in 1986. In 1995, the Navlab 5 vehicle completed a cross-country trip from Pittsburgh to San Diego, covering 2,849 miles with the steering handled autonomously for 98.2% of the journey (though a human controlled speed and braking). The project was called "No Hands Across America" [4].
The DARPA Grand Challenge, organized by the U.S. Defense Advanced Research Projects Agency, was a pivotal turning point in autonomous vehicle development. DARPA's longer-term goal was to accelerate the development of autonomous vehicles that could substitute for personnel in hazardous military operations, such as supply convoys.
2004 Challenge. The first DARPA Grand Challenge took place on March 13, 2004. Fifteen autonomous vehicles attempted to navigate a 142-mile course across the Mojave Desert from Barstow, California to Primm, Nevada. No vehicle completed the course. The best performer, Carnegie Mellon University's Red Team entry called Sandstorm, traveled only 7.32 miles (11.78 km) before getting stuck. The $1 million prize went unclaimed [5].
2005 Challenge. DARPA announced a second challenge almost immediately. On October 8, 2005, five vehicles out of 195 entrants successfully completed a 132-mile course through the Nevada desert. Stanford University's entry, "Stanley," a modified Volkswagen Touareg, finished first with a time of 6 hours and 53 minutes, winning the $2 million prize. Stanley was led by Sebastian Thrun, a Stanford professor who would later become a central figure in the development of commercial self-driving cars [5].
2007 Urban Challenge. The third iteration moved from the desert to an urban environment, requiring vehicles to navigate a 60-mile course through a simulated city with traffic, intersections, and parking. Carnegie Mellon's "Boss" vehicle won, completing the course in approximately 4 hours. This challenge was significantly more complex, requiring vehicles to obey traffic laws, merge with moving traffic, and navigate intersections [5].
The DARPA challenges were transformative for the field. They demonstrated that autonomous driving on public roads was technically feasible, attracted major media attention, and created a community of researchers and engineers who would go on to found or lead commercial autonomous vehicle companies.
In 2009, Google launched its self-driving car project, led by Sebastian Thrun, the Stanford professor who had led the Stanley team to victory in the 2005 DARPA Grand Challenge. Several other DARPA Challenge veterans joined the project, including Chris Urmson from Carnegie Mellon's Boss team.
Google's project began by retrofitting Toyota Prius vehicles with sensors and computing hardware. By 2012, the fleet had logged over 300,000 miles of autonomous driving on public roads in California and Nevada. In 2014, Google unveiled a custom-built prototype with no steering wheel or pedals, designed from the ground up for autonomous operation.
In December 2016, Google spun off the project as Waymo, a subsidiary of Alphabet Inc. Waymo launched its first commercial robotaxi service, Waymo One, in Phoenix, Arizona in December 2018, initially with a safety driver present. By 2020, Waymo began offering fully driverless rides (with no safety driver) to members of the public in a limited area of suburban Phoenix [6].
As of late 2025, Waymo operates fully driverless robotaxi services in Phoenix, Los Angeles, San Francisco, Austin, and Atlanta, providing over 250,000 paid rides per week. The company operates a fleet of over 3,000 vehicles and has driven more than 127 million rider-only miles (miles without a human safety driver). In November 2025, Waymo began offering freeway routes in the San Francisco, Phoenix, and Los Angeles markets. For 2026, Waymo has announced plans to expand to Dallas, Denver, Detroit, Houston, Las Vegas, Miami, Nashville, Orlando, San Antonio, San Diego, and Washington, D.C. The company also announced plans to launch in London, which would be its first international market [1].
Cruise was founded in 2013 by Kyle Vogt and Dan Kan and was acquired by General Motors in 2016 for approximately $1 billion. GM subsequently invested billions of dollars into Cruise, eventually spending more than $10 billion on the robotaxi venture.
Cruise began offering fully driverless rides to the public in San Francisco in 2023. However, in October 2023, a serious incident occurred: a pedestrian who had been struck by a hit-and-run driver was thrown into the path of a Cruise vehicle, which then dragged the person approximately 20 feet before stopping. The incident led to the suspension of Cruise's California operating permit, and the National Highway Traffic Safety Administration (NHTSA) fined Cruise $1.5 million after the company failed to fully disclose details of the crash [7].
In December 2024, General Motors announced it would stop funding Cruise's robotaxi development, effectively shutting down the program. GM dismissed approximately 1,000 employees, halving Cruise's workforce. The company stated it would "realign its autonomous driving strategy" to focus on advanced driver-assistance systems (ADAS) for personal vehicles rather than robotaxis [7].
The autonomous vehicle industry includes technology companies, traditional automakers, and startups. The following table summarizes the major players and their current status as of early 2026.
| Company | Parent/Affiliation | Approach | Current status (2026) |
|---|---|---|---|
| Waymo | Alphabet (Google) | Lidar + cameras + radar; Level 4 robotaxi | Operating driverless service in 5 U.S. cities; 250,000+ rides/week; expanding to 10+ additional cities and London in 2026 [1] |
| Tesla FSD | Tesla | Vision-only (cameras); aiming for Level 4/5 | FSD (Supervised) available to consumers; 8.3 billion miles driven with FSD; limited robotaxi service launched in Austin January 2026 [2] |
| Cruise | General Motors | Lidar + cameras + radar; Level 4 | Robotaxi program shut down December 2024; GM pivoted to ADAS for personal vehicles [7] |
| Zoox | Amazon | Custom vehicle with no steering wheel; lidar + cameras + radar | Testing bidirectional robotaxi in several U.S. locations; preparing for commercial launch [8] |
| Baidu Apollo | Baidu | Lidar + cameras + radar; Level 4 | Apollo Go operating approximately 1,000 vehicles in multiple Chinese cities; achieved city-wide unit economics breakeven in select markets [9] |
| Motional | Hyundai (majority) / Aptiv | Lidar + cameras + radar; Level 4 | Restructured ownership; Hyundai committed approximately $1 billion; shifted focus from immediate robotaxi deployment to longer-term Level 4 development [10] |
| Aurora Innovation | Independent (public company) | Lidar + cameras + radar; autonomous trucking | Launched driverless trucking on Dallas-Houston and Fort Worth-El Paso routes; surpassed 100,000 driverless miles; expanding with next-gen hardware in 2026 [11] |
| Pony.ai | Independent (public company) | Lidar + cameras + radar; Level 4 | Included in MSCI China Index (February 2026); dual-listed in Hong Kong; achieved unit economics breakeven in Guangzhou with Gen-7 platform [12] |
An autonomous vehicle's software stack can be broadly divided into four major subsystems: perception, localization, planning, and control. These subsystems work together in a continuous loop, processing sensor data, understanding the vehicle's environment, deciding on actions, and executing those actions.
Perception is the process by which the vehicle senses and interprets its surroundings. Autonomous vehicles use multiple sensor modalities, each with different strengths and weaknesses.
| Sensor | How it works | Strengths | Weaknesses |
|---|---|---|---|
| Cameras | Capture 2D images using visible light; typically 6 to 12 cameras provide 360-degree coverage | Rich color and texture information; can read signs, traffic lights, lane markings; inexpensive | Affected by lighting conditions (glare, darkness); limited depth perception without stereo or ML |
| Lidar | Emits laser pulses and measures the time for reflections to return; generates detailed 3D point clouds of the environment | Precise 3D distance measurements; works in darkness; less affected by lighting variation | Expensive (though costs have fallen significantly); degraded by heavy rain, snow, or fog; does not capture color/texture |
| Radar | Emits radio waves and measures reflections; operates at millimeter-wave frequencies for automotive applications | Works in all weather conditions (rain, snow, fog); excellent at measuring velocity via Doppler effect; long range | Lower spatial resolution than lidar or cameras; cannot read signs or detect color |
| Ultrasonic sensors | Emit sound waves and measure reflections at short range | Very inexpensive; reliable for close-range detection | Very short range (typically under 5 meters); slow update rate |
Sensor fusion combines data from multiple sensor types to create a unified model of the environment that is more robust and accurate than any single sensor could provide. For example, cameras provide rich visual context while lidar provides precise distance information; combining both yields a detailed, 3D-aware scene understanding.
Perception algorithms process raw sensor data to detect and classify objects (vehicles, pedestrians, cyclists, traffic signs, lane markings), estimate their positions and velocities, and predict their future trajectories. Modern perception systems rely heavily on convolutional neural networks (CNNs) for object detection and semantic segmentation, with architectures like YOLO, SSD, and PointPillars processing camera images and lidar point clouds.
Localization determines the vehicle's precise position within its environment, typically to centimeter-level accuracy. Autonomous vehicles combine GPS data with high-definition maps and real-time sensor matching. HD maps contain detailed 3D representations of road geometry, lane markings, traffic signs, and other infrastructure. The vehicle compares its live sensor data against the HD map to precisely determine its location, a process known as map matching or localization.
Simultaneous Localization and Mapping (SLAM) techniques allow vehicles to build or update maps while simultaneously tracking their own position. This is particularly important in areas where HD maps may be outdated or unavailable.
The planning subsystem decides what the vehicle should do, translating the perception system's understanding of the environment into a series of actions. Planning typically operates at multiple levels:
Planning algorithms must handle an enormous range of scenarios, from routine highway driving to complex urban intersections with pedestrians, cyclists, construction zones, and unpredictable behavior by other drivers. This combinatorial complexity is one of the core challenges of autonomous driving.
The control subsystem translates the planned trajectory into physical actuator commands: steering angle, throttle position, and brake pressure. Control algorithms, often based on model predictive control (MPC) or PID controllers, ensure the vehicle accurately follows the planned path while maintaining stability and passenger comfort.
Autonomous vehicles are among the most demanding applications of artificial intelligence, requiring real-time processing of high-bandwidth sensor data, robust decision-making under uncertainty, and safe operation in an open-ended environment.
Convolutional neural networks (CNNs) form the backbone of most perception systems in autonomous vehicles. CNNs excel at processing grid-like data such as images and are used for object detection (identifying and localizing vehicles, pedestrians, cyclists, and other objects), semantic segmentation (classifying each pixel in an image by category), and lane detection.
Common CNN architectures used in autonomous driving include ResNet, EfficientNet, and specialized real-time architectures like YOLO (You Only Look Once) and SSD (Single Shot MultiBox Detector). For lidar data, architectures like PointNet, PointPillars, and VoxelNet process 3D point clouds directly [13].
Transformer architectures, originally developed for natural language processing, have increasingly been applied to autonomous driving. Vision transformers (ViTs) process image data using self-attention mechanisms rather than convolutions, enabling them to capture long-range dependencies in visual scenes.
TransFuser, a notable research architecture, uses transformer modules at multiple resolutions to fuse perspective-view camera features with bird's-eye-view lidar features, enabling end-to-end driving through imitation learning. The attention mechanism in transformers has proven effective at aggregating context across different sensor inputs, improving both perception accuracy and planning quality [14].
BEVFormer and similar architectures use transformers to construct bird's-eye-view representations from multi-camera inputs, providing a unified spatial representation that simplifies downstream planning tasks.
Imitation learning, particularly its sub-category behavior cloning, trains driving policies by having the model learn to mimic the behavior of expert human drivers. The model is given sensor inputs (camera images, lidar data) paired with the corresponding driving actions (steering angle, acceleration, braking) taken by a skilled human driver, and learns to map inputs to actions through supervised learning.
The advantage of imitation learning is its simplicity and the ability to leverage large datasets of human driving. Tesla's FSD system, for example, is trained in part by learning from billions of miles of driving data collected from Tesla vehicles on public roads. The weakness is that imitation learning can struggle with situations that are rare in the training data, and small errors can compound over time (a problem known as distributional shift or covariate shift) [14].
Reinforcement learning (RL) trains driving agents through trial and error in simulated environments. An RL agent receives a reward signal based on driving quality (smooth driving, maintaining safe distances, reaching the destination) and learns a policy that maximizes cumulative reward. RL is particularly useful for learning complex behaviors like negotiating unprotected left turns, merging onto highways, and interacting with aggressive drivers.
In practice, RL for autonomous driving is typically conducted in simulation (using environments like CARLA, NVIDIA DRIVE Sim, or Waymo's internal simulator) because real-world training would be unsafe and prohibitively expensive. The challenge lies in transferring policies learned in simulation to the real world, a problem known as the sim-to-real gap [13].
A growing trend in autonomous driving research is end-to-end learning, where a single neural network maps directly from raw sensor inputs to driving actions, without explicitly separate perception, planning, and control modules. End-to-end systems can potentially learn more efficient representations and avoid information loss that occurs when separate modules communicate through predefined interfaces.
Tesla's FSD system has moved increasingly toward an end-to-end architecture. In early 2024, Tesla deployed FSD v12, which replaced much of the previous rule-based planning code with a neural network trained on millions of video clips of human driving. This shift represented a significant architectural change, relying on the neural network to learn driving behavior implicitly rather than through hand-coded rules [2].
However, end-to-end systems face challenges in interpretability (it is difficult to understand why the system made a particular decision), verification (proving the system is safe is harder when there are no modular components to test individually), and robustness (a single model failure can affect the entire driving stack) [14].
One of the most debated topics in autonomous driving is whether lidar is necessary for safe autonomous driving or whether camera-based (vision-only) systems can achieve equivalent or superior performance.
Tesla, under the leadership of CEO Elon Musk, has been the most prominent advocate of a vision-only approach. Since 2021, Tesla has progressively removed radar and ultrasonic sensors from its vehicles, relying entirely on eight external cameras and a neural network-based perception system called Tesla Vision. Musk has called lidar "a fool's errand" and argued that since humans drive using only vision, a sufficiently advanced vision system should be capable of full autonomy [15].
Tesla's arguments for the vision-only approach include cost (cameras are far cheaper than lidar, enabling Tesla to sell vehicles at lower price points while still offering autonomy features), scalability (Tesla's fleet of millions of vehicles generates training data at a scale that lidar-equipped fleets cannot match), and the thesis that vision provides fundamentally richer information than lidar (including color, texture, and the ability to read signs and traffic lights) [15].
In August 2025, Musk stated that "lidar and radar reduce safety due to sensor contention," arguing that when lidar/radar disagree with cameras, the resulting ambiguity increases rather than decreases risk [15].
As of February 2026, Tesla reported that vehicles had driven 8.3 billion miles with FSD (Supervised). In late January 2026, Tesla launched a limited robotaxi service in Austin without a Tesla employee in the vehicle. However, Musk's repeated timelines for achieving full autonomy have been missed; he did not fulfill his stated goal of removing safety drivers from Austin robotaxis before the end of 2025 [2].
Waymo, Baidu, Aurora, Zoox, and virtually every other autonomous vehicle company besides Tesla use lidar as a core sensor alongside cameras and radar. Proponents argue that lidar provides critical capabilities that cameras alone cannot reliably deliver.
Lidar produces precise, direct 3D distance measurements that do not depend on lighting conditions. It works equally well in darkness, direct sunlight, and varying weather (though heavy rain and snow degrade performance). Cameras, by contrast, must infer depth from 2D images using neural networks, a process that can fail in low-light conditions, against bright glare, or when encountering objects not well-represented in training data.
In a widely discussed comparative test in early 2025, tech YouTuber Mark Rober tested a lidar-equipped Lexus SUV against a Tesla running FSD across six challenging scenarios. The lidar vehicle passed all six tests; the Tesla passed three [15].
Multi-sensor fusion, the combination of cameras, lidar, and radar, provides redundancy and complementary capabilities. If one sensor modality fails or is degraded (camera blinded by sun, lidar obscured by rain), the others can compensate. Waymo's system, for example, uses cameras for rich visual context, lidar for precise 3D geometry, and radar for velocity measurement and all-weather detection, fusing these inputs to create a robust environmental model [6].
Critics of Tesla's approach cite edge-case vulnerabilities in vision-only systems, including poor low-light performance, difficulty detecting stationary objects that the neural network has not been trained on, and challenges with unusual road configurations [15].
Safety is the central concern for autonomous vehicles. Companies, regulators, and researchers use various metrics to compare autonomous vehicle safety to human driving performance.
Waymo has published the most comprehensive safety data of any autonomous vehicle company. A peer-reviewed study published in 2025, examining 56.7 million rider-only miles through January 2025, found statistically significant safety advantages compared to human drivers on the same roads [16]:
| Metric | Reduction compared to human drivers |
|---|---|
| Crashes causing serious injury or worse | 91% fewer |
| Crashes causing any injury | 80% fewer |
| Injury-involving intersection crashes | 96% fewer |
| Crashes involving pedestrian injuries | 92% fewer |
| Crashes involving cyclist injuries | 82% fewer |
| Crashes involving motorcyclist injuries | 82% fewer |
Through September 2025, Waymo had driven 127 million rider-only miles without a human safety driver. Between July 2021 and November 2025, there were 1,429 reported incidents involving Waymo vehicles, though the majority were caused by other drivers (human-driven vehicles colliding with Waymo vehicles) rather than by the Waymo system itself [16].
Comparing autonomous vehicle safety to human driving is methodologically challenging. Human-caused crashes in the United States result in approximately 40,000 deaths annually, with human error cited as a contributing factor in approximately 94% of serious crashes according to NHTSA. Proponents of autonomous vehicles argue that even imperfect autonomous systems could save tens of thousands of lives if they are meaningfully safer than average human drivers.
However, critics note that autonomous vehicles currently operate in a narrow set of conditions (good weather, well-mapped urban areas, moderate traffic) and that their safety record in these favorable conditions should not be directly compared to the full range of conditions human drivers face. As autonomous vehicles expand to more challenging environments, their safety performance will face sterner tests.
Regulation of autonomous vehicles in the United States is split between federal and state authorities. NHTSA, the federal agency responsible for vehicle safety, sets Federal Motor Vehicle Safety Standards (FMVSS), while states regulate licensing, registration, insurance, and traffic law.
AV STEP Program. In January 2025, NHTSA proposed the ADS-equipped Vehicle Safety, Transparency, and Evaluation Program (AV STEP), a voluntary review and reporting framework for vehicles equipped with automated driving systems. AV STEP is open to vehicle manufacturers, ADS developers, fleet operators, and system integrators [17].
AV Framework. In April 2025, NHTSA announced its new AV Framework, with stated principles of prioritizing safety, enabling innovation, and facilitating commercial deployment. The agency streamlined crash reporting requirements for ADAS and ADS and included domestically produced vehicles in the Automated Vehicles Exception Program. In September 2025, NHTSA announced three rulemakings to update FMVSS for ADS-equipped vehicles that lack conventional controls (steering wheels, pedals) [17].
State-level regulation. As of 2025, over 30 states have enacted legislation or executive orders related to autonomous vehicles. California, Arizona, Texas, and Nevada have been the most permissive in allowing testing and commercial deployment. California's Department of Motor Vehicles operates a rigorous permitting program that separately licenses testing (with and without safety drivers) and commercial deployment of autonomous vehicles.
The United Nations Economic Commission for Europe (UNECE) Working Party on Automated/Autonomous and Connected Vehicles (GRVA) approved a Global Technical Regulation on Automated Driving Systems in early 2025, after approximately 10 years of development. This regulation provides a framework for validating autonomous vehicles using a "safety case" approach [17].
The EU has also updated its General Safety Regulation to accommodate higher levels of automation, though fully driverless passenger vehicles are not yet widely permitted on European roads. Germany was the first EU member state to create a legal framework for Level 4 autonomous vehicles with its Autonomous Driving Act of 2021, allowing driverless vehicles in defined operational areas.
China has pursued an aggressive regulatory approach to enable autonomous vehicle deployment. Several cities, including Beijing, Shanghai, Guangzhou, and Shenzhen, have issued permits for driverless testing and commercial robotaxi services. Shenzhen enacted China's first municipal regulation specifically governing intelligent connected vehicles in 2022, establishing a legal framework for liability when autonomous vehicles are involved in accidents [9].
The most fundamental challenge for autonomous vehicles is handling the enormous variety of unusual situations that can occur on public roads. These "edge cases" or "corner cases" include unusual objects in the road (debris, animals, overturned vehicles), ambiguous traffic situations (construction zones with conflicting signals, police officers directing traffic), extreme weather conditions, and the unpredictable behavior of other road users.
While autonomous vehicles perform well in common driving scenarios, it is the rare and unexpected situations that prove most dangerous. The challenge is sometimes described as the "long tail" problem: there are a virtually infinite number of unusual scenarios, each individually rare but collectively frequent. No amount of testing can cover every possible edge case, making formal verification and robust generalization critical unsolved problems [13].
Current autonomous vehicle systems perform best in clear weather conditions. Heavy rain, snow, fog, and dust can degrade the performance of cameras (reduced visibility), lidar (light scattering), and even radar (ground clutter). Most commercial autonomous vehicle services, including Waymo's, restrict or limit operations during severe weather events.
Level 4 autonomous vehicles typically rely on high-definition maps that must be created and maintained for their operational domains. These maps require regular updates to reflect road construction, new signage, and other changes. This dependence on HD maps is one reason why Level 4 systems operate in geofenced areas rather than driving anywhere; creating and maintaining maps for the entire road network would be an enormous ongoing effort.
Autonomous vehicles are complex networked computer systems that present potential targets for cyberattacks. Researchers have demonstrated various attack vectors, including spoofing sensor inputs (projecting fake objects using laser emitters to confuse lidar, or placing adversarial patches on road signs to fool camera-based perception), compromising vehicle-to-infrastructure communication, and exploiting software vulnerabilities. Ensuring cybersecurity across the vehicle's entire software stack, sensor suite, and communication channels is an ongoing challenge [13].
When an autonomous vehicle causes an accident, questions of legal liability become complex. Is the manufacturer liable? The software developer? The fleet operator? The vehicle owner? Different jurisdictions are developing different approaches. Some, like Germany, have established liability frameworks that place primary responsibility on the vehicle operator or manufacturer when the autonomous system is engaged. Others, like many U.S. states, are still developing their legal frameworks.
The insurance industry is adapting to autonomous vehicles, with some insurers developing products specifically for autonomous fleets. Waymo, for example, carries significant insurance coverage for its robotaxi operations.
Public attitudes toward autonomous vehicles remain mixed. Surveys consistently show that a significant portion of the population is uncomfortable with the idea of riding in a driverless vehicle. High-profile incidents, such as the 2018 fatal crash involving an Uber autonomous test vehicle in Tempe, Arizona, and the Cruise incident in San Francisco in 2023, have affected public confidence. Building public trust requires not only demonstrating strong safety records but also transparent communication about capabilities and limitations.
The autonomous vehicle industry in 2025 and 2026 is characterized by a maturing competitive landscape with clear leaders, some high-profile exits, and significant expansion of commercial services.
Waymo is the clear market leader. With over 3,000 vehicles, 250,000+ weekly rides, and operations in five U.S. cities (and growing), Waymo has established the most extensive and commercially advanced autonomous ride-hailing service in the world. The company's planned expansion to more than 10 additional U.S. cities and its first international market (London) in 2026 signals a transition from proving the technology to scaling it commercially [1].
Tesla is pursuing a fundamentally different path. Rather than deploying a small fleet of purpose-built robotaxis in mapped areas, Tesla is attempting to achieve broad autonomy across its millions of consumer vehicles using a vision-only system and end-to-end neural networks. While Tesla's data advantage (8.3 billion FSD miles as of February 2026) is enormous, the company's timeline for achieving fully unsupervised autonomy has repeatedly slipped [2].
Autonomous trucking is gaining traction. Aurora Innovation launched driverless commercial trucking routes in Texas in 2025, marking a significant milestone for the logistics industry. Autonomous trucks operate in a more constrained environment than urban robotaxis (primarily highway driving with planned routes), making them a potentially earlier path to commercial viability [11].
China is a major proving ground. Baidu's Apollo Go and Pony.ai are operating large-scale autonomous ride-hailing services in multiple Chinese cities, with supportive regulatory frameworks enabling rapid deployment. Pony.ai's inclusion in the MSCI China Index reflects growing investor confidence in the Chinese autonomous vehicle sector [9] [12].
The Cruise shutdown was a significant setback. GM's decision to abandon its $10 billion+ investment in Cruise demonstrated that the path to autonomous vehicle commercialization is not guaranteed, even for well-funded incumbents. The incident highlighted the importance of safety culture and transparent incident reporting in maintaining regulatory and public trust [7].
Safety data is becoming more robust. Waymo's peer-reviewed safety studies, showing reductions of 80% to 96% in injury-causing crashes compared to human drivers, provide the strongest evidence to date that autonomous vehicles can be significantly safer than human driving. As more companies publish safety data and more miles accumulate, the empirical case for autonomous vehicle safety will continue to strengthen [16].
The technology for autonomous driving in defined conditions has been proven. The remaining challenges are scaling to broader geographies and conditions, reducing costs to achieve profitability, navigating an evolving regulatory landscape, and building sufficient public trust to enable widespread adoption.