Autonomous vehicles are no longer a concept confined to research labs or science fiction. Robotaxis carry paying passengers through city streets, self-driving trucks haul freight across highways, delivery robots navigate sidewalks to drop off groceries, and autonomous tractors work farm fields with no one in the cab. The technology varies widely in maturity depending on the application, but concrete, operating examples already span tens of millions of miles and multiple industries.
Robotaxis Carrying Passengers in Cities
The most visible example of autonomous vehicles today is the robotaxi. Waymo, a subsidiary of Alphabet, operates a commercial driverless ride-hailing service in several U.S. cities, including San Francisco, Phoenix, Los Angeles, and Austin. An analysis of data reported to the California Public Utilities Commission found that Waymo completed more than 14 million robotaxi trips totaling over 86 million vehicle miles between August 2023 and December 2025.1Findings. Millions of Trips, “Waymo” Empty Miles: California’s First Thousand Days of Commercial Robotaxi Service Those numbers make it, by a wide margin, the largest driverless passenger service in the country.
Cruise, backed by General Motors, also ran a robotaxi program in San Francisco before suspending operations in late 2023 following a pedestrian-dragging incident and a regulatory review. Before the suspension, both Waymo and Cruise were accumulating crash and performance data that researchers have used to compare driverless systems against human-driven rideshare vehicles in the same city.2arXiv. Initial Indications of Safety of Driverless Automated Driving Systems Other companies, including Zoox (owned by Amazon), are testing robotaxi prototypes in smaller pilot areas, though none yet match Waymo’s commercial scale.
Autonomous Trucking on Highways
Long-haul trucking is widely seen as the next big commercial frontier for self-driving technology. Companies like Aurora, Kodiak Robotics, and Torc (a Daimler subsidiary) have been running supervised and semi-supervised autonomous trucks on highway corridors, primarily in the southern United States where weather conditions are more predictable. The operating model most of these companies are pursuing is a “transfer hub” approach: autonomous trucks handle the long, relatively simple highway segment of a route, while human drivers take over for the more complex first-mile and last-mile legs through city streets.
A study using U.S. Commodity Flow Survey data estimated the potential scale of this shift. If autonomous trucking technology matures to handle all weather conditions, up to 94% of long-haul tractor-trailer operator-hours could be affected. If the technology remains restricted to the warmer southern states where most testing is currently happening, only about 10% of those hours would be impacted.3Humanities and Social Sciences Communications. Impact of automation on long haul trucking operator-hours in the United States That gap underscores how much the real-world viability of autonomous trucks depends on solving the weather problem, a point that comes up repeatedly across applications.
Several autonomous truck developers have begun or are planning revenue-generating freight runs. Aurora, for instance, has announced plans for a commercial launch of driverless trucks on a Texas highway corridor. These operations remain geographically limited, but the economic pressure is intense: a truck that can drive continuously without rest breaks or hours-of-service regulations could dramatically change shipping economics.
Sidewalk Delivery Robots
At the other end of the size spectrum, small autonomous delivery robots roll along sidewalks in dozens of U.S. cities and university campuses. These are pedestrian-sized vehicles, typically about the size of a large cooler, that carry items like food, groceries, and packages directly to customers without a human delivery person involved.4Transportation Research Record: Journal of the Transportation Research Board. Study of Sidewalk Autonomous Delivery Robots and Their Potential Impacts on Freight Efficiency and Travel Companies operating in this space include Starship Technologies, which has completed millions of deliveries on college campuses, and Serve Robotics, which delivers for Uber Eats in Los Angeles.
The appeal of these robots is straightforward: last-mile delivery is the most expensive leg of the shipping chain, and a small electric robot is far cheaper per delivery than a human driver in a car or van. The technology is relatively simpler than a highway-speed robotaxi because the robots move slowly, typically under 6 mph, and operate in pedestrian environments where the consequences of a mistake are less catastrophic. Cities have had to write new rules for these devices, and the regulatory patchwork is still uneven. Some municipalities have embraced them, while others have restricted or banned sidewalk robots over concerns about pedestrian accessibility and sidewalk crowding.
Agriculture and Mining
Farming is one area where autonomous vehicles have been quietly advancing for years, partly because a tractor in the middle of a private field does not need to share space with unpredictable human drivers. Companies like John Deere, CNH Industrial, and several startups sell or are developing autonomous tractors and sprayers that use GPS-guided navigation and sensor arrays to plow, plant, spray, and harvest with minimal human oversight.
The economics, though, are more complicated than they first appear. Research from the United Kingdom found that requiring full on-site human supervision of autonomous farm equipment nearly wipes out the economic benefit for small and medium farms, while amplifying the advantage that large farms already enjoy from economies of scale.5Applied Economic Perspectives and Policy. Lessons to be learned in adoption of autonomous equipment for field crops In other words, the regulatory framework matters almost as much as the technology itself. If regulators insist on a human babysitter for every robot tractor, the savings disappear for the farmers who need them most.
Mining is another sector where autonomous vehicles operate at large scale. Companies like Caterpillar and Komatsu have deployed autonomous haul trucks in open-pit mines in Australia, Canada, and Chile. These trucks follow fixed routes in controlled environments, hauling ore around the clock. The mining use case shares the same advantage as agriculture: the vehicles operate on private land, at relatively low speeds, without public traffic.
How Safe Are They So Far
Safety data is the question everyone asks about autonomous vehicles, and the best evidence so far comes from robotaxis, where the most miles have been logged. A peer-reviewed study comparing Waymo’s rider-only (fully driverless) crash data against a human benchmark found substantial differences over 7.1 million miles of driving. The rate of crashes involving any reported injury was about 0.6 per million miles for the Waymo system versus 2.8 per million miles for human drivers, a roughly 80% reduction. For police-reported crashes more broadly, the Waymo rate was 2.1 per million miles compared with 4.68 for human drivers, a 55% reduction.6PubMed. Comparison of Waymo rider-only crash data to human benchmarks at 7.1 million miles
Those numbers look impressive, but researchers caution that comparing autonomous vehicles to human drivers is harder than it sounds. A separate study focused on the methodology of these comparisons and found that failing to account for geographic region, road type, and vehicle type can substantially bias the results. Human crash rates pulled from police reports are also known to be underreported, because many minor crashes are never filed. Without correcting for that underreporting, autonomous systems look safer than they might actually be relative to a true apples-to-apples benchmark.7PubMed. Benchmarks for retrospective automated driving system crash rate analysis using police-reported crash data
The honest summary is that the early data points in a favorable direction for driverless systems, especially for injury crashes, but the sample sizes and operating environments are still narrow enough that sweeping declarations of superiority are premature. Waymo’s vehicles operate mostly in dry, well-mapped cities with good road infrastructure. Comparing that to the full U.S. driving population, which includes rural roads, snowstorms, and impaired drivers, is inherently lopsided.
Weather and Perception Limits
All autonomous vehicles rely on sensors to perceive the world: cameras for color and detail, lidar for precise 3D distance measurements, and radar for detecting objects through some obstructions. Each sensor type has weaknesses, and bad weather exposes them all. Light rain barely affects lidar, but heavy rain with uneven precipitation rates can create clumps of water that register as fake obstacles. Snow is even worse, because solid snowflakes can form into larger masses that either trigger false detections or block the sensor’s line of sight entirely.8ISPRS Journal of Photogrammetry and Remote Sensing. Perception and sensing for autonomous vehicles under adverse weather conditions: A survey
Cameras suffer in fog, low light, and glare. Radar handles rain and fog better than the other sensors but offers lower spatial resolution, making it harder to distinguish between, say, a pedestrian and a mailbox. The upshot is that no single sensor works well in all conditions, which is why autonomous vehicle developers fuse data from multiple sensor types. A review of perception methods noted that weather limitations and equipment constraints are a primary reason that autonomous driving remains confined to specific, favorable scenarios for now.9PubMed Central. Perception Methods for Adverse Weather Based on Vehicle Infrastructure Cooperation System: A Review
Sensor fusion strategies vary in complexity. Some systems combine raw data from cameras and lidar at the data level, but this creates enormous volumes of information and heavy computing loads. Others run each sensor through a separate detection pipeline and then merge the final decisions, though this can produce conflicting results. The approach gaining the most traction fuses data at an intermediate feature level, where the raw information has already been compressed into useful patterns but before a final decision is made, which strikes a balance between computational efficiency and accuracy.10Scientific Reports. Real time object detection using LiDAR and camera fusion for autonomous driving Even with these techniques, no system has fully cracked the problem of reliable all-weather autonomy, which is why you will not find robotaxis operating in Minneapolis in January.
Vehicle-to-Everything Communication
Autonomous vehicles do not have to rely solely on their own onboard sensors. Vehicle-to-everything communication, usually shortened to V2X, allows cars to exchange data with each other, with traffic infrastructure like signal lights, and with pedestrians’ smartphones. The idea is to give a vehicle information it cannot physically see: a car approaching a blind intersection could receive a warning from another vehicle already in the cross street.
Simulations of V2X-equipped vehicles at intersections have shown meaningful safety gains. In one modeling study, equipping just the turning vehicle with intersection automated emergency braking reduced crashes by about 59%. When both vehicles at the intersection were equipped, crash avoidance jumped to 77%.11ScienceDirect. Vehicle-to-everything (V2X) in the autonomous vehicles domain – A technical review of communication, sensor, and AI technologies for road user safety The catch is that V2X only works at scale when a critical mass of vehicles and infrastructure are equipped, and deployment remains patchy. The United States has gone back and forth on mandating V2X technology, and automakers have been slow to include the necessary hardware in production vehicles.
Who Pays When a Driverless Car Crashes
The existing legal framework for car accidents in the U.S. was built around the assumption that a human driver is at fault. Most states allow injured parties to bring a negligence claim against the driver, and insurance operates on the premise that human error causes the crash. In practice, driver error accounts for roughly 90% of all accidents, with vehicle defects responsible for as little as 2%.12The American Journal of Comparative Law. Autonomous Vehicles and Liability Law
Autonomous vehicles flip that ratio on its head. If there is no human driver, traditional negligence claims against a driver make no sense. The liability question shifts toward the manufacturer, the software developer, or the company operating the fleet. Some legal scholars have argued for strict product liability, treating a crash caused by software as fundamentally the same as one caused by a defective brake. Others propose new insurance models where the autonomous vehicle operator carries the liability rather than any individual. A handful of U.S. states have started writing legislation that addresses this, but there is no national consensus. In the meantime, companies like Waymo self-insure their fleets, absorbing liability for any crashes their vehicles cause.
The twelve U.S. states with no-fault insurance systems present an additional wrinkle. In those states, crash victims are largely limited to claiming against their own insurance, which sidesteps the question of driver fault. But autonomous vehicles could still face product-liability suits when injuries exceed the no-fault threshold, and how those suits play out against a software system rather than a human is untested legal territory in most jurisdictions.
Cybersecurity Risks
A vehicle controlled entirely by software is, by definition, a target for hacking. The attack surface of an autonomous vehicle includes its electronic control units, its GPS receiver, and every one of its sensors. Researchers have demonstrated that GPS signals can be spoofed to feed a vehicle false location data, and that sensor values can be altered through physical or electronic interference.13ScienceDirect. A review of cyber attacks on sensors and perception systems in autonomous vehicle A lidar sensor, for example, can be tricked by shining a laser at it to create phantom objects, potentially causing the vehicle to brake hard or swerve for no reason.
These attacks have so far been demonstrated mostly in controlled research settings, not in real-world exploits against commercial fleets. But the stakes grow with every mile of autonomous driving added to public roads. V2X communication, while offering safety benefits, also creates new entry points for cyberattack if the communication channels are not properly secured. The industry is investing heavily in encryption, anomaly detection, and redundant sensor checks designed to flag data that does not match across multiple systems. Whether these defenses can keep pace with increasingly sophisticated attack methods remains an open and active question.
Traffic Congestion and Empty Miles
One concern that has emerged as robotaxis scale up is whether driverless vehicles actually make traffic better or worse. Unlike a personal car, a robotaxi does not park at the destination. After dropping off a passenger, it either picks up another rider nearby or drives empty to reposition itself. These “deadhead” or empty miles add vehicle traffic to roads without moving anyone. The study of Waymo’s California operations specifically analyzed the composition of vehicle miles traveled, including how many of those 86 million-plus miles were driven without a passenger in the vehicle.14Findings. Millions of Trips, “Waymo” Empty Miles: California’s First Thousand Days of Commercial Robotaxi Service
A broader economic modeling study looked at what happens to travel patterns when people can be productive during their commute, as they could in a fully autonomous car. The model found that households across all regions increased the number of trips they took when travel time became less burdensome. Overall welfare for the city rose, but some households in outlying regions actually experienced a welfare loss because increased travel by everyone else created worse congestion on the roads they depended on.15Multimodal Transportation. Assessing the economic impacts of labour time in autonomous vehicles The implication is counterintuitive: making driving more pleasant could generate so much additional traffic that the net effect is negative for people in certain parts of a metro area.
This dynamic is sometimes called “induced demand,” and it is a familiar problem in transportation planning. Widening a highway encourages more people to drive, eventually filling the new capacity. Autonomous vehicles could trigger a similar cycle, where the comfort and convenience of hands-free travel pull people out of transit and into cars, or encourage them to take trips they otherwise would have skipped. Cities that welcome robotaxis without planning for the traffic consequences could find that the congestion benefits promised by optimized autonomous driving are eaten up by sheer volume.
Where the Technology Stands Across Applications
The range of autonomous vehicle examples in operation today reflects a consistent pattern: the technology works best in controlled, predictable environments and struggles as complexity and unpredictability increase. Mining trucks running fixed routes on private land have been commercially viable for over a decade. Farm equipment operating in open fields with GPS guidance is increasingly practical, at least for large operations with the capital to invest. Sidewalk delivery robots, moving slowly among pedestrians, have scaled into millions of deliveries.
Highway trucking sits at an intermediate point. The driving environment is more structured than city streets, with consistent lane markings, limited access points, and relatively predictable traffic patterns, but weather, construction zones, and the sheer kinetic energy of a loaded tractor-trailer raise the stakes for any sensor failure. Robotaxis face the hardest version of the problem: dense urban traffic with jaywalking pedestrians, double-parked cars, construction detours, and edge cases that no training dataset fully covers. That Waymo has managed to scale to millions of trips in this environment is genuinely noteworthy, even as the geographic and weather constraints remain tight.
The next few years will likely determine whether autonomous vehicles remain clustered in favorable niches or begin to break into more challenging territory. Solving the weather-perception gap, building out V2X infrastructure, settling the liability framework, and managing the congestion effects of empty repositioning miles are all open problems. None of them are purely technical. Each one involves regulators, insurers, city planners, and the public in ways that make autonomous vehicles as much a policy challenge as an engineering one.