Autonomous Vehicles: what is actually working now?
In our autonomous vehicle market deck, you will find everything you need to understand the market
SUMMARY
Autonomous vehicles are genuinely working now, but the strongest systems succeed by narrowing the driving problem rather than solving every road, weather condition and use case at once.
Robotaxis have crossed the clearest commercial threshold. Waymo and Baidu already complete driverless passenger rides at volumes measured in hundreds of thousands per week, while Pony.ai and WeRide are building fleets measured in the thousands.
The most important change is that autonomous-driving evidence is now operational. Passenger rides, freight miles, fleet utilization, crash outcomes, revenue and unit economics tell us far more than another prototype or test permit.
Geofencing is a core part of the current business model. The companies moving fastest are expanding prepared operating domains city by city, which means wider coverage has become a practical measure of progress rather than a side detail.
China and the US are scaling in different ways. The US market is dominated by Waymo, while China has several large operators expanding at the same time, creating a broader competitive field in robotaxis.
Autonomous trucking works for a similar reason: predictable, repetitive routes make every extra operating hour valuable. Aurora, Kodiak and Gatik are already moving freight without a driver in the cab on selected routes and industrial corridors.
Industrial autonomy is further ahead than city driving. Caterpillar's autonomous mining trucks have moved more than 11 billion tonnes of material, showing how quickly autonomy becomes ordinary infrastructure when roads, traffic and access are controlled.
The consumer-car story is different. Tesla FSD Supervised and GM Super Cruise operate across enormous mileage, but the driver still carries responsibility; Mercedes Drive Pilot goes further by taking over the driving task under tightly defined Level 3 conditions.
Economics are beginning to separate from technical feasibility. Pony.ai says some Gen-7 robotaxis have reached city-level unit-economics breakeven, yet autonomous-vehicle companies still spend heavily on engineering, fleets and expansion, so vehicle-level viability does not mean the wider business is profitable.
The clearest pattern across robotaxis, freight, mines and consumer systems is that autonomy gets stronger as the operating domain gets narrower. The technology already solves several valuable transport problems, while the cheap personal car that can drive almost anywhere without supervision remains out of reach.
This market map, featured in our autonomous vehicle market deck, highlights top companies and startups in the autonomous vehicle market
Are autonomous vehicles actually working now?
Autonomous vehicles are working today in several real businesses, but the answer changes completely depending on whether we mean robotaxis, trucks, consumer cars or industrial vehicles.
The easiest way to see the difference is to look at who is still sitting behind the wheel.
Waymo and Baidu are already carrying paying passengers with nobody driving. Aurora, Kodiak and Gatik are moving freight without a driver in the cab. Mercedes sells a Level 3 system that can legally take over the driving task under specific conditions. Tesla and GM have consumer systems used across billions of miles, although their drivers still carry responsibility.
A few years ago, most autonomous-driving claims leaned heavily on test fleets and demonstrations. We now have enough commercial mileage, passenger rides, freight deliveries, fleet utilization and safety data to judge individual use cases on their own.
The pattern is already fairly clear. Autonomous driving works best today when a company can limit where the vehicle goes, understand that environment extremely well and repeat the same kind of trip thousands of times.
Are driverless robotaxis actually carrying normal passengers?
Driverless robotaxis are already a real transport service, with Waymo and several Chinese operators now completing millions of rides without a human driver.
Waymo currently provides more than 500,000 fully autonomous rides each week, according to its latest public figures. The company says it operates in more than 15 major US cities, although public availability varies by market.
Houston is a useful example of how these launches have changed. Waymo initially opened the service to selected riders, carried more than 100,000 people there, and later opened the app to everyone. That looks much closer to a normal ride-hailing rollout than an engineering trial.
The annual numbers tell the same story. Waymo completed about 15 million rides during 2025 alone, more than triple the previous year's volume. Baidu's Apollo Go completed another 3.2 million fully driverless rides during the first quarter of 2026, with weekly volume peaking above 350,000. Baidu said cumulative public rides had passed 22 million by April.
Those volumes are too large to write off as demonstrations. People are already using driverless cars for ordinary trips at meaningful scale.
| Robotaxi operator | Recent real-world evidence |
|---|---|
| Waymo | 500,000+ fully autonomous rides per week |
| Baidu Apollo Go | 3.2m driverless rides in Q1 2026 |
| Pony.ai | 1,975 robotaxis at the end of Q2 2026 |
| WeRide | 1,800+ robotaxis reported this summer |
As this chart shows, and as featured in our autonomous vehicle market deck, search interest in autonomous vehicles has continued to rise
Has Waymo really reached scale yet?
Waymo has reached meaningful commercial scale in autonomous driving, even though its service is still small beside Uber or the wider taxi market.
The growth rate is more revealing than the headline fleet size. Waymo was doing around 250,000 paid trips per week in May 2025. By early 2026, the company had crossed 400,000. Its latest published material now puts the figure above 500,000.
That roughly doubles weekly volume in little more than a year.
The annual comparison is just as useful. Waymo delivered around 15 million rides during 2025. A sustained pace of 500,000 rides per week would equal roughly 26 million rides over a full year.
Expansion has changed too. Phoenix and San Francisco once made up most of the story. Waymo then moved into Los Angeles, Austin and Atlanta before opening or preparing services across a much wider list of US cities. Its current material says the Waymo Driver operates in more than 15 major US cities.
Waymo's latest operating disclosure also gives us a sense of fleet productivity. In February 2026, the company said roughly 3,000 vehicles were completing more than 400,000 weekly rides and driving over four million miles per week. That works out to more than 130 rides per vehicle per week across the fleet, although utilization will obviously vary by city and vehicle.
Waymo has moved past proving that one fleet can work in one favorable market. The harder test now is whether the same system keeps working while the fleet and the number of cities grow together.
If you want more recent data on this point, please see our latest autonomous vehicle market report.
Is China already competing with Waymo on robotaxis?
China is already operating robotaxis at a scale comparable with the largest US deployments, and the biggest difference is that several Chinese companies are scaling at once.
Baidu's Apollo Go is the clearest starting point. During the first quarter of 2026, Apollo Go completed 3.2 million fully driverless rides. Weekly volume peaked above 350,000 in March, while cumulative public rides passed 22 million the following month.
Baidu also reported more than 220 million fully driverless kilometres across its global fleet by May. That gives Apollo Go a substantial operating history rather than a recent burst of deployment.
Pony.ai is growing quickly as well. The company's robotaxi fleet reached 1,975 vehicles at the end of June 2026, up from 961 near the end of 2025. Pony.ai is targeting more than 3,500 vehicles by year-end.
WeRide adds a third scaled operator. The company reported more than 1,800 robotaxis in a global Level 4 fleet of roughly 3,400 vehicles during the summer. Average utilization exceeded 21 rides per robotaxi per day during the second quarter and reached 28 on peak days.
The market structure looks different from the US. America currently has one obvious leader in Waymo. China has several companies trying to build similar density at the same time.
This chart, included in our autonomous vehicle market deck, illustrates yearly VC funding for autonomous vehicle startups
Is Tesla's Robotaxi finally real?
Tesla's Robotaxi has finally crossed into real passenger service, although Tesla still has far less driverless operating history than Waymo or Baidu.
Tesla currently advertises autonomous Robotaxi rides in Austin, Dallas, Houston, Miami, Orlando and Tampa. Most of that network uses Model Y vehicles.
The bigger change is Cybercab. Tesla has started producing its purpose-built two-seat autonomous vehicle, which has no steering wheel or pedals. Public Cybercab rides are currently available in limited parts of Austin.
Cybercab was previously something investors could mostly judge through prototypes, factory plans and demonstrations. Customers can now actually ride in one.
Tesla's own disclosures still show how early this phase is. During the second quarter, the company said production had started and that employee rides had begun on the Gigafactory Texas campus in July. Commercial availability followed in a restricted Austin area.
The comparison with Waymo remains lopsided. Waymo already has hundreds of millions of fully autonomous miles behind it. Tesla is only beginning to build comparable driverless mileage with paying passengers.
Tesla Robotaxi is operational now. Calling it proven at large scale would still run ahead of the evidence.
If you want more recent data on this point, please see our latest autonomous vehicle market report.
Are robotaxis safer than human drivers yet?
Waymo now has enough real-world data to make a strong safety claim: its driverless system has produced far fewer injury crashes than human drivers covering comparable roads.
Waymo's latest safety analysis covers 220.6 million rider-only miles through the end of March 2026 across Phoenix, San Francisco, Los Angeles, Austin and Atlanta.
Compared with human drivers covering the same distances in those areas, Waymo reported 82% fewer crashes involving any reported injury. Crashes causing serious or fatal injuries were 94% lower. Airbag-deployment crashes were down 82%.
The pattern also holds for vulnerable road users. Waymo found 93% fewer injury-causing crashes involving pedestrians, 84% fewer involving cyclists and 84% fewer involving motorcyclists.
The sample is large enough to change the discussion. Two hundred million driverless miles still leave room for methodological arguments, especially because human crash databases and autonomous-vehicle telemetry are collected differently. They leave much less room for the idea that Waymo's safety result comes from a lucky small sample.
There is one important boundary around this conclusion. The data supports Waymo's system inside Waymo's operating areas. It does not tell us that every autonomous-driving product is safer than a human.
| Crash measure | Waymo versus comparable human driving |
|---|---|
| Any reported injury | 82% fewer |
| Serious or fatal injury | 94% fewer |
| Airbag deployment | 82% fewer |
| Pedestrian injury | 93% fewer |
| Cyclist injury | 84% fewer |
This chart, included in our autonomous vehicle market deck, shows how Waymo is winning in autonomous vehicles
Does geofencing make robotaxis less impressive?
Geofencing is a big part of why robotaxis work today, and expanding those boundaries is one of the best ways to measure real technical progress.
Waymo, Baidu, Pony.ai, WeRide, Zoox and Tesla all limit where fully autonomous rides can happen. Inside those zones, the company has already studied road layouts, pickup areas, traffic rules, charging needs, construction patterns and local operating procedures.
The useful question is how much those zones can grow without service quality falling apart.
Waymo has gradually moved from Phoenix into denser environments such as San Francisco and Los Angeles, then into Austin, Atlanta, Houston and other markets. It is also preparing deployments in places such as Denver, where winter weather adds another layer of difficulty.
China shows a similar progression. Pony.ai has expanded coverage across major parts of Guangzhou and Shenzhen, including busy transport hubs. WeRide says its Guangzhou operating area has tripled since the end of 2025 and supports round-the-clock service.
The areas are getting larger and more varied, which is genuine progress. They are still prepared operating domains.
A human can land in an unfamiliar city, rent a car and attempt almost any legal road immediately. Commercial Level 4 robotaxis still need far more preparation than that.
Can robotaxis actually make money now?
Individual robotaxis are starting to cover their direct operating costs in China, although the companies building them are still spending far more money than their fleets earn.
Pony.ai gives us unusually specific evidence. The company says its seventh-generation robotaxis reached city-wide unit-economics breakeven in Guangzhou late in 2025 and then in Shenzhen during 2026.
During the Shenzhen breakeven month, each Gen-7 robotaxi generated around RMB338 of daily net revenue from roughly 23 orders.
Pony.ai says its calculation includes vehicle and autonomy-hardware depreciation, charging, maintenance, remote operations, insurance, labor, parking and network infrastructure. The company's latest hardware generation costs roughly 70% less than its previous autonomous-driving package, which helps explain the improvement.
The revenue growth is now large enough to notice. Pony.ai reported $12.1 million of robotaxi-services revenue in the second quarter of 2026, up 691% year over year. Fare-charging revenue grew 849%.
Then we get to the expensive part.
Pony.ai generated only $36.2 million of total company revenue in that quarter while continuing to spend heavily on engineering and expansion. Aurora has the same basic problem in trucking: the company expects only $14 million to $16 million of 2026 revenue while using roughly $190 million to $220 million of cash per quarter.
The first economic milestone has arrived. Some autonomous vehicles can apparently pay for their own daily operation in dense deployments. The wider companies still have years of expensive development and fleet expansion to finance.
If you want more recent data on this point, please see our latest autonomous vehicle market report.
This chart, included in our autonomous vehicle market deck, illustrates yearly funding for autonomous vehicle startups
Is autonomous trucking actually working now?
Autonomous trucking is already working commercially on selected routes, and the economics can be unusually attractive because driverless trucks can stay productive for far longer each week.
Aurora's work with Werner gives us one of the clearest examples. Its driverless trucks have been averaging more than 4,000 miles per week. Aurora says that translates into an annual run rate above 225,000 miles per truck.
That kind of utilization is difficult with a conventional long-haul driver because working hours are regulated.
Aurora expects more than 200 driverless trucks to be operating by year-end. At that scale, the company estimates an annualized transportation-as-a-service revenue run rate of roughly $80 million.
Kodiak is using autonomy in a different freight environment. Atlas Energy had 28 Kodiak-powered driverless trucks running in the Permian Basin by the end of March 2026. The partners now plan to expand the fleet to 100 autonomous trucks by mid-2027. Today those vehicles mainly operate in industrial logistics; public-road expansion is planned for 2027 if regulatory and operational milestones are met.
Gatik gives us another model: repetitive regional deliveries between distribution centres and stores. The company recently reported 85,000 completed fully driverless orders, more than $600 million in contracted revenue and a 99% on-time delivery rate.
Across all three companies, the commercial logic is similar. Autonomy is especially attractive when freight moves along predictable, high-frequency routes and every extra operating hour has obvious value.
| Company | What is already working |
|---|---|
| Aurora | 4,000+ miles per week on some driverless trucks |
| Kodiak / Atlas | 28 driverless trucks in Permian operations |
| Gatik | 85,000 driverless commercial orders completed |
| Pony.ai | Robotruck revenue reached $13.3m in Q2 2026 |
Are mines and quarries actually ahead of robotaxis?
Autonomous mining trucks remain the most mature large-scale example of machines doing productive transport work without human drivers.
Caterpillar currently has nearly 700 autonomous haul trucks operating around the world. Together, those vehicles have moved more than 11 billion tonnes of material.
That is a huge amount of real work.
The conditions explain why mining got there first. Operators control the roads, traffic patterns and access to the site. Trucks repeat similar routes for long periods. Removing driver shifts can directly increase vehicle utilization.
Caterpillar is now applying the same model beyond giant mines. At Luck Stone's Bull Run quarry in Virginia, four autonomous 100-ton Cat 777 trucks were enough to create a working autonomous fleet. The operation reached staffed-machine productivity shortly after deployment and moved its first million tonnes within months.
Mining gives the wider AV industry a useful reference point. The more controlled the environment and the more repetitive the journey, the easier autonomy is to turn into an ordinary business tool.
This chart, included in our autonomous vehicle market deck, compares the main business model options for autonomous trucking companies
What autonomous driving is actually working in normal cars?
Hands-free driver assistance already works at mass-market scale, with Tesla and GM collecting billions of real driving miles while keeping the human responsible for supervision.
GM's Super Cruise recently passed one billion hands-free miles. Nearly 750,000 Super Cruise-equipped vehicles are now on North American roads across more than 20 models, and the compatible road network covers more than 600,000 miles.
Those are consumer-product numbers, not pilot numbers.
Tesla operates at even larger software mileage scale. Its latest public safety material says FSD Supervised has accumulated more than 12 billion miles, including billions of miles on city streets.
The experience can feel surprisingly close to autonomous driving. These systems can steer, accelerate, brake and handle long stretches of a journey with very little physical input from the driver.
Responsibility is where the difference becomes clear. GM still monitors the driver's attention with Super Cruise. Tesla tells FSD Supervised users to remain actively engaged and ready to take control.
For today's consumer, this is the version of automated driving that has genuinely reached mass adoption: the car does a large share of the mechanical driving, while the person remains the fallback.
If you want more recent data on this point, please see our latest autonomous vehicle market report.
Can you buy a car that really lets you stop watching the road?
Mercedes-Benz already sells a genuine Level 3 system that lets the driver look away from the road under approved conditions, but the usable situations remain tightly defined.
Drive Pilot can take over the driving task on Germany's Autobahn network at speeds up to 95 km/h when its operating conditions are met. Mercedes offers the system on the S-Class and EQS.
The legal difference from Tesla FSD Supervised or GM Super Cruise is important. When Drive Pilot is properly engaged, Mercedes assumes the dynamic driving task. The driver can use the time for other activities and only needs to take back control when requested.
Mercedes uses more than 35 sensors, including cameras, radar, ultrasound and lidar, along with detailed high-definition maps and precise positioning.
The limitation appears as soon as the required conditions disappear. Road type, lane, traffic situation and other system requirements all affect whether Drive Pilot can activate. The driver also needs to remain available for a takeover request.
So eyes-off driving already exists in a production consumer car. The usable slice of driving remains much narrower than what most people imagine when they hear "self-driving car."
This chart, featured in our autonomous vehicle market deck, shows the share of revenue generated by each customer segment in the autonomous vehicle market
Have autonomous vehicles figured out bad weather and weird road situations?
Autonomous vehicles handle far more difficult conditions than they used to, but unusual events and extreme weather still slow down expansion.
We can see that in where companies choose to launch.
Waymo started building commercial scale in Phoenix, where weather is relatively predictable. It later expanded into the dense streets of San Francisco and Los Angeles. The company is now preparing markets such as Denver, where snow and colder conditions make the job harder.
Waymo's sixth-generation Driver was specifically designed to support a wider range of weather while using a cheaper sensor setup.
Chinese operators are pushing in the same direction. Pony.ai has reported 24-hour operation through heavy rain and has run its fleets during snow in Beijing. Its vehicles increasingly serve airports, railway stations and dense central-city districts, where driving is much less predictable than a suburban test route.
The remaining problem comes from events that happen rarely but still require a correct decision: a police officer manually directing traffic, a lane suddenly blocked by debris, strange temporary road markings, flooding, emergency vehicles or construction layouts that barely resemble the map.
Companies deal with those cases through simulation, cautious fallback behaviour and remote assistance. Waymo said in February 2026 that around 70 remote-assistance agents were on duty worldwide at a given time for a fleet of roughly 3,000 vehicles. Those agents do not remotely drive the cars; they can provide information when the autonomous system requests help.
That ratio is interesting. Roughly one active remote-assistance worker for every 43 vehicles shows that large driverless fleets can operate without recreating one human operator for every car.
Why can't a successful robotaxi just launch everywhere?
Launching an autonomous vehicle in a new city still requires a lot of local work, although companies are clearly getting faster at it.
The driving software is only one piece. A commercial fleet needs charging, cleaning, maintenance, depots, roadside support, remote assistance, regulatory approval, airport access, sensible pickup areas and enough cars in one place to keep passenger waiting times reasonable.
Waymo's recent expansion suggests that this process is becoming more repeatable. The company is entering several markets in parallel and manufacturing vehicles at higher volume with Magna in Arizona. Its newer autonomy hardware is also designed to cost less than previous generations.
Pony.ai is attacking the same problem from the vehicle side. Its seventh-generation autonomous-driving hardware costs roughly 70% less than the previous version, and the company says its complete domestic robotaxi bill of materials should fall below RMB230,000 by mid-2027.
Partnerships may speed things up further. Uber can supply customer demand and local ride-hailing infrastructure while autonomous-driving companies provide the actual Driver. Waymo, WeRide, Pony.ai and other AV developers increasingly use versions of this model.
Autonomous vehicles these days look more like a transport network that has to be built city by city than a software feature that can simply be switched on worldwide.
This chart, included in our autonomous vehicle market deck, shows how robotaxi platform technology has evolved over time
So what is actually working in autonomous vehicles right now?
Autonomous vehicles are genuinely working today in robotaxis, selected freight routes, industrial sites and advanced driver assistance; unrestricted self-driving for ordinary consumers remains out of reach.
The strongest evidence comes from repeated commercial use.
Waymo and Baidu now complete driverless passenger rides at volumes measured in hundreds of thousands per week. Pony.ai and WeRide operate fleets measured in thousands of robotaxis and are improving vehicle utilization rapidly.
Freight autonomy has crossed the same line in narrower environments. Aurora moves loads without a driver and gets more than 4,000 miles per week from some trucks. Kodiak-powered vehicles are already working in the Permian Basin. Gatik has completed 85,000 driverless commercial deliveries.
Industrial autonomy is further ahead again. Caterpillar's autonomous trucks have already moved more than 11 billion tonnes of material.
Consumer cars show a different level of progress. Super Cruise and FSD Supervised work across enormous mileage, while the driver continues to supervise. Mercedes has pushed further with genuine eyes-off Level 3 driving, though only under tightly defined conditions.
The combined evidence is pretty decisive. Autonomous driving has already solved several valuable transport problems. Geofenced city driving works. Repetitive freight works. Controlled industrial transport works extremely well. Hands-free consumer assistance works at mass scale.
What we still cannot buy is the version people were promised most often: a reasonably priced personal car that can drive almost anywhere, through almost anything, with nobody paying attention.
For now, autonomous vehicles succeed fastest when the problem gets narrower. Every serious commercial deployment reviewed here follows that pattern.
If you want more recent data on this point, please see our latest autonomous vehicle market report.
OUR METHODOLOGY
This analysis asks whether autonomous vehicles are actually working now by separating robotaxis, autonomous freight, industrial vehicles, supervised consumer systems and Level 3 consumer driving. We treat "working" as something that has to show up in real operation, not just in a test, prototype or announcement.
We broke the question into the dimensions that best show whether autonomy has moved into practical use: commercial operation, scale and utilization, safety, economics, operating-domain breadth, and the role still required from a human driver. We gave more weight to paid rides, fully driverless mileage, completed freight movements, active fleet size, utilization, revenue, unit economics and observed crash outcomes than to future deployment targets.
We kept different levels of automation separate. Supervised systems such as Tesla FSD Supervised and GM Super Cruise were assessed as supervised systems. Mercedes Drive Pilot was assessed around the transfer of the driving task under approved Level 3 conditions. Level 4 fleets were assessed inside the operating domains for which they were designed, using the SAE and NHTSA framework to keep responsibility clear.
Geofencing was treated as a measure of operating scope rather than a reason to dismiss a deployment. Expansion into larger, denser and more varied environments gives a more useful indication of progress than asking whether a vehicle can already drive everywhere.
For safety, we gave the most weight to real-world exposure and comparable crash outcomes. For economics, we separated the economics of an individual autonomous vehicle or service from the profitability of the company developing it, so unit-level progress is not confused with a mature profitable business.
No single deployment determined the conclusion. We looked for patterns that repeat across passenger transport, freight, industrial operations and consumer vehicles, and prioritized recent operating and financial disclosures wherever possible.
Key sources used for this analysis include Waymo on current ride scale and city expansion, Waymo's 2025 operating review, Waymo's latest safety analysis, Waymo on remote assistance, Baidu's Q1 2026 results for Apollo Go, Pony.ai's Q2 2026 results, Pony.ai on Gen-7 robotaxi breakeven in Shenzhen, WeRide's Q2 2026 operating update, Aurora's Q1 2026 shareholder letter, Kodiak and Atlas Energy on driverless freight operations, Gatik on fully driverless commercial orders, Caterpillar on autonomous haulage scale, GM on Super Cruise usage, Tesla's FSD Supervised safety and mileage material, Tesla's Robotaxi service page, Mercedes-Benz on Drive Pilot, NHTSA's automated-vehicle framework, and SAE J3016.
In our autonomous vehicle market deck, we identify pain points entrepreneurs should prioritize
Related blog posts
- Autonomous Vehicles: what are the biggest challenges now?
- Autonomous Vehicles: what are the biggest unsolved problems?
- Autonomous Vehicles: what is getting real adoption now?
- The latest funding news in the autonomous vehicle market
