Signals · Article · Automotiveシグナル · 記事 · 自動車
Autonomous Driving [Explained]
The standard that defines driving automation discourages the word autonomous. Understanding why is the fastest way to understand the technology.

Here is an awkward fact for anyone who writes about this subject, including me. SAE J3016, the document almost everyone cites when they talk about levels of self-driving, lists “autonomous” and “self-driving” among its deprecated terms. I put the word in this headline anyway, because it is what people search for. But the standard has a point, and following it clears up most of the confusion around robotaxis, driver assistance and the cars in between.
The useful question is not how smart the car is. It is who is responsible for driving at this moment, under what conditions, and what happens when those conditions end. Answer those three and almost every headline about the technology sorts itself out. This article explains the levels, how an automated driving system sees and decides, where driverless vehicles actually operate in late 2026, and why proving they are safe is harder than building them.
Six levels that describe who is driving
SAE published J3016 in 2014 and revised it in April 2021 together with ISO, which publishes the same content as ISO/SAE PAS 22736. It defines six levels, numbered 0 to 5. The levels do not grade how capable or how good a system is. They describe how the driving task is divided between a human and the machine.
Two terms make the levels readable. The dynamic driving task is the real-time work of driving, steering, braking, accelerating, watching the road and responding to what happens. Planning the route is not part of it. The operational design domain, usually shortened to ODD, is the set of conditions a feature is designed for, such as road types, speeds, weather, time of day or a mapped city area.
At Level 0 the human does all of the driving, even if warning and emergency braking systems help. Level 1 handles either steering or speed. Level 2 handles both at once, but the human must supervise continuously and remains the driver. At Level 3 the system performs the whole driving task inside its ODD, and a human in the seat must be ready to take over when the system asks or when a vehicle failure is evident. At Level 4 the system also handles its own fallback inside the ODD, bringing the vehicle to what J3016 calls a minimal risk condition, usually a safe stop, without needing anyone. Level 5 is the same as Level 4 with no ODD limit at all.
Two consequences follow. A very capable Level 2 system is still Level 2 if a human has to watch it, which is why Tesla now names its product Full Self-Driving (Supervised). And there is no Level 2+ in the standard. That label is marketing vocabulary for Level 2 systems with more features.
J3016 also says it is technical rather than legal. It describes roles. It does not decide who pays when something goes wrong. Courts, regulators and insurers do that, and they do not always line up with the levels.
The difference between Level 2 and Level 3 is a person
Level 3 is the first point where the human is allowed to stop watching the road, and that makes it the hardest commercial step. Honda got there first. In March 2021 it began leasing 100 Legend sedans in Japan with a Traffic Jam Pilot system approved for conditional automation in congested highway traffic, each priced at 11 million yen. Mercedes-Benz followed with Drive Pilot, approved by Germany’s KBA in December 2021 for heavy traffic at up to 60 km/h. Nevada and California approved it in 2023 at up to 40 mph on specified freeways, and in December 2024 Germany cleared an update allowing up to 95 km/h in the right-hand lane.
The international rule behind these approvals is UN Regulation 157 on Automated Lane Keeping Systems. When it took effect in early 2021 it capped operation at 60 km/h. An amendment adopted in June 2022 raised the ceiling to 130 km/h and allowed lane changes. BMW added its own Level 3 Personal Pilot to the 7 Series in Germany in 2024.
China granted its first Level 3 product approvals in December 2025, for a Changan Deepal model limited to 50 km/h on designated Chongqing roads and a BAIC Arcfox model limited to 80 km/h on designated Beijing expressways. CnEVPost reported that the driver remains responsible for potential accidents, which sits uneasily with how J3016 describes Level 3. It is a reminder that the level describes the engineering, while liability is decided by local law.
Notice what all of these systems share. Each one works only in a narrow, well-defined slice of driving. That narrowness is deliberate. A system that knows exactly where its boundaries are can hand back control in time, and one that does not cannot.
How the car sees and decides
Most automated driving systems are described as a chain of jobs. Sensors capture the world. Localization works out exactly where the vehicle is, often against a detailed map. Perception identifies what is around it, such as cars, cyclists, lane lines and traffic lights. Prediction estimates what each of those things will do next. Planning chooses the vehicle’s own path and speed. Control turns that plan into steering, braking and acceleration commands.
The sensors each cover the others’ weak spots. Cameras read color, signs and lane markings but struggle in glare and darkness. Radar measures distance and speed through rain and fog but sees in low resolution. Lidar builds a precise three-dimensional point cloud with its own laser light, at a higher cost. Mercedes equips Drive Pilot with lidar alongside cameras and radar, and pairs them with redundant steering, braking and electrical supply so a single failure does not leave the car without a way to stop. Tesla takes the opposite position and relies on cameras.
The control end of that chain depends on the car’s own actuators accepting electronic commands, which is where automated driving meets drive-by-wire. The companion explainer in this series covers how those commands reach the brakes and the steering rack, and why redundancy there matters more than any sensor choice.
One network or many modules
The classic stack builds each job as its own module, with clear interfaces between them. Engineers can test perception separately from planning and see which part made a mistake. The newer approach trains one large neural network to go from camera pixels to driving commands, learning the intermediate steps from huge amounts of driving data. Tesla moved its Full Self-Driving software to this end-to-end design with version 12 in 2024, and UK-based Wayve has built its business around the same idea.
A third approach adds reasoning to the learned model, so it can explain why it chose an action. NVIDIA’s Alpamayo models are the most visible example, and I looked at what they change for developers in Alpamayo Moves the Cost of Autonomy. In practice most serious programs now mix the approaches, using learned models where they perform best and keeping explicit safety checks and rules where engineers need to prove behavior.
Alpamayo Moves the Cost of Autonomy
Where driverless rides actually run
Waymo is the clearest measure of progress. It reported 500,000 paid rides a week in March 2026, up from about 50,000 in May 2024, across what is now 14 to 15 US metro areas, with a fleet of about 4,000 vehicles by September 2026. Its stated goal is one million paid weekly rides by the end of this year, and it plans a Tokyo service with GO and Nihon Kotsu in 2027. Its vehicles had driven 271.3 million miles with no one in the driver’s seat through June 2026.
Baidu’s Apollo Go is the largest operator in China. It passed 23 million cumulative rides across 28 cities by mid-2026, though quarterly volume fell to about one million rides in the second quarter of 2026 from more than three million in the fourth quarter of 2025. Baidu cited operational adjustments in some cities, and Caixin reported a regulatory pause on robotaxi expansion in China in May 2026. Pony.ai reported 1,975 robotaxis at the end of June 2026 and WeRide more than 1,800, across China and several overseas markets, and both still report losses.
Tesla launched a robotaxi service in Austin in June 2025 with a safety monitor in the front passenger seat. It began rides with no employee inside for part of its Austin fleet around January 2026 and now runs rides without an employee aboard in six cities across Texas and Florida. Electrive reported that remote intervention and chase vehicles remained part of the operation. Public rides in the purpose-built Cybercab, which has no steering wheel or pedals, began in Austin in early September 2026. Within hours NHTSA opened an audit into Tesla’s self-certification that the Cybercab meets federal safety standards, and Tesla now has until October 30 to answer.
Amazon’s Zoox opened free public rides on the Las Vegas Strip in September 2025 in a vehicle with no manual controls. On July 30, 2026 it received the first federal exemption allowing paid passenger service in such a vehicle, and paid rides followed in August. In freight, Aurora launched driverless trucks between Dallas and Houston in May 2025, then put a human observer back in the cab at the request of its truck partner PACCAR. In 2026 it added a second-generation fleet with no one behind the wheel, and its second-quarter shareholder letter reported nearly 440,000 driverless miles through June.
Not every program survived. In October 2023 a Cruise robotaxi in San Francisco struck a pedestrian who had been thrown into its path by a human driver, then dragged her about 20 feet while trying to pull over. California suspended Cruise’s permits, the company later paid a 500,000 dollar criminal fine for a false report to federal investigators, and in December 2024 GM stopped funding the robotaxi business after investing more than 10 billion dollars.
The people behind the curtain
Driverless does not mean unsupervised by humans. Most operators run remote assistance centers. J3016 draws a sharp line here. In remote assistance, a person gives the vehicle information or advice, for example confirming that a lane is blocked, while the automated system still does the driving. In remote driving, a person controls the vehicle in real time, and the standard says plainly that remote driving is not driving automation.
That line matters for accountability. A remote assistant can be wrong, and the system still acts on the advice it receives. So the safety case has to cover the software and the people supporting it, including how quickly they respond and what they are allowed to tell the vehicle.
The ratio of remote staff to vehicles is one of the least disclosed and most important numbers in the business. It decides both the cost per ride and how gracefully a fleet behaves on a bad day.
Proving it is safer takes more miles than anyone has
In 2016 RAND researchers Nidhi Kalra and Susan Paddock calculated that showing an automated fleet matches the US human fatality rate with 95 percent confidence would take about 275 million failure-free miles. That is 100 cars driving around the clock for about 12.5 years. Showing that it is 20 percent better would take around 11 billion miles. Driving alone cannot carry the full burden of proof, which is why simulation, scenario testing and formal safety cases matter so much.
Waymo has the largest public record. Using data through June 2026, it reports 95 percent fewer crashes causing serious injury or worse and 82 percent fewer injury crashes than human benchmarks for the cities it serves. Earlier versions of this analysis were peer reviewed in the journal Traffic Injury Prevention, and a separate study with the reinsurer Swiss Re found large reductions in insurance claims.
Waymo is also direct about the limits. Its published benchmarks cover surface streets and leave out freeway driving. It still does not have enough miles for a statistically meaningful comparison of fatal crashes. And it says the comparison does not by itself establish that the system is safe enough. That honesty is worth noticing, because 271 million rider-only miles is the largest driverless record any operator has published, and RAND’s bar of roughly 275 million fatality-free miles just to show parity rises with every fatal crash.
Regulators rely heavily on what companies report. NHTSA’s Standing General Order requires operators to report crashes involving automated systems, but the reports are company-written and some, including Tesla’s, are heavily redacted. The Cruise case showed the risk. Its deferred prosecution agreement with the US Justice Department centered on a report that left out the dragging.
What autonomous really means
Every driverless service operating today works inside a bounded domain, whether that is a set of city districts, a freight corridor, a single lane of Autobahn or a casino strip. Nobody deploys Level 5, and nobody has a credible date for it. What has changed is that the domains are getting larger, the fleets are getting bigger, and a few operators now publish enough data for outsiders to check their claims.
So when a new announcement arrives, the useful reading is the one J3016 suggests. Who is driving, the human or the system? Within what ODD? What happens at the edge of that ODD? And what evidence, beyond the company’s own description, shows the answer holds?
The word autonomous will stay in headlines, including mine. The engineering will keep being decided by those four questions.
Words you will hear
SAE J3016. The SAE recommended practice that defines Levels 0 to 5 of driving automation. Current edition April 2021, aligned with ISO/SAE PAS 22736.
ADS. Automated driving system. The hardware and software that perform the whole driving task at Levels 3 to 5.
DDT. Dynamic driving task. The real-time steering, braking, accelerating, monitoring and responding involved in driving.
ODD. Operational design domain. The conditions a feature is designed to operate in, such as roads, speeds, weather and geography.
Minimal risk condition. A stable, usually stopped state that the system or the human brings the vehicle to when driving cannot continue.
Remote assistance. A remote human giving information or advice while the automated system keeps driving. Not the same as remote driving.
End-to-end. A design where one learned model maps sensor input to driving output, instead of separate hand-built modules.
This article is part of the AutomotiveCloudWatch Explained series. Before it, read Software-Defined Vehicle [Explained] and Zonal Architecture [Explained].
If this helped, subscribe free at automotivecloudwatch.com to get each new explainer when it publishes.
All opinions are my own and do not reflect those of my employer.
Get the next Executive Briefing by email
Each new briefing, in your inbox.
Subscribe




