The Wheel Turns on Its Own. The Passenger Relaxes.

The steering wheel turned by itself, and nobody was holding it.

That was the first thing a South China Morning Post reporter noticed from the back seat of a Pony.ai robotaxi in Shenzhen, in an account the paper published on November 7, 2026. ^[1] The driver's seat was empty. The wheel moved anyway. The reporter felt a flicker of nervousness.

Then, according to the account, something happened that nobody sells as a feature but many passengers will recognise. There was no driver to make conversation with, and the silence turned out to be a pleasure. ^[1]

The same account covers a second ride in Beijing, in a car from Baidu's Apollo Go service. ^[1] Whether a staff member sat in that car, what either trip cost and how long the reporter waited are details this paper does not have yet.

Two moments in a back seat cannot say whether driverless taxis are becoming ordinary transport. They are still worth having, because the question is usually argued in numbers, and numbers do not tell you what it is like to sit there.

What the back seat tells you

The reporter's reaction is a small thing, and it is one person's. It would be a mistake to read it as a verdict from the public. But the order of feeling, first unease and then relief at being left alone, matches how new technology tends to settle into daily life. The strange part is noticed in the first minute. The convenience is noticed for the rest of the trip.

The account also describes a trade-off between comfortable private cabins and practical compromises. ^[1] The text this paper has seen does not say what those compromises are. That is where a reader deciding whether to trust the service would want detail, and it is the part still missing.

The scale the companies report

The companies have a lot to say about scale. Pony.ai said its robotaxi fleet passed 1,700 vehicles as of May 24, 2026. It said its average weekly paid orders in May were 119 percent higher than in January. ^[2]

Its robotaxi services revenue for the first quarter of 2026 was $8.57 million, against $1.73 million a year earlier, nearly five times as much. ^[2] The company has set a target of more than 3,500 robotaxis by the end of 2026, more than double the fleet it reported in May. ^[2]

Baidu's figures are larger. It said Apollo Go completed 3.2 million fully driverless rides in the first quarter of 2026, and had given the public more than 22 million rides in total as of April. ^[4] Spread over the quarter's 90 days, that averages roughly 35,000 driverless rides a day.

These are the companies' own numbers. Pony.ai's come from its filing with the US Securities and Exchange Commission, and Baidu's from its own announcement. ^[2][4] They count fleets, orders and rides, which are different things: a fleet is vehicles, an order is a paid request, and a ride is a trip completed. They do not count failures.

None of them tells you whether the person next to you at a Shenzhen bus stop can summon one, or whether she would want to.

A permit with edges

The edges of the service matter as much as its size. CNBC reported that Pony.ai held a Shenzhen permit to charge fares for fully driverless rides, with no staff inside, in parts of the Nanshan district. Listed service hours ran from 7:30 a.m. to 10 p.m. ^[3]

The permit did not cover the whole district, and CNBC reported that Pony.ai did not say how many cars it could run in the permitted area. ^[3] A city does not become a robotaxi city when a company is allowed to charge fares on some streets for some hours. It is still an enormous step from a test track.

The account does not say whether the reporter's Shenzhen ride fell inside that zone, or whether a driver's seat left empty by design means the same thing across a whole city.

The morning the cars stopped

The Wuhan episode is the sharpest reminder that a smooth ride and a safe system are different things.

In late March 2026, several Apollo Go vehicles stopped in the road in Wuhan. Police said initial findings pointed to a system malfunction. They said passengers got out safely, that no injuries had been reported, and that the cause was still under investigation. ^[5] The BBC reported that Baidu did not immediately respond to its request for comment. ^[5]

The police account was preliminary, and this paper has not seen a final finding. What it shows is a kind of failure no human fleet has. One taxi driver can have a bad day. A shared software fault can stop many cars at once.

Passengers in Wuhan left their cars safely, according to police. A system that fails this way can still fail safely. It is also the kind of event that a fleet of 1,700 or a count of 22 million rides can hide.

What the safety review has and has not shown

There is also a national safety review of self-driving cars. Pony.ai chief executive James Peng told Reuters that the company had completed evaluations connected to it and that its business had not been affected. He said the review did not amount to a suspension of licences. ^[6]

That is the company's account of its own position. It is not an independent finding that the system is safe, and Peng did not claim it was.

The figures the companies have released include no accident or injury rate that can be set against human-driven taxis. ^[2][4] That is the number a passenger most needs. Until a comparable rate is published, claims that robotaxis are safer than people are claims about the future.

What the rides add up to

My view is that the rides prove the service is real. A person can order a car, sit in it and arrive. They do not prove it is routine, and they do not prove it is safe, because routine means something specific. It means a service that works for most people in most places at most times, and that fails rarely enough that nobody thinks about it.

The next reporting that would settle things is plain. It needs crash and injury records, kilometres driven, a clear count of how incidents are logged, and the final word on Wuhan. Until then, readers can know what two rides felt like and what the companies say about their own scale. They cannot yet know how often these cars fail, or whether they fail less often than people do.