Spend a week in China riding in driverless cars, meeting the companies that build them and the authorities that regulate them, and something unexpected happens: you stop asking whether autonomous driving works. You come home with fewer doubts about the technology, and a much longer list of questions about what we should actually do with it.

In August 2026, I was invited by the Mobility Masterclass to join their China Study Tour, organised by Espaces-Mobilités with the support of CCI France Chine. Over seven days, we travelled through Beijing, Guangzhou and Shenzhen to experience autonomous mobility in China first-hand.

The programme covered virtually every layer of the ecosystem: Pony.ai, Neolix, the Beijing M-Zone, TransInfo, DiDi, Beijing’s High-level Automated Driving Demonstration Zone, Guangzhou’s Intelligent Connected Vehicle Demonstration Zone, WeRide, XPENG, AION, DeepRoute.ai and Huawei.

We experienced robotaxis and automated buses, visited autonomous delivery fleets, saw monitoring and operations centres, explored OEM production and tested increasingly advanced passenger vehicles.

I came back with ten observations on the state of autonomous mobility in China. Here they are.

1. A robotaxi ride is wonderfully boring, until human drivers make sure it doesn’t last

I expected the technology to be impressive. I did not expect it to become positively boring quite so fast. The robotaxis we rode were smooth, predictable and reassuring, the word that kept coming up among us. By the third or fourth ride, the empty driver’s seat had stopped being the interesting part of the journey.

The contrast only really landed when we went back to conventional taxis. Human driving suddenly felt coarse: harder acceleration and braking, more honking, more improvisation. Next to that, the automated vehicles seemed calm and unusually anticipatory.

One moment in a WeRide vehicle captured it. An oncoming human driven car cut across our lane although it had to yield. Our vehicle registered the situation, sounded its horn, kept monitoring the other car and, when the driver came through anyway, braked hard and avoided the collision. Technically, it behaved perfectly.

And yet I got out of that car with an uncomfortable question. How will people behave around vehicles they know will almost always back down? Once you have learned that an automated car will brake rather than fight you for priority, sooner or later somebody will take advantage of it. Autonomous mobility is not only about machines learning to read humans. It is just as much about humans learning to read machines, and about the rules we put in place before that becomes a habit.

2. Autonomous mobility in China has moved past pilot projects

The most striking difference between autonomous mobility in China and Europe is not the technology. It is the scale, and the fact that nobody there talks about pilots any more.

The numbers make the point. As of August 2026, Pony.ai reported more than 1’900 robotaxis, over 200 robotrucks and more than 100 million autonomous kilometres on public roads. WeRide crossed 1’000 robotaxis in January 2026 and expects more than 2’600 this year, on the back of 2’000 mass-produced GXR vehicles built with Geely’s Farizon. Baidu’s Apollo Go reported more than 22 million public rides across 27 cities by April 2026, 330 million autonomous kilometres, 220 million of them fully driverless, and weekly rides peaking above 350’000 in March.

DiDi, the Chinese equivalent of Uber or Bolt, has built an entire operational ecosystem around autonomous mobility. Its 24-hour Huiju Port brings safety, user services, operations, maintenance and fleet support together in one place, while its latest robotaxi platform combines 33 sensors with an interior designed around the passenger rather than the driver.

In logistics, the scale is even harder to miss. At Neolix, we were not shown a single prototype under a spotlight in a demonstration hall. We walked past rows of autonomous delivery vehicles built for different jobs: compact units for dense streets, larger ones for heavier loads, and specialised versions doing things like mobile retail.

China already has more than 60’000 urban autonomous logistics vehicles on its roads, with Neolix and Zelos leading the market at roughly 25’000 each. That near-duopoly looks less like a technological head start than a regulatory artefact: each district writes its own rules for autonomous logistics, and tends to admit one operator at a time. It is unlikely to hold: the competition is closing fast.

Once you have tens of thousands of these vehicles on public roads, the difficult questions stop being technical and become industrial:

• How do we operate thousands of them?
• How do we maintain them?
• Where do they stop?
• How do authorities supervise them?
• How are they integrated into the wider mobility system?

That is a completely different conversation from the one we are having in Europe, where most projects are still asking whether the vehicle can do it at all. In China that question has largely been answered, and the hard work has moved somewhere else: to the depot, the control room and the maintenance schedule. That shift, from proving the vehicle to running the system around it, may be the most important thing we brought home.

3. Beijing’s most valuable lesson was about the infrastructure: what not to build

One of the most fascinating visits of the week was Beijing’s M-Zone. Its roads and intersections are lined with intelligent roadside infrastructure: cameras, radars, LiDARs and communication equipment. At first sight, it is the textbook vision of connected autonomous mobility, intelligent vehicles in constant dialogue with intelligent infrastructure.

Then we asked what it is actually used for. The vehicle developers told us plainly that their cars do not need it to drive. The authorities confirmed it.

The explanation was honest. Around 2020, Beijing invested heavily in vehicle-to-infrastructure systems, at a moment when nobody knew how much automated vehicles would depend on roadside perception and communication. Six years later, on-board intelligence has advanced so fast that the dependency has largely evaporated. Parts of the infrastructure are therefore being scaled back, while what remains is being put to a use it turns out to be genuinely good at: observing traffic and understanding how automated and conventional vehicles interact.

For Europe, that is worth a great deal. We should be careful about committing large sums to infrastructure built on assumptions about what tomorrow’s vehicles will require. Roadside sensing and connectivity can genuinely improve traffic management, and that alone can justify targeted investment. But automated vehicles have to be able to drive without it, or they will never scale.

Sometimes the most useful thing you learn from a pioneer is not what they are building. It is what they have discovered they no longer need.

4. Regulators don’t need all the data, they need an interface to the right data

The regulatory monitoring infrastructure impressed me as much as the vehicles. In Beijing and in Guangzhou, we stood in front of command-centre dashboards giving authorities a live overview of autonomous and connected vehicle activity. The Guangzhou platform displayed hundreds of vehicles at once: operating status, mileage, vehicle category and individual vehicle data, in real time.

Beijing pairs this with a governance idea worth borrowing. Rather than obliging operators to open up all their internal systems to the state, the authorities can require a dedicated on-board unit that collects precisely the datasets needed for regulatory oversight. The result is a clean, defined interface between operator and authority.

Which data that unit should collect in Europe is a separate conversation, and one that would have to start from our privacy and data-protection rules. But the principle travels well:

Authorities need the right data, not necessarily all the data.

In Europe, where authorities will be the ones authorising and supervising L4 deployments, designing that interface deserves attention now rather than after the first fleets are running. Requiring a standardised on-board unit in automated vehicles strikes me as the right place to start.

5. Operations may end up mattering more than the driving system itself

The observation that came back most often during the week had nothing to do with driving. It was how much work sits behind the vehicle. Autonomous mobility at scale needs far more than good perception algorithms: the cars have to be cleaned, charged, maintained, dispatched, remotely supported, supervised and recovered when something goes wrong.

The depot at DiDi made this concrete. Rows of automated vehicles stood next to purpose-built operational infrastructure, including an automated car wash, because at that fleet size, even cleaning becomes a process you industrialise.

And the depot is only one floor of the building DiDi is putting up. The company is positioning itself along the entire vertical: the app that hails the ride, the operator that runs the fleet, the maintenance and cleaning behind it, and increasingly the autonomous driving technology in the vehicle itself. Imagine Uber owning its own robotaxis, servicing them in its own depots and developing the driving system that moves them. If the operational layer is where an L4 service is won or lost, owning every floor of it is a rational strategy, and an extremely difficult one to compete against.

This matters for Europe more than we tend to admit, and not only competitively. Our attention goes almost entirely to the autonomous driving system, because that is the part that feels like the hard engineering problem. But the economic viability of an L4 service will increasingly be decided by the operational layer around the vehicle, and by whether we can automate and optimise that too.

6. The next disruption may not come from robotaxis

For years, autonomous mobility has been illustrated with the same picture: a purpose-built robotaxi bristling with sensors. China is quietly pointing at a different path, in which the technology simply moves into ordinary, mass-produced passenger cars.

DeepRoute.ai is the clearest example. Its end-to-end systems work without high-definition maps, and the company sits directly alongside OEMs, building ever more capable driving technology into series-production cars. Its own figures point to more than ten production models and over 200’000 units planned for the mass market in 2026.

Huawei plays a similar game from a different starting point, supplying software, advanced driving systems, intelligent cockpits and other components to OEM partners through its automotive ecosystem.

At XPENG, that future stopped being an argument and became a test drive.

Sitting in an ordinary car that drives itself is a fundamentally different experience from sitting in a purpose-built robotaxi. It feels less like a demonstration and more like a preview of your own next car, which is precisely why it may be the more disruptive of the two paths. If L4 becomes a feature embedded in millions of private vehicles, autonomous mobility will not arrive through professional fleets that authorities can license, monitor and steer. It will arrive one privately owned car at a time.

For Europe, and for Switzerland in particular, that raises questions we have barely begun to discuss.

• What does remote supervision mean for an individually owned L4 vehicle?
• How should its operating area be authorised by authorities?
• And what happens to traffic volumes and modal split if private car travel suddenly becomes much more attractive because travellers can work, rest or sleep instead of driving?

7. The most impressive mobility system we observed was Chinese public transports

We arrived assuming that China’s mobility future would revolve around the car. It does not. The public transport systems we used in Guangzhou and Shenzhen were extensive, fast and remarkably cheap. Our conclusion as a group was that robotaxis are far more likely to replace trips made today by conventional taxi or private car than to pull significant numbers of passengers off an efficient metro.

That distinction matters enormously, because it is where policy actually bites. Autonomous driving is a technology, nothing more; its societal impact is decided by how we choose to deploy it.

A private autonomous car carrying one passenger still generates congestion: it simply removes the driver from it. A shared autonomous vehicle feeding passengers into a railway station strengthens public transport. An autonomous shuttle can serve places where a conventional bus route will never pay for itself. An autonomous delivery vehicle can take inefficient logistics trips off the road altogether. Same technology, four completely different outcomes for the transport system.

8. Private L4 is coming, which is exactly why the shaping has to happen locally

If L4 really does arrive in ordinary private cars as well as in professional or public fleets, the mix will not settle itself. Shared L4 integrated into public transport? Private L4 in everyone’s garage? Some combination of the two, and if so, on which roads, in which areas and at what times? None of these are technical questions, and none of them will be answered by the vehicle developers.

My conviction is that those answers have to be given locally. A canton or a city knows its own mobility system, travel demand, spatial constraints and public transport network far better than any authority above it, and therefore knows which of those futures would strengthen its territory and which would simply fill its streets with empty seats.

And local approval is not a rubber stamp, it is a design tool. By deciding where a driverless service may operate and which type of vehicle is allowed to operate there, a local authority is in practice deciding what its mobility system will look like: shared shuttles feeding a railway station in one area, on-demand ridepooling in a low-density valley in another, autonomous delivery on industrial roads, and perhaps no private L4 at all on a congested city axis at peak hours. The operational domain is the instrument; the shape of the mobility system is the result.

Switzerland’s case is a good example. With the automated driving ordinance of March 2025, it is the cantons that authorise driverless operation on their own territory, route by route, area by area, service by service. That makes them considerably more than a licensing desk: it hands them a real lever over what autonomous mobility ends up contributing to the transport system. If the technology keeps advancing at the pace we saw in China, that lever will have to be used sooner than most people expect.

9. Even China has not solved everything

In Europe, anyone who combines modes ends up with a phone full of apps: one for the train, one for the city network, one for ridehailing, one for the shared bike, each with its own account and its own payment method. A single interface for planning, booking and paying across all of them is one of our most stubborn unsolved problems. Given how much of daily life in China runs through one or two super-apps, we half expected to find it solved there, one platform through which you could book any vehicle, autonomous or not.

There is no such thing. Each provider still runs its own app, and commercial competition and platform ownership stand in the way of one universal interface, exactly as they do here.

Pick-up and drop-off points are also a critical issue for autonomous mobility in China. Most robotaxis work with designated digital pick-up and drop-off points, guiding passengers to virtual stations where the vehicle can stop safely, rather than allowing stops anywhere.

But some of those points were distinctly clumsy in practice. We were set down at spots with no proper access to the pavement, which meant stepping out into the carriageway or walking along the kerb to find a way up. It sounds like a detail, but it is the difference between a service that feels convenient and one that feels awkward, and it has to be designed into the street, not into the vehicle.

These are not footnotes. Pick-up and drop-off management, ticketing, digital integration and intermodality will largely decide whether automated vehicles end up as isolated mobility products or as a real component of an integrated transport system. China is still working on that, which means we are not as far behind as it sometimes feels.

10. We went expecting secrecy and found openness

Technology aside, what left the deepest impression was the openness. Every company and every authority we met answered our questions, including the uncomfortable ones and the things that had simply not worked. The conversation about Beijing’s roadside infrastructure was the clearest example. Instead of defending an earlier strategy, the authorities simply said that technological progress had invalidated the assumptions behind it. I am not sure how often we would hear that in Europe.

That openness also corrected a perception I had brought with me. In Europe, Chinese automated vehicles are frequently discussed as instruments of data collection, cars that come to watch us. That is not what I saw. These are commercial operations trying to build a viable business, and the clearest indication came from the regulatory side: the authorities require their own dedicated on-board unit in the vehicle precisely because they do not have a direct pipe into the operator’s systems. If the data flowed to the government anyway, that unit would be pointless.

What I met instead was a remarkable engineering culture: brilliant people, working extraordinarily hard, who mostly want to expand internationally, despite the geopolitical tension and the trust deficit we Europeans bring to the conversation. We are entitled to set our own conditions on data protection and oversight, and we should. But we should set them on the basis of what these companies actually do, not on assumptions we have never tested.

There is also an unmistakable bias towards implementation. Technology is tested, then industrialised, then deployed, then measured, then improved, and the cycle runs fast enough that Beijing’s infrastructure strategy from 2020 can already be openly described as obsolete.

Guangzhou alone reports more than 1’200 test and demonstration vehicles across passenger mobility, urban transport, campuses, logistics and vehicle-road coordination.

Europe should not copy China’s approach. Our institutions, transport systems, privacy frameworks and societal objectives are different, and in several respects deliberately so. But we can learn a great deal from its ability to learn from the experimentation phase and quickly get into implementation.

What autonomous mobility in China means for Europe and Switzerland

After a week of robotaxis, automated buses, logistics vehicles, OEMs, AI companies, monitoring centres and long conversations with authorities, my main conclusion is almost embarrassingly simple. The decisive question is no longer whether autonomous driving works. It is:

What do we want to do with it?

China is not showing us one future, it is showing us four at once: large robotaxi fleets; autonomous logistics working quietly in the background; shared autonomous shuttles; and, probably the most disruptive of them, millions of increasingly automated private cars that turn travel time into usable time. None of these futures excludes the others, and that is exactly why the decisions we take now matter so much. Whichever mix we end up with will be the result of choices, not of technology.

For SAAM, the objective was never simply to bring autonomous vehicles to Switzerland. It is to make sure that when they arrive, they make mobility safer, more accessible, more efficient and more sustainable than it is today. Autonomous mobility in China has shown what becomes possible once the technology reaches scale. It is now up to Switzerland to decide what we want that technology to do for the transport system we are trying to build.

A final personal note

Beyond the technology, this was simply an extraordinary week.

Experiencing these systems yourself changes the conversation completely; you cannot get the same understanding from a presentation or a report. And doing it alongside a group of mobility professionals, arguing about what we had just seen after every ride and every visit, was arguably as valuable as the visits themselves.

A sincere thank you to Espaces-Mobilités, Mobility Masterclass, Xavier Tackoen, Dario Deserranno, Quentin Colombier, CCI France Chine, all the organisations that opened their doors to us.

And, of course, all the participants who made the exchanges so rich: Safa Alkateb, Tim Asperges, Hugo Bosc Ducros, Wernher Brucks, Irene Campione, Samira Chadli, Filip Delannoy, Thierry Devresse, Nathalie Duplex, Mireille Fauconnier, Sylvain Guillaume-Gentil, Taina Haapamaki, Thibault Lecuyer, Armoghan Mohammed, Wiebke Muller, Oliver Nahon, Cedric Nzuzi Bunyezi, Eileen Pace, Judith Schuermans, Nicolas Talpe, Daniel Vargas

I went to China wanting to understand how far autonomous mobility had progressed. I came back thinking about something else entirely: what happens after it succeeds.

---

Follow us on LinkedIn to stay up-to-date with SAAM news!