On Tuesday evening in Wuhan, more than 100 of Baidu’s Apollo Go robotaxis stopped moving. They did not pull over. They did not activate an emergency protocol. They simply froze, scattered across the city’s roads and elevated highways, some in the middle lane of ring roads with traffic streaming past on both sides. Passengers trapped inside called the police. Videos circulating on Weibo showed Apollo Go vehicles stranded at intersections, hazard lights blinking, going nowhere. One clip appeared to show the outage causing a highway collision, though Wuhan police said no injuries were reported and all passengers exited their vehicles safely.
The Wuhan traffic police said in a statement posted on Weibo that initial findings suggested a “system malfunction” had caused the vehicles to halt. The cause, they added, was still under further investigation. Baidu did not immediately respond to requests for comment from multiple news organisations.
Wuhan is not a small-scale test market. It is the largest deployment of Apollo Go’s fleet in China, with more than 1,000 driverless vehicles operating across the city. As of Baidu’s most recent earnings report in February, Apollo Go’s services had been deployed in 26 cities globally, with cumulative orders exceeding 20 million. In the fourth quarter of 2025 alone, the service delivered 3.4 million fully driverless rides, with weekly rides surpassing 300,000 during peak weeks. This is not an experimental programme. It is a commercial operation running at genuine scale, which makes the simultaneous failure of more than 100 vehicles in a single city something qualitatively different from an isolated incident.
Jack Stilgoe, professor of science and technology policy at University College London, told BBC News that while driverless technology may be safer on average than human drivers, this incident demonstrated it could “still go wrong in completely new ways.” The distinction matters. A human driver who suffers a medical episode creates one hazard. A fleet-wide system failure creates dozens or hundreds simultaneously. “If we’re going to make good choices about this technology, we need to understand entirely new types of risk,” Stilgoe said.
The 💜 of EU tech
The latest rumblings from the EU tech scene, a story from our wise ol’ founder Boris, and some questionable AI art. It’s free, every week, in your inbox. Sign up now!
The Wuhan incident is the most dramatic in a pattern of failures that has shadowed the robotaxi industry’s rapid expansion. In December 2025, a major power outage at a Pacific Gas and Electric substation in San Francisco knocked out electricity across roughly a third of the city, disabling traffic lights in several neighbourhoods. Waymo’s robotaxis, which treat dead traffic signals as four-way stops, began requesting confirmation checks from the company’s fleet response team. The concentrated spike in requests overwhelmed the system, and dozens of vehicles stopped in place, blocking streets and contributing to significant traffic jams. Waymo suspended its service and later shipped a software update to help its vehicles navigate disabled traffic infrastructure more decisively.
Three months before that, in August 2025, an Apollo Go robotaxi carrying a passenger in Chongqing drove into a construction pit. The vehicle bypassed barriers and warning signs before dropping into a trench, and the passenger, a woman, was rescued unharmed by nearby residents using a ladder. The incident highlighted a known vulnerability in autonomous driving systems: their difficulty detecting large, irregular road hazards that fall outside trained datasets.
The timing of the Wuhan outage is acutely uncomfortable for Baidu. In December 2025, Uber and Lyft both announced partnerships with the Chinese technology company to bring Apollo Go vehicles to UK roads, with trials initially targeting London. Uber’s pilot was expected to begin in the first half of 2026 using Apollo Go RT6 vehicles, purpose-built electric cars with detachable steering wheels. Lyft’s testing would start with a fleet of 50 vehicles, part of the operational rollout following its $197 million acquisition of the FreeNow mobility platform. Both companies still require regulatory approval before trials can begin, and the footage from Wuhan of frozen robotaxis blocking traffic is unlikely to accelerate those conversations.
Apollo Go has also expanded into the Middle East, launching fully autonomous ride-hailing services in Abu Dhabi with AutoGo and obtaining the first fully driverless testing permit in Dubai. In recent weeks, it began offering rides through the Uber app in Dubai. These are markets where the regulatory environment has been deliberately welcoming, positioning Gulf cities as global hubs for autonomous mobility. But a mass fleet failure in a flagship Chinese deployment raises questions that hospitality alone cannot answer.
The deeper issue is systemic. When autonomous vehicles operate as individual units, a software bug or sensor failure produces a single-point event. When they operate as centrally managed fleets, connected to shared backend systems and reliant on common infrastructure, they introduce a category of risk that has no analogue in conventional transport: correlated failure. Every vehicle fails in the same way, at the same time, for the same reason. The result is not a car accident. It is a traffic system event.
Regulators are watching. Tesla has faced investigations from US safety authorities over erratic robotaxi behaviour, and the broader industry debate over whether the robotaxi model was ever the right bet continues to sharpen. Waymo operates with safety drivers available remotely. Baidu’s Wuhan fleet runs fully driverless. The difference, in an outage scenario, is the difference between inconvenience and entrapment.
What Wuhan demonstrated on Tuesday was not that robotaxis are unsafe. Statistically, they may well be safer than human drivers across millions of miles. What it demonstrated is that autonomous fleets fail differently, at a scale and in a manner that existing regulatory frameworks, built around individual vehicle defects and driver error, were not designed to address. The question is no longer whether the technology works. It is whether the systems built to govern, deploy, and recover it can keep pace with the speed at which it is being rolled out across the world.
