100 Baidu Robotaxis Stall Mid-Traffic in Wuhan: The Correlated Failure Problem
Just before 9 PM on March 31, 2026, more than 100 Baidu Apollo Go driverless taxis came to an abrupt stop across Wuhan, China. They halted at intersections, on elevated ring roads, and in the fast lanes of busy arterials — simultaneously. Passengers could not exit unassisted. Police and Baidu staff mobilized to help, and no injuries were reported. But the incident exposed a fundamental vulnerability in large-scale autonomous vehicle deployment: the correlated failure.
What Happened
Wuhan police issued a statement describing the event as a "system malfunction" affecting Apollo Go's fleet management platform. The simultaneity of the failures rules out sensor errors or isolated mechanical issues. Every affected vehicle received the same bad signal — or lost connectivity to the same backend — at the same moment.
Videos posted to Chinese social media showed dozens of white Apollo Go vehicles frozen in place, hazard lights blinking, blocking multiple lanes. Passengers pressed in-car emergency buttons with mixed results. Many eventually had to wait for police or Baidu support staff to arrive and manually assist them out of the vehicles.
Some vehicles stalled on elevated highways, creating traffic jams that took hours to clear. The full disruption to Wuhan traffic lasted well into the early hours of April 1.
The Correlated Failure Problem
In traditional software engineering, a single backend outage might take down a website. In autonomous vehicle fleets, the same principle can freeze hundreds of cars in traffic simultaneously. This is called a correlated failure — and it is considered one of the hardest unsolved problems in large-scale AV deployment.
Autonomous vehicles rely on constant communication with cloud-based fleet management systems for route updates, perception model patches, and safety override signals. When that cloud connection degrades or sends a bad state, every vehicle on the same system update cycle is affected at once. Individual sensor failures are rare and isolated. Backend software failures can be instantaneous and fleet-wide.
Waymo experienced a milder version of this in San Francisco in December 2025, when a cluster of robotaxis repeatedly circled a city block after receiving conflicting routing instructions. The Wuhan incident is an order of magnitude larger — and more dangerous, given that it occurred on elevated expressways at night.
Timeline of Major AV Incidents
| Date | Company | Location | Incident | Vehicles Affected |
|---|---|---|---|---|
| Mar 31, 2026 | Baidu Apollo Go | Wuhan, China | Fleet-wide stall, correlated system failure | 100+ |
| Dec 2025 | Waymo | San Francisco, USA | Cluster looping due to conflicting routing | ~12 |
| Aug 2025 | Baidu Apollo Go | Chongqing, China | Vehicle drove into construction pit | 1 |
| Jun 2024 | Cruise (GM) | San Francisco, USA | Vehicle dragged pedestrian after accident | 1 |
What Baidu Said
Baidu has not released a detailed technical post-mortem. The company stated it is "actively investigating" the root cause and dispatched staff overnight to assist affected passengers. Wuhan police confirmed no criminal investigation is underway, classifying it as a technical malfunction.
Apollo Go currently operates in Wuhan, Beijing, Shenzhen, Guangzhou, and Chongqing. The service has logged over 10 million driverless rides and has international expansion agreements with Uber and Lyft for London operations, as well as active deployments in the Middle East.
What This Means for Autonomous Vehicle Deployment
The Wuhan incident will likely accelerate regulatory pressure on AV operators in China to demonstrate "fail-operational" architectures — systems that continue to function safely even when the cloud backend is unreachable. Current regulatory frameworks in China require AVs to pull over safely when connectivity is lost, but the Wuhan vehicles appear to have stopped in active traffic lanes rather than executing a safe-stop maneuver.
For Waymo, which is aggressively expanding in Austin, Atlanta, and Tokyo, the incident is a useful contrast. Waymo has consistently emphasized hardware redundancy and onboard compute that allows safe-stop execution without cloud dependency. The company's incident rate per million miles is far lower than Apollo Go's, partly because Waymo operates in more geofenced, mapped environments.
Baidu's model prioritizes rapid fleet scaling and cost efficiency through cloud-heavy architecture — which enables faster deployment but increases correlated failure risk. The tension between scale and resilience is the central engineering challenge for the next phase of autonomous vehicle commercialization.