The Fleet That Froze

A hundred robotaxis stopped thinking at the same time. That's not a bug report. That's a design review.

On the evening of March 31, Baidu's Apollo Go fleet — the largest commercial robotaxi service in China — suffered a simultaneous system failure across Wuhan. Over 100 autonomous vehicles froze mid-route on elevated highways and city streets, trapping passengers inside, blocking traffic lanes, and causing at least three collisions when human drivers encountered the stationary fleet.

Passengers reported the in-car SOS buttons returned "unavailable." Calls through the vehicle's built-in screen were automatically disconnected. The customer service hotline offered assurances that a technician was en route. None arrived for hours.

One passenger, stranded on Wuhan's Third Ring Road for nearly two hours, described "large trucks speeding past on both sides" while the car sat dead in the middle lane. Another was charged the full fare after being trapped for 90 minutes on an elevated highway. Police began receiving emergency calls around 9 PM and ultimately had to physically extract passengers from vehicles that their own safety systems had rendered inoperable.

The Architecture Problem

We've been testing autonomous vehicles as an individual reliability problem. Can this car navigate this intersection? Can this sensor detect this pedestrian? Can this algorithm handle this edge case? The testing frameworks, the safety certifications, the regulatory approvals — they all evaluate single-vehicle performance.

But Baidu didn't deploy individual vehicles. They deployed a fleet — centrally managed, network-dependent, software-updated as a unit. And a fleet connected to one system fails as a fleet. Not one car in a thousand having a bad day. A hundred cars, simultaneously, becoming obstacles.

Apollo Go's customer service representatives told local media the failure was caused by "network issues" triggering a "driving system abnormality." An industry insider described it as "unexpected circumstances that triggered a safety self-check mechanism." In other words: the system designed to protect passengers from unsafe driving conditions activated across the entire fleet and then couldn't communicate with anyone about what to do next.

The safety system worked exactly as designed. It just worked on all the cars at once, and nobody designed for that.

The Pattern

This reportedly mirrors a December 2025 incident where multiple Waymo vehicles in San Francisco entered "minimal risk condition" simultaneously during a power outage — pulling over and stopping, which sounds safe until you realize a dozen vehicles simultaneously stopping in a dense urban grid creates its own hazard.

The pattern is the same both times: individual-vehicle safety logic, applied to a centrally-managed fleet, produces fleet-level paralysis. The thing that makes one car safe makes a hundred cars dangerous.

Coherenceism calls this nested coherence — reliability at one scale creating fragility at the scale above it. Each vehicle is doing the right thing. The fleet is doing the catastrophic thing. And the architecture can't tell the difference because it was never designed to think at fleet scale.

As of publication, Baidu had not released an official statement. Service resumed the following day.

One hundred cars. One failure. Zero surprises — for anyone looking at the right level.

Sources:

• Apollo Go's robotaxi fleet suffers mass paralysis, stranding passengers on elevated highways — CarNewsChina, 2026-04-01
• Baidu Apollo Go robotaxis hit by rare mass paralysis in Chinese city — CnEVPost, 2026-04-01
• Baidu robotaxis reportedly halted mid-traffic causing crashes in Wuhan, China — CNBC, 2026-04-01