Bloomberg reported on April 29, with The Verge picking it up the same day, that China has suspended new licenses for autonomous vehicles after dozens of Baidu's Apollo Go robotaxis ground to a halt in Wuhan traffic last month, creating chaos. The new restrictions block any company from adding driverless cars to existing fleets, expanding into new cities, or starting new test projects. There is no announced timeline for when licenses will start being issued again. The Verge notes this is at least the second time Chinese regulators have intervened after a Baidu-related incident; Baidu's Wuhan operations remain on pause while local authorities investigate.
"Dozens froze in traffic" is the failure mode that matters more than the regulatory response. Most autonomous-vehicle planning systems handle the "individual vehicle gets stuck" case with a fallback to remote operator or roadside assistance. What appears to have happened in Wuhan is fleet-scale coordination failure — multiple vehicles in close proximity hitting the same logic state at the same time and stopping. That is a different category of failure from individual-vehicle malfunctions, and it is exactly the kind of failure that scales nonlinearly with deployment density. Once you have hundreds of robotaxis in one city, a single bug or upstream service degradation can stall a non-trivial fraction of the fleet simultaneously. Chinese regulators have apparently concluded the same thing — and have decided the right intervention is on new license issuance, not on operational requirements for existing fleets.
Two patterns matter. First, this is the highest-profile regulatory intervention against autonomous driving in 2026 so far, and it is China — the country with the most permissive robotaxi rollout regime through 2025. Beijing freezing licenses sends a market signal that robotaxi expansion will not be unilaterally upward, even in jurisdictions that were pushing fastest. Waymo, GM Cruise, and Tesla face the same fleet-scale-coordination-failure risk in any US city where they hit deployment density; the regulatory-response template Beijing just set will get cited the next time it happens in a US market. Second, Baidu's repeat-offender status matters. The Verge's "at least the second time" framing is consistent with reporting that Apollo Go has had multiple incidents over the past 12 months. Single-incident investigations are routine; pattern-of-incidents investigations are how operators lose permits permanently. Watch whether Wuhan is the limit or the precedent for further Baidu suspensions.
For builders in autonomous vehicles or any large-scale agentic deployment, three concrete things. First, fleet-scale coordination failures are a separate category from per-unit failure modes and need their own monitoring. If your system has shared state, shared services, or shared model versioning across many active agents, you can stall the entire fleet on one bug. Test for it. Second, regulatory response to AI-system deployment failures is becoming faster and more aggressive across jurisdictions. Beijing froze licenses — that is a policy lever US states have but rarely use; expect them to use it in 2026 if a domestic fleet incident reaches a similar magnitude. Third, the connection to today's Eka and JAL coverage is concrete: the robotics-deployment frontier is no longer a question of "can we build it" but "what happens when a fleet-scale system fails publicly." Eka's tabletop dexterity and JAL's airport humanoid are testing the frontier in low-stakes contexts. Robotaxis are testing it at high stakes, and the failure has now produced a regulatory cost.