Waymo Suspends Service in Four Cities After Robots Keep Driving Through Floods

Waymo robotaxis continued driving through flooded streets during a period of severe weather — and that failure forced service suspensions across four cities. The incident is one of the most significant operational safety failures for an autonomous vehicle company operating at commercial scale.

What Happened

During periods of heavy rainfall, Waymo vehicles either failed to detect flooded road sections or detected them and proceeded anyway. Passengers were transported through conditions that posed clear safety risks, and the behavior was repeated across multiple vehicles and locations. Waymo responded by halting service in affected cities while engineering teams investigated the sensor and decision-making failures.

Autonomous vehicle systems are generally well-tested in moderate rain conditions. But compound weather scenarios — flooded roadways combined with poor visibility, debris, and rapidly changing road surfaces — remain a documented weak point in current AV systems.

Regulatory Implications

The incident is likely to accelerate NHTSA review of autonomous vehicle operational safety requirements, particularly around Operational Design Domain (ODD) definitions. ODD specifies the conditions under which a self-driving system is certified to operate. Current ODD frameworks may be insufficiently precise about extreme weather thresholds, allowing vehicles to operate in conditions their systems cannot reliably handle.

For Waymo, Cruise, Zoox, and other commercial AV operators, this incident could mean stricter weather-condition verification requirements before market expansion.

The Commercialization Reality Check

Waymo operates what is widely regarded as the most advanced commercial robotaxi service in the world. That makes this failure notable: if the industry leader has a systematic gap in extreme weather response, the gap is an industry-wide problem. Full autonomy in all real-world conditions — including the ones that don’t appear in training data — remains the field’s hardest open problem.