San Francisco's Fourth of July fireworks turned into gridlock, and Waymo's robotaxis froze in place for hours, the second mass stalling in two years and a failure mode sensors cannot prevent.
Waymo's robotaxis did not stall on the night of July 4 because of a software bug or a sensor failure. They stalled because a fireworks show turned San Francisco's streets into a gridlocked maze the cars could not drive through.
National and local coverage over the weekend focused on the spectacle: cars blocking intersections, riders waiting hours for tows, one Waymo apparently driven through an erupting box of fireworks, another catching fire. The pictures made it look like another AV embarrassment. The actual mechanism points at a category of failure autonomous fleets keep running into, one the industry's standard driving tests are not designed to catch.
According to Business Insider's reporting, Waymo spokesperson Chris Bonelli attributed the disruptions to "severe traffic congestion" and said the company's roadside-assistance team worked with local authorities and emergency services to clear vehicles. Bonelli added that Waymo is "always evaluating ways to strengthen Waymo's resilience in major traffic disruptions." That phrasing is not about improving the driving model. It places the failure in the surrounding environment rather than in the cars.
The trigger was the city, not the cars. The same night, the Golden Gate Bridge fireworks show turned into a foggy, logistically chaotic mess, with unplanned road closures and traffic the city's event playbook could not absorb. Waymo's vehicles entered that environment, made choices their software was not designed for, and stopped. Multiple cars blocked intersections. Eyewitnesses documented a Waymo being towed out of the Presidio and asked publicly what was happening with the fleet. One X user described waiting three to four hours for a tow with a family, an account that has not been independently verified.
A separate SF Chronicle report adds a vehicle fire to the operational picture and compounds the question of what the cars did when road conditions collapsed around them.
Waymo has been here before. In 2024, a San Francisco blackout caused Waymo robotaxis to stall en masse, and the company's explanation then also pointed at the surrounding environment rather than the driving stack. Two years later the pattern has not changed in shape, only in cause. The 2024 stalling was triggered by an external utility failure. The 2026 stalling was triggered by a routine city event that spiraled into gridlock. Both produced the same result: cars that work in normal traffic and freeze when the city's logistics break down around them.
Sensor quality and motion planning have advanced to the point that AVs handle ordinary driving well. The gap is somewhere else. An autonomous fleet cannot yet improvise. A human driver faced with an unscheduled road closure, an intersection full of confused pedestrians, and a dispatcher who has gone silent will make judgment calls: pull into a side street, double-park briefly, accept an unsafe-but-mobile state to escape a worse one. A robotaxi operating under conservative safety policies will not. It will stop, ask for help, and wait.
July 4 is not a black-swan event. It is a known annual stress test, one of dozens a dense city runs every year, alongside marathons, bridge closures, protests, parades, and storm-driven gridlock. If robotaxis cannot distinguish between "drive normally" and "find a way out" when the city's logistics collapse, the ceiling is not perception. It is policy.
Bonelli's promise to evaluate resilience is a commitment that the next incident will be handled better. The structural question stays: a fleet that cannot improvise inside a city that frequently changes shape will, on a long enough timeline, stall again.