The edge case that never was: why robotaxis still can't handle an emergency scene

There is a document trail that should worry every autonomous vehicle operator in America, and it isn't held by a regulator or a rival. It's held by the San Francisco Fire Department.

Since spring 2022, San Francisco firefighters have been instructed to file a report any time an autonomous vehicle changed what they would otherwise have done in the field. The resulting log — at least 85 interference incidents between January 2023 and early 2024, with roughly 25 in August 2023 alone — is the closest thing the industry has to a public intervention dataset for emergency scenes. The companies themselves have never published one.

That log is now shaping federal policy. On 8 July, NHTSA administrator Jonathan Morrison wrote to AV developers citing vehicles that drove into active emergency scenes, blocked ambulances and fire engines, and failed to recognise flashing lights, flares, smoke and traffic cones. His framing matters more than his examples.

> [QUOTE] "The inability to detect and appropriately respond to such situations represents a functional insufficiency." — Jonathan Morrison, NHTSA Administrator

Functional insufficiency, not software anomaly. Emergency scenes, Morrison argued, are not rare or extreme edge cases — they are a routine feature of the urban environments these vehicles were certified to operate in. Companies were given until the end of July to explain their fixes.

What the incident reports actually describe

The failure modes in the fire department's records are specific and repetitive. Vehicles stopping perpendicular across narrow streets. Vehicles that customer-service teleoperators cannot remotely reposition, leaving firefighters to climb into the driver's seat mid-callout — in one December incident, a Waymo en route to a stabbing response reportedly drove forward when placed in reverse. Robotaxis picking up and dropping off passengers directly outside fire stations and ambulance depots, a pattern fire chief Dean Crispen singled out publicly last week.

Then there is 4 July 2026: dozens of Waymos immobilised in holiday traffic, several depleting their batteries where they stood, blocking streets they could not clear.

None of these are perception problems at the frontier of machine learning. They are operational design decisions — where to stage vehicles, how quickly remote operators can intervene, what a vehicle does when its planner runs out of options. Which is precisely why the "edge case" defence has expired.

The legislative response

Representative Kevin Mullin's AV Emergency Response Coordination Act would direct NHTSA to set minimum national emergency-response standards, require published protocols for first responders, and mandate a 24-hour hotline giving public officials direct access to operators. Locally, San Francisco supervisor Connie Chan is drafting an ordinance to fine companies whose vehicles obstruct responders, expected before the Board of Supervisors within months.

Note what is being legislated: not perception benchmarks, but disclosure and accountability infrastructure. Hotlines. Protocols. Fines that create a paper trail.

The lesson for the rest of the sector

Every autonomy vertical — drones, maritime, warehouse AMRs — should read this as a case study in what happens when an industry declines to publish its own intervention data. The information vacuum does not stay empty. It gets filled by whoever is keeping records, and the party keeping records here is the fire brigade.

Waymo operates driverless in eleven US cities with four more planned. At that scale, emergency-scene interactions are a daily statistical certainty. The operators who thrive under the coming rules will be the ones who measured this failure mode before a firefighter measured it for them.