What actually happens when a robotaxi crashes?
Every crash on the federal record. Most robotaxi coverage quotes a total. The total is not the story — the distribution is.
Every outcome, one mark per report
Severity is the highest injury alleged by any party, as filed by the operator. It is not an adjudication and no fault is assigned anywhere in this dataset.
What each party was doing at the moment of impact
In 50.5% of reports the robotaxi was stopped or parked when it was struck. Any comparison of raw crash counts between operators, or against human drivers, that ignores this is comparing different kinds of event.
Why this is not a finding of fault
These are two coded fields describing movement immediately before impact. A stationary vehicle can still have created the conflict — by stopping abruptly, by stopping somewhere it should not have, or by blocking a lane.
The federal record contains no determination of fault. What it contains is what each party was doing.
Which end of the robotaxi got hit
The two sides are near mirror images. The robotaxi is struck at the back and the other vehicle makes contact with its front, which is the signature of a rear-end collision. Counting only the reports where that pattern is unambiguous, the robotaxi was rear-ended 450 times and ran into the back of somebody 90 times — a ratio of 5.0 to one.
Share of each operator’s reports that are unambiguous rear-end impacts, with the counts alongside. Operators below twenty reports are left out; at that volume the share is one or two events.
Why being rear-ended is not the same as being blameless
A car that brakes harder than the driver behind expected gets rear-ended, and the contact zones look identical whether the braking was necessary or not. This is the central criticism of autonomous vehicles in traffic — that they stop for things a human would have driven through — and the geometry cannot distinguish it from a following driver who was simply too close.
The federal record contains no determination of fault, and no field for it. It contains which panels of each vehicle made contact, which is a fact about the crash rather than a judgement about it.
Where crashes happen
The roadway type named on every filed report. An intersection is a small share of road surface and a large share of these crashes.
28% of reports place the crash at an intersection. Weather is stated on 1,214 of 3,446 reports — Clear 983, Cloudy 174, Rain 46. A blank is an unanswered field, not fair weather.
The same vehicles, again
Every filing carries a stable identifier for the physical vehicle involved, so the record can be read per car rather than per crash. It identifies a vehicle and nothing else — no person, no place.
2,251 distinct vehicles account for 3,439 reports. 825 of them appear more than once, and between them they carry 2,013 reports — 59% of the record. The busiest single vehicle appears 9 times.
A high count is not by itself a bad vehicle. A car that works more hours is exposed to more incidents, and this file carries no per-vehicle mileage to divide by — so this is a distribution, not a ranking of which car is worst. What it does show is that crashes are not spread evenly across a fleet.
The California file
12,149 incident records filed with California regulators across 6 quarters, 2025Q1 to 2026Q2. Every row carries its own flags for what was struck, how badly anyone was hurt, and what was complained about. It is a state record and does not move with the operator and state filters above.
Who or what the robotaxi hit
One mark per collision record. A record names every party involved, so a car struck with a passenger aboard sets two of these lanes.
41 of the 1,453 collision records name a cyclist or scooter rider, 4 a pedestrian and 1 a rail vehicle. 10 of the cyclist records and 301 of the other-vehicle records carry the pick-up or drop-off flag.
How badly anyone was hurt
Outcome flags across the 1,453 collision records. A record can carry two — a possible injury and a minor one — so the lanes total more than the 119 records that name any outcome at all.
1,334 collision records name no injury and no damage at all. At the other end, 3 severe injuries and 1 fatality. The severe and fatal lanes are short, not blank.
What riders and the public complained about
2,425 complaint records, each filed under exactly one heading.
| Safety | 1,141 | |
|---|---|---|
| Other | 938 | |
| Pick-up or drop-off | 326 | |
| Customer service | 15 | |
| Accessibility | 4 | |
| Wheelchair-accessible service | 1 | |
| No rows match this filter. | ||
The file, by carrier
One row per carrier that put its name on a California filing.
| Waymo | 12,124 | 1,438 | 8,271 | 2,415 | 321 | 3 | 1 |
|---|---|---|---|---|---|---|---|
| Zoox | 25 | 15 | 0 | 10 | 0 | 0 | |
| No rows match this filter. | |||||||
When the collisions happened
Collisions by hour of day, from the years California asked for a quarterly total rather than a row per event. 18 of 21 filings published the histogram, covering 2022Q1 to 2024Q4.
The busiest hour is 16:00. The nine hours from 21:00 to 06:00 are 38% of the day and hold 93 of 313 collisions, 30% — so the night is under-represented, and not by much. The shape follows service hours far more than it follows darkness.
This column exists nowhere else. The per-incident file that replaced this era carries a time of incident on every row and redacts it on every row, so the hour a robotaxi crashed is public only for the quarters filed this way. Read the shape, not the level: these are counts from a smaller fleet in earlier years, and there is no exposure figure by hour to divide them by.