Which Driverless Cars Actually Drive Themselves?

California issues two separate testing permits, and the gap between them is the gap between research and a product. One lets a company test an autonomous vehicle with a trained human in the driver seat, ready to take over. The other lets it test with nobody there at all.

Twenty companies filed mileage for the year to 30 November 2024. Sixteen of them reported only the first kind. Four reported any of the second: Waymo, Zoox, WeRide and Nuro.

The volumes make the gap starker than the count does. Those four between them reported 552,896 driverless miles, and 516,572 of those were Waymo’s. WeRide managed 471. Nuro managed 169, which is about the distance from San Francisco to Sacramento and back.

This is why a chart of total miles flatters the field. Apple drove 152,924 miles in California that year, more than every company except Waymo, Zoox and Nuro, and never once took the driver out.

Waymo reported 2,389,565 supervised miles against 516,572 driverless ones. So even at the front of the field, 82 per cent of testing still had a human in the seat.

That is not a criticism; it is what testing is for. But it does answer the question people actually ask. The company furthest along still runs four supervised miles for every one it runs alone, and it has been at this since 2009.

The number that shows how far ahead Waymo is sits outside this chart entirely. Its commercial robotaxi service, which operates under a deployment permit rather than a testing one, carried passengers 16,579,725 miles across 4,220,075 trips in the single quarter to 30 June 2026. That is thirty times its entire year of testing mileage, and the state does not require it in these files.

These reports have been the public record of autonomous driving in America since 2014, and California is about to retire them. The DMV says its 2026 regulations will replace the disengagement reporting requirement with a different set of measures entirely.

What replaces it is more useful. The proposed rules ask for vehicle miles travelled during testing AND deployment, which would close the gap this page has had to work around: for the first time, commercial robotaxi mileage would sit in the same return as testing mileage. They also ask for crash reporting aligned to the federal standing order, vehicle immobilisations, and heavy braking events above a defined threshold.

That matters beyond bookkeeping. Disengagements measured how often a human rescued the car, which made sense when every car had a human in it. Most of the questions people now ask about driverless vehicles are about cars with nobody aboard, and a disengagement cannot happen in one.

So this chart is close to the last of its kind. It is a decade-long record of an industry learning to drive, ending at the point where the thing being measured stopped matching the thing being built.

The same files carry disengagement counts, and it is tempting to divide miles by disengagements and call the result a safety league table. We have not, because the DMV says plainly that the reports "are intended to provide insight into the operational activities of each permit holder and are not designed for comparative analysis across companies".

The reason is that companies define a disengagement differently, test in different places and in different conditions, and report it themselves. A company testing in heavy city traffic will look worse than one testing on an empty road, and neither number says anything about the product a passenger would ride in.

Miles driven is a different kind of figure. It is an activity count, it is reported the same way by everyone, and dividing it between companies says who is doing the work. That is all this page claims.

One more limit worth stating: these are the figures for the year to November 2024. The DMV announced in February 2026 that the following year’s total exceeded nine million miles, roughly double, but it has not published the per-company files for that year, and its own link to them points at a page that does not contain them.