Are robotaxis safe? What the data says
It's the first question everyone asks, and the honest answer needs care. Here is what each operator and regulator actually publishes — with the fine print that usually gets dropped.
Read this first: you can't line these up
Every figure below uses a different yardstick — different miles, different definitions of a crash, different roads, self-reported versus regulator-collected. Treat each as its own lens with its denominator attached. Even the regulators say their own numbers aren't built to rank companies. Anyone handing you a single “robotaxis are X times safer” headline is flattening all of that.
Waymo — the most rigorous data
Waymo publishes the most detailed public safety data, and independent peer-reviewed studies back the core of it. As of March 2026 it reports 220.6 million fully-driverless miles. Against human drivers in the same areas over the same period, the reductions are large:
Waymo vs. human drivers, same cities and period (peer-reviewed):
The fine print
Waymo picks the benchmark and only counts its own operating area — a fair, location-matched comparison, but still the operator grading its own homework. The peer-reviewed papers (at 56.7M and 7.1M miles) are the sturdiest version to rely on.
California DMV — the regulator's log
California makes permit-holders file every disengagement and every collision. As of April 2026, 978 AV collision reports had been filed; permit-holders logged over 9 million test miles in the year to November 2025.
The fine print
Two catches. This covers only testing-permit miles — the paid, fully-driverless rides run under a different agency (the CPUC) and aren't in this dataset. And the DMV itself states these reports “are not designed for comparative analysis across companies.” The disengagement metric is being retired and replaced.
NHTSA — the national crash order
Federally, operators must report any crash where the automated system was engaged within 30 seconds of it — a nationwide log, refreshed monthly.
The fine print
It's a count, not a rate: there's no mileage attached, it's self-reported, and completeness varies by operator. NHTSA cautions that raw counts shouldn't be used to compare manufacturers — a bigger fleet with better crash-detection simply reports more. The large per-operator totals that circulate online are third-party tallies of the raw file, not official figures.
Tesla — a different measurement entirely
Tesla publishes a Vehicle Safety Report: for Q4 2025, one crash per 5.39 million miles with Autopilot engaged, against roughly one per 702,000 miles for the US average.
The fine print
This is supervised driving — a human ready to take over — on Tesla's own chosen miles and crash threshold, so it can't be compared with driverless robotaxi data. Tesla publishes no separate safety rate for its actual Robotaxi service.
Chinese operators — self-reported milestones
Baidu says its Apollo Go fleet has driven 200+ million kilometers with one airbag-deployment incident per roughly 10.1 million km and no reported major-injury accidents; Pony.ai cites 50+ million kilometers.
The fine print
These are company statements to press and investors, given in kilometers, with no independent regulator dataset behind them like the US has. Treat them as milestones, not verified rates — the transparency is lower.
The bottom line
Waymo has by far the most rigorous, peer-reviewed safety data, and it shows large injury-crash reductions in the places it drives. Everyone else publishes less, on different terms. That's the honest state of it in 2026 — real, encouraging evidence where it exists, and a lot of numbers that aren't the same thing.
Saw something yourself? Report an incident →