Which robotaxi is safest?

Verdict up front

The Waymo Driver, riding on the Waymo Driver 6th generation platform, is the only fully driverless robotaxi with a published, peer-reviewed safety record measured in the tens of millions of paid, rider-only miles. Every other operator either (a) has orders of magnitude fewer autonomous miles on the road, (b) still has a human safety driver most of the time, or (c) has been through a serious safety-driven wind-down. In August 2026, safest robotaxi is Waymo, by a very wide margin, and it is not particularly close.

That answer comes with an important asterisk that most rankings miss. Waymo is safest in the specific sense that matters to a rider: fewer injury crashes, fewer serious-injury crashes, and far fewer at-fault collisions per mile than a typical human driver on the same streets. It is not safest by every metric a regulator might chart. Waymo has been through a fleet-wide freeway recall as recently as June 2026 (recall filing) and briefly suspended freeway service in four cities in May 2026 to retrain construction-zone behavior. Being the safest robotaxi today still means shipping software that occasionally has to be pulled back.

---

What "safe" actually measures

There are three metrics people confuse.

Miles per disengagement, from the California DMV Autonomous Vehicle Disengagement Reports, is a self-reported count of how often a safety driver had to take over during permitted testing. Every California AV testing permit-holder has to file one every February for the prior calendar year. It is the most-cited "safety" statistic on the internet, and it is also the most misleading one. The reports cover testing miles with a safety driver in the seat, not commercial driverless miles. Once an operator moves to rider-only commercial service under a CPUC deployment permit, most of its miles disappear from that particular filing entirely.

Crash rate per million miles, benchmarked against a human-driver baseline for the same city and road type, is what Waymo publishes in Traffic Injury Prevention and on its Safety Impact page. It is the metric that maps to "am I likelier to get hurt in this thing than in an Uber." NHTSA's Standing General Order 2021-01 requires every ADS operator to report every crash within one day, which is where cross-operator comparisons actually become possible.

Recall count and severity, the NHTSA recall database, is the third leg. A recall is not a crash; it is a filing that acknowledges a safety defect and commits to a fix. High-mileage operators accumulate more recalls simply because they are on the road; low-mileage operators can look pristine right up until they aren't.

A defensible ranking of "safest" has to weight all three, and it has to weight them by exposure. A robotaxi that has driven 400,000 miles without incident tells you almost nothing; a robotaxi that has driven 70 million rider-only miles and can quote a peer-reviewed injury rate tells you a great deal.

---

Why Waymo wins

Waymo's advantage is exposure combined with disclosure. As of its 2025 safety updates, the company has published rider-only mileage totals into the tens of millions of miles across Phoenix, San Francisco, Los Angeles, Austin, and expansion markets, with the flagship Waymo San Francisco marquee running roughly 500,000 weekly rides through mid-2026. Against a human-driver crash benchmark for those exact street networks, the Waymo Driver shows large reductions in airbag-deploying, injury-reported, and police-reported crashes, published in peer-reviewed venues rather than press releases alone.

On the California disengagement metric, Waymo has led every filed report since 2020, with miles-per-disengagement figures in the tens of thousands on the years it still ran meaningful California testing miles. The methodological caveat cuts both ways: the number is self-reported and easy to game by declining to file "disengagements" for things that arguably should count, but Waymo has been the most-scrutinized filer in the program and independent audits have not found the sort of category-shifting other operators have been accused of.

Insurance and litigation data, tracked in Waymo's Swiss Re collaboration, reports property damage and bodily injury claim frequencies well below human-driver benchmarks per million miles. It is the closest thing to an actuarial verdict the industry has produced. See also the DEPLOY explainers on where Waymo operates, Waymo fatal crashes, and what happens if a Waymo crashes.

---

Runners-up, ranked

1. Zoox

The Zoox Robotaxi is the only other purpose-built driverless robotaxi cleared for commercial paid service in the United States as of August 2026, after NHTSA granted the first-ever commercial FMVSS exemption for a purpose-built robotaxi on July 30, 2026, authorizing up to 2,500 vehicles per year for two years. On per-vehicle miles it is a fraction of Waymo, and its fleet-wide software recall of all 105 vehicles for smoke-detection failure on July 17, 2026 (NHTSA 26E044) is the sort of edge-case failure that shows up when driverless miles start scaling. Zoox is the credible number-two, but not yet number-one on exposure.

2. Apollo Go (Baidu)

Baidu's Apollo RT6, running the Apollo ADFM foundation stack, powers the Apollo Go Wuhan fleet, the largest single-city robotaxi deployment on Earth: 1,000+ vehicles, 24/7 fully driverless. On raw commercial mileage in a single city, only Waymo is at a comparable scale. Apollo Go's safety story is complicated by the March 31, 2026 fleet-wide system failure, the first publicly-documented outage of that kind at a commercial robotaxi, and by the fact that Chinese domestic crash and disengagement disclosure is not directly comparable to NHTSA SGO filings.

3. WeRide

WeRide runs the Robotaxi GXR on the WeRide One foundation stack, with commercial deployments in Guangzhou, Beijing, Abu Dhabi, Riyadh, and Dubai. WeRide is one of the very few operators to have been granted L4 permits across four regulatory regimes, an implicit safety endorsement even in the absence of consolidated per-mile data.

4. Pony AI

Pony AI runs the Pony Gen-7 Robotaxi on its PonyWorld virtual driver. The Toyota bZ4X robotaxi joint venture and the Uber partnership point to a maturing operational base. The Beijing test-vehicle fire of May 14, 2025 was contained by the emergency-stop procedure with no injuries, which is a passable outcome for a lithium-ion incident but not a proof of freeway-scale safety. See the Chinese robotaxi comparison for context.

5. Motional

Motional and the Motional IONIQ 5 Robotaxi at Las Vegas retain a strong patent estate (see risk navigation and remote assistance) but have scaled back commercial ambition, and the Motional Driver stack sits at claimed rather than verified maturity today. Fewer commercial miles, fewer publicly-reported incidents; the sample is too small for a per-mile safety verdict.

---

What beats it but does not count

Cruise's pre-ban disengagement numbers

Before October 2023, Cruise filed California disengagement reports that on paper matched Waymo's, and its Cruise AV fleet in San Francisco briefly ran fully driverless commercial rides. The October 2, 2023 pedestrian-dragging incident, the California DMV permit suspension, and the eventual GM decision to wind down the robotaxi program erased that record. The Cruise Origin purpose-built vehicle was cancelled before commercial deployment. Read what happened to Cruise for the full narrative. A safety record is not a snapshot; it is a running average that includes the last bad day.

Tesla Robotaxi's marketing claims

The Tesla Cybercab is at research maturity with no public rider-only mileage at all. The Tesla Robotaxi (Model Y ADS) pilot in Austin launched June 22, 2025 and, through August 2026, runs with a Tesla safety monitor in the front passenger seat with an emergency-stop control. That is not the same product as fully driverless service, and it disqualifies the comparison until Tesla publishes rider-only mileage under NHTSA SGO 2021-01. The DEPLOY explainers on how safe is Tesla Robotaxi, Waymo vs Tesla Robotaxi safety, and Tesla Robotaxi vs Waymo work through why the foundation model for robotics architecture differences (camera-only end-to-end vs sensor-fusion with lidar and HD map) matter to how the numbers should be read.

Small-fleet operators with pristine records

The Avride Robotaxi, Didi Robotaxi, MoveEz Robotaxi, and Momenta Robotaxi all have per-mile safety records that look clean, in the sense that no serious incident is publicly on file. The exposure is far too low to draw a conclusion. A fleet that has driven a few thousand paid miles and one that has driven 70 million paid miles are not comparable on any per-mile metric.

---

Gotchas that trip up "safest robotaxi" answers

Rider-only miles are usually invisible in the CA DMV disengagement report. Once an operator moves off a testing permit onto a CPUC deployment permit, the miles a rider actually pays for stop counting toward miles-per-disengagement. The most-cited "safety" chart is measuring the wrong thing.

Disengagement is self-classified. Operators decide which take-overs are disengagements and which are "planned tests." Two operators with equal actual safety can look very different in the filing.

NHTSA SGO reports include secondary collisions. The 2026 Pico-Union crashes in Los Angeles (one, two, three) and the Las Vegas rear-end will appear against Waymo in the raw crash count even though the vehicle was struck. At-fault classification, not raw count, is the metric that matters.

Regulatory action is not the same as unsafe operation. 83 parking citations in Austin is a curb-behavior problem, not a crash-safety problem, and conflating them is a common mistake in "worst robotaxi" content.

A recall is a feature, not a bug. Waymo's June 2026 freeway recall of nearly 3,800 vehicles and Zoox's July 2026 smoke-detection recall both mean the operator noticed a defect and filed the fix. Silence is not safety. If a fleet with tens of millions of miles has zero recalls on file, the more likely explanation is that it has not yet noticed the defects. See also the DEPLOY treatment of is Waymo actually driverless.

Bottom line: safest robotaxi in 2026 is Waymo, because it is the only operator with tens of millions of rider-only miles, peer-reviewed injury-rate publication, and a benchmark that shows fewer serious-injury crashes than a human on the same streets. Zoox is the runner-up on regulatory posture; Apollo Go on raw mileage; the rest are too small to rank on safety alone.

Sources: California DMV Autonomous Vehicles, NHTSA SGO 2021-01, Waymo Safety Impact, Swiss Re x Waymo, NHTSA recall lookup.