Robotaxis were sold as a near term fix for human fallibility, a software upgrade for the street that would make crashes rarer and commutes smoother. Early operating data tells a more complicated story, where some fleets are already beating human drivers on key safety metrics while others are struggling just to match them, let alone replace millions of people behind the wheel. The gap between the hype and the numbers is now large enough that regulators, investors, and riders are all being forced to recalibrate expectations.
Instead of a clean handoff from human drivers to code, the first wave of robotaxis is exposing how messy that transition will be. Safety performance varies sharply between companies, the technology still stumbles in edge cases like school zones, and even the most advanced systems are learning in public. Replacing human drivers at scale is proving to be a long, uneven grind rather than an overnight software update.
Robotaxis meet the real world, and the numbers bite back
The most jarring reality check is coming from Tesla, which has pitched its Robotaxi vision as a software unlock on existing cars. According to Jan disclosures highlighted in its Q4 2025 earnings materials, a chart of cumulative Robotaxi miles shows the fleet has traveled approximately 50 million miles, yet its own Robotaxi crash rate data show performance roughly three times worse than human drivers even with a safety monitor in the car. Separate filings summarized by Jan commenters note that, according to figures submitted to the National Highway Traffic, Tesla Robotaxis were involved in nine crashes over a relatively small mileage base, roughly one incident every 55,000 miles. That aligns with a separate analysis of Tesla’s own Robotaxi crash rate, which cites one crash every and concludes the system is still far from matching human baselines.
Those numbers have not stopped markets from betting on the long term story. Around the time this safety data surfaced, Shares of the electric vehicle maker were up 2.1% at $425.26 even as S&P 500 and Dow Jones Industrial futures were down 0.5, a sign that investors are still treating the setbacks as growing pains rather than existential flaws. Yet even sympathetic analysts concede that, Overall, Tesla is not yet delivering the safety uplift that would justify removing human supervisors. That tension is sharpened by a Monday note that a Monday report from Elektrek found Tesla Robotaxis are crashing much more frequently than cars driven by humans just as the company moves to pull safety monitors, a decision critics argue is premature given the current crash rate.
Waymo’s safer record still collides with messy edge cases
On the other side of the ledger, Waymo is starting to show what a more mature robotaxi stack can do, at least on paper. Internal and third party analyses of the Waymo Driver suggest that, compared with human benchmarks, the system is INVOLVED SIGNIFICANTLY LESS, with one study citing an 85% reduction in certain crash types when measured against a carefully defined benchmark. A separate review of Waymo’s transparency reports notes that its autonomous fleet has logged 56.7 m miles in driverless mode and dramatically cut police reported collisions compared with human drivers in the same cities, reinforcing the idea that the technology can outperform people when it is deployed conservatively. Consulting work that digs into Waymo’s own safety impact, laid out in a detailed Table of Contents, stresses that When we think about the future of transportation, safety is often the first concern that comes to mind and that Aft er millions of miles the data now supports a real reduction in harm to vulnerable road users.
Yet even with that track record, Waymo is not immune to the hardest edge cases. Federal investigators are now examining how its vehicles behave around school buses and children after a crash in Austin where, in a statement, Waymo said a child “suddenly entered the roadway from behind a tall SUV” and that its vehicle immediately braked. Safety experts like Phil Koopman have questioned whether that Robotaxi crash was really unavoidable, pointing to NTSB scrutiny of how the system handled stopped school buses and whether human drivers in those situations would have behaved differently, as outlined in a Jan analysis that quotes Waymo saying There have been no injuries in similar events with human drivers in those situations. Even sympathetic commentators concede that while Waymo reports more dramatic safety improvements than rivals, as one Waymo focused essay notes, the company is still working through how its perception and planning stack handles rare but high stakes scenarios in real world conditions.
Benchmarks, disengagements, and the limits of lab-style metrics
Part of the confusion around robotaxi safety comes from how performance is measured. Waymo and its research partners have invested heavily in constructing a rigorous HDV benchmark, short for “latest generation human driven vehicles,” that compares autonomous performance not to an average driver in an old sedan but to people using newer, technologically advanced cars with modern safety features. Another industry analysis of the Waymo Driver stresses that its crash reductions are calculated against human benchmark rates, which is crucial context when interpreting claims of an 85% reduction or 6.8 times fewer certain crash types. These carefully constructed baselines matter because they determine whether a robotaxi is truly safer than the best human drivers on the road or only looks good compared with a national average that includes drunk and distracted motorists.
At the same time, traditional testing metrics like disengagements per mile are starting to show diminishing returns as a proxy for real world safety. A comprehensive academic review of autonomous vehicle performance cites a rate of 0.0028 disengagements per mile, meaning that for every mile driven there are approximately 0.0028 instances where a human safety driver has to take over. That looks impressive on paper, but as fleets move from test loops to dense urban service, the rare interventions that remain tend to cluster around complex, high risk situations like unprotected left turns or interactions with emergency vehicles. Waymo’s own research on whether autonomous vehicles outperform latest generation human driven vehicles over 25 million miles underscores that its benchmark methods are designed to capture those rare but severe events, yet even that work acknowledges that safety estimates will evolve as fleets continue to scale and develop.
Regulators, roads, and the slowdown in testing
While the technical debate plays out, regulators and city officials are quietly reshaping the environment in which robotaxis operate. In California, the epicenter of early autonomous testing, total reported autonomous miles fell sharply as scrutiny increased. One analysis notes that autonomous vehicle testing in the state dropped 50% year over year, with total logged distance shrinking to just 552,895 miles, a reversal detailed in a report that explains Here why companies pulled back. A related summary of the same trend, which credits Kirsten Korosec and notes the 4:16 PM PST timestamp alongside Image Credits for Image Credits photographer David Paul Morris at Bloomberg, underscores how local politics and public backlash can slow deployment regardless of what the safety data says. As robotaxis move from pilot projects to real services, every high profile crash or traffic jam involving a stalled vehicle becomes fodder for hearings and moratorium proposals.
That regulatory friction is not limited to the United States. A global review of the sector notes that, Yet in recent years, Chinese players have rapidly closed the gap with American rivals and that Since 2023, Chinese and American companies have actively explored Robotaxi commercialization, but their approaches diverge dramatically. In China, city governments have in some cases carved out dedicated zones and more permissive rules for robotaxis, while in U.S. hubs like San Francisco and Austin, services are often subject to overlapping state and municipal oversight. That patchwork means the same technology can look ready for prime time in one jurisdiction and stalled in another, complicating any simple narrative about when human drivers will be replaced.
More from Morning Overview