Waymo Robotaxis Crash 68% Less Than Humans. Just Not in Texas
Waymo’s robotaxis posted a substantially lower crash rate than human drivers across a new four-city study, though Austin stood out as the exception in the data. The Insurance Institute for Highway Safety analyzed 50 million miles of Waymo driving from 2021 through 2024 in San Francisco, Los Angeles, Phoenix, and Austin, then compared that record with 222 billion miles driven by humans in the same cities over the same period.
Across the full sample, Waymo vehicles had a 68 percent lower crash rate than human drivers. The IIHS data also showed the autonomous vehicles were involved in 85 percent fewer single-vehicle crashes and 81 percent fewer injury crashes than human-operated vehicles. The study focused on Waymo’s Level 4 vehicles in driverless operation and compared them with crash patterns from broadly similar deployment areas.
The city-level results were less uniform. In Los Angeles, Waymo vehicles recorded a 71 percent lower crash rate than human drivers. In Phoenix, the figure was 76 percent lower. San Francisco was closer, at 35 percent lower. Austin moved in the other direction, with Waymo vehicles showing a crash rate 4 percent higher than human drivers, which is partly attributed to the small sample size.
The study counted 736 reported crashes involving Waymo autonomous vehicles during the period analyzed. The IIHS determined that only 22 percent of those were likely to be “police-reportable,” and just 64 occurred while the vehicle was operating in autonomous mode. Of those autonomous-mode crashes, 31 happened with no occupant in the Waymo vehicle, while 25 involved property damage only.
The IIHS also cautioned that comparing robotaxi crash records with human-driver data remains difficult. Human crash totals are incomplete because many incidents never make it into police databases. Roughly half of all crashes and about one-third of all injuries are never reported to police, which limits the precision of any direct comparison between autonomous and human-driven fleets.
Even with that limitation, the IIHS said the findings supported a lower crash-involvement rate for Waymo’s driverless vehicles in the areas studied. The institute concluded that the study “provides further evidence that Waymo’s L4 vehicles in driverless operation have lower crash involvement rates than human drivers in reasonably similar deployment areas and they are involved in less egregious types of crashes.”
Lead author Eric Teoh also warned that the available reporting framework may struggle as autonomous fleets scale. “The present data collection system isn’t good enough to allow continuous monitoring of a large-scale expansion,” Teoh said.
The study gives Waymo one of the more concrete public comparisons yet between its robotaxis and conventional traffic, with the major caveat that human crash reporting remains incomplete. The 50-million-mile Waymo sample showed strong aggregate results, especially in Phoenix and Los Angeles, while Austin demonstrated that the numbers can shift sharply when the local dataset is smaller.
AutoGuide's daily news hound. I report on the latest breaking news in the auto industry across the US and around the world.
More by Marcus Lewinsky
Comments
Join the conversation
Why are they comparing WAYMO at only 50 million miles, and humans at 222 billion miles? Why not compare only 50 million miles of humans driving so that it's on an even scale?
If WAYMO racked up that many less miles than humans, then of course it's going to show less crashes. Apples to apples.
I agree it IS "crash rate" but it's also like saying a car driver (that does 1,000 miles a YEAR in commutes) is the same as a truck driver that does 50,000 miles a year (1,000 miles is ALMOST achievable daily). Until/unless you have "apples to apples" any data is suspect, how many humans vs. how many AI? - I could opt to look at 50 million miles of HUMAN traffic and possibly find 0 (ZERO) incidents, meaning crash rate WILL scale - the comparison is USELESS without an "apples to apples" (time of day, route, (therefore) vehicles on the road (AND on THAT particular road), etc. comparison) - doing 50 million miles at 03:00, when you're the only thing moving, does NOT compare to doing 50 million miles in bumper-to-bumper rush-hour traffic, so it CAN NOT be a valid comparison, even saying "I've got 1 car (robotaxi) on the road in rush hour traffic (which is about .02% of the total number of vehicles on the road) versus 4,440 cars (regular taxis) on the road during that time - of course it's smaller accident rate, it's as valid as saying "robotaxi has 68% fewer accidents than Cadillacs." It has less to do with the number of incidents than with the number of "targets". My step-father was a professional truck driver for FORTY YEARS (and many of the companies he worked for failed to track (and report) his safe driving while he was working there). He received awards from the ATA for 1 million AND 2 million miles of accident free driving (yes he had speeding tickets, not an accident and therefore did not impact his "safe" mileage award), make him (and fellow professional drivers) the LIMIT of your comparison, then you MIGHT get close to a valid comparison. The IIHS even admits it's data is "flawed" as many factors impact reporting (humans may not report at all, but the machine (hopefully) is compelled to report every little "curb check" - admittedly that suggest the data may be even better for the robot (as it's below humans, which is "artificially low" already) but you're also looking at 50 MILLION vs 222 BILLION - that's 4,440 TIMES less sample. The author "warns that the available reporting framework may struggle as autonomous fleets scale." put 4,440 times as many autonomous miles/drivers in the mix and then we'll talk (meaningful) numbers.