Uber’s Robotaxi Partner Crashed 16 Times in Four Months. The Safety Driver Intervened Once.
NHTSA opened an investigation into Avride, the Nebius subsidiary that operates robotaxis through the Uber app in Dallas, after identifying 16 crashes in four months. A safety monitor sat behind the wheel in every incident and attempted to intervene in exactly one. An original crash-rate analysis estimates Avride's collision frequency at 7 to 12 times the human driver baseline.
Sixteen crashes. One intervention. That is the ratio NHTSA reported on May 8, 2026, when the agency's Office of Defects Investigation opened a formal probe into Avride, the autonomous vehicle company whose robotaxis carry paying passengers booked through the Uber app in Dallas, Texas. Between December 2025 and March 2026, Avride's fleet of Hyundai Ioniq 5 vehicles changed lanes into the path of adjacent traffic, failed to slow for stopped vehicles ahead, and struck stationary objects including a dumpster. Every vehicle had a trained safety monitor in the driver's seat, and in fifteen of sixteen incidents, that monitor did nothing.
Mileage Nobody Published
Avride has not disclosed its total mileage, while Waymo publishes quarterly, and that asymmetry is exactly why the comparison matters.
Avride's fleet reached 200 Hyundai Ioniq 5 vehicles across Austin and Dallas by spring 2026, ramping from roughly 50 at the December 2025 Dallas launch. Using a conservative average fleet size of 125 vehicles operating over 120 days at 50 to 100 miles per vehicle per day in a nine-square-mile commercial zone, the fleet likely accumulated between 750,000 and 1.5 million miles during the crash period. Sixteen crashes over that range yields an estimated rate of 10.7 to 21.3 per million miles.
Now compare. Waymo, operating 3,000 vehicles across four cities, disclosed in March 2026 that it has driven 170.7 million rider-only miles at roughly 4 million miles per week. ARK Invest reported in late 2025 that Waymo averaged approximately 500,000 miles between police-reported collisions: about 2 per million. NHTSA tallies roughly 6 million police-reported crashes per year on 3.3 trillion vehicle miles traveled nationally, producing a human baseline of about 1.82 per million.
| Operator | Est. Crashes per Million Miles | Mileage Base |
|---|---|---|
| U.S. human drivers (2024) | ~1.82 | 3.3 trillion VMT |
| Waymo (2025–2026) | ~2.0 | 170.7 million miles |
| Avride (Dec 2025–Mar 2026) | ~10.7–21.3 | ~0.75–1.5 million miles (estimated) |
Even the most generous end of the estimate puts Avride at roughly six times the human baseline and five times Waymo's, a gap so wide that even tripling the mileage denominator would not close it. At the midpoint, the ratio is eight times both.
A 6.25 Percent Intervention Rate
Safety monitors exist for one reason: when software fails, a human grabs the wheel, steers the car back into safety, and spares whatever pedestrian or pickup truck or parked van happens to be in the wrong lane at the wrong second.
Avride's monitors were present in all 16 crashes, and in fifteen of those incidents, the monitor did not attempt to intervene, producing a 6.25 percent intervention rate, or a 93.75 percent failure rate for the single control mechanism that is supposed to stand between buggy software and a bent fender, between a lane-change algorithm that misjudges closing speed and a van driver who cannot swerve out of the way fast enough.
Avride declined to explain why the monitors did not act, telling TechCrunch it had "implemented targeted technical and operational mitigations" and that "the frequency of incidents relative to our mileage has steadily declined." Uber did not respond to requests for comment.
NHTSA reviewed video of each crash, and the agency's descriptions are damning: vehicles changing lanes directly into adjacent traffic, failing to slow for stopped vehicles in their own lane, striking stationary objects partially obstructing the road, behaviors that would earn a human driver a suspended license in most jurisdictions and a failed driving test in all of them. One December 2025 collision in Dallas involved an Avride vehicle clipping the open driver's-side door of a parked pickup, causing a minor injury that did not require hospitalization, while another involved the car attempting to change lanes around a parked truck and turning squarely into a van beside it.
Precedent matters here, because in October 2023, a single Cruise robotaxi in San Francisco dragged a pedestrian 20 feet after a separate human-driven vehicle knocked her into the AV's path, and California's DMV revoked Cruise's permit weeks later for the company's failure to disclose the dragging. Cruise's fleet had no safety monitors at all, while Avride's had monitors in every vehicle who sat through fifteen collisions without reaching for the wheel.
Uber's Platform Bet
Between 2015 and 2020, Uber spent billions building an in-house autonomous program that ended when a test vehicle struck and killed pedestrian Elaine Herzberg in Tempe, Arizona in March 2018, the first autonomous car to kill a pedestrian on a public road. Uber sold its tech to Aurora Innovation and pivoted to what it calls the platform model: let other companies build the robots, integrate them into Uber's network, take a cut of every ride without bearing the cost or the liability of developing the technology yourself.
No sensor development costs, no fleet liability, no sequel to Elaine Herzberg on your balance sheet. In Q1 2026, autonomous trips grew tenfold year-over-year on Uber's platform, with partners including Avride, Waymo, Nuro, Volkswagen's MOIA, and Wayve operating across eight cities and targeting fifteen by year's end.
But platforms cannot outsource experience. When a Dallas rider books UberX and matches with an Avride car, they see Uber's interface, pay Uber, rate the trip on Uber, and if that vehicle clips a pickup truck door, the passenger was in an Uber regardless of which company's name appears on NHTSA's investigation. The rider remembers Uber. Brand exposure travels with the passenger while balance-sheet exposure stays with the partner.
From Yandex to Dallas
Avride is a subsidiary of Nebius Group, the Amsterdam-based company formed when Yandex NV sold its Russian operations in 2024 for $5.2 billion. Founder Arkady Volozh relaunched with 1,300 employees, $2.5 billion in cash, and businesses spanning data infrastructure, edtech, and autonomous driving. Avride inherited Yandex's self-driving stack, under development since 2017, and is headquartered in Newburyport, Massachusetts, with teams in Texas, Israel, Serbia, and South Korea.
Uber and Nebius committed up to $375 million to scale Avride to 500 vehicles. Dallas launched December 3, 2025: a nine-square-mile zone covering downtown, Uptown, Turtle Creek, and Deep Ellum, with each Ioniq 5 carrying 13 cameras alongside lidar, radar, and additional sensors while a backseat tablet shows passengers the route in real time. Avride also runs sidewalk delivery robots for Uber Eats in Austin, Dallas, and Jersey City, and for Grubhub at Ohio State and other campuses.
What This Analysis Did Not Prove
Five limitations constrain the crash-rate comparison. First, Avride has not published total mileage, so the denominator in our per-million-mile calculation is an estimate derived from fleet size, operating area, and deployment timeline. If Avride's actual mileage is significantly higher than 1.5 million miles, the crash rate moves closer to the human baseline. If it is lower than 750,000 miles, the rate worsens further.
Second, NHTSA's Standing General Order on automated driving requires reporting at a lower threshold than the police-reported crashes used for the human baseline. Some of the 16 Avride incidents might not appear in police reports. That asymmetry overstates Avride's relative rate. But the gap is six to twelve times worse. Definitional differences alone cannot close it.
Third, Waymo has operated commercially since 2020 and accumulated 170.7 million miles, while Avride has four months in a single city, and every AV operator's earliest miles are the worst without exception in the industry's history, though Waymo worked through its worst period without charging passengers for the ride.
Fourth, NHTSA's crash descriptions note that "multiple crashes involved other vehicles turning into the Avride robotaxis," meaning fault allocation is unclear and some collisions may reflect other motorists' behavior more than Avride's system failures, though the investigation specifically targets "the competence of" Avride's software.
Fifth, this analysis cannot determine why Avride's safety monitors failed to act, whether they were undertrained, overwhelmed by the speed of events, or operating under policies that set the intervention threshold too high, and each explanation implies a different systemic failure that the available data cannot distinguish among.
The Strongest Case for Avride
Every robotaxi company that now operates safely once operated dangerously, and Waymo's early Firefly vehicles in 2015 averaged interventions every few hundred miles when Google's self-driving project logged 424 disengagements across 424,331 miles in California, roughly one per thousand miles, a rate translating to approximately 1,000 per million if each had become a crash. Safety drivers caught them. By 2026, Waymo's crash rate had dropped to about 2 per million, a three-order-of-magnitude improvement in a decade that is the core argument for the autonomous driving learning curve.
Avride's statement supports this framing: "Our total operations have continued to grow, while the frequency of incidents relative to our mileage has steadily declined." If Avride can demonstrate a Waymo-like improvement trajectory, these 16 crashes become an embarrassing early chapter rather than evidence of fundamental incompetence, but the honest question is whether that chapter should be written with paying passengers in the back seat when the safety monitor meant to protect them fails to act 93 percent of the time.
What You Can Do
If you ride in autonomous vehicles: Know which company built the technology driving you. When you book through Uber or Lyft and are matched with a third-party AV, you can decline and request a human driver at no additional cost. Do this if you are uncomfortable with an operator whose safety record you cannot verify.
If you are a regulator or policymaker: Require mileage disclosure alongside crash reporting, because crash counts without mileage denominators are fundamentally meaningless for cross-operator comparison. NHTSA's Standing General Order captures crash data but does not require operators to publish total miles driven, which makes it impossible for the public, journalists, or other regulators to calculate per-mile rates and means that a company operating a thousand vehicles over millions of miles looks identical to one running ten cars in a parking lot. Waymo publishes voluntarily while Avride does not, and mandating mileage disclosure would transform crash reports from anecdotes into evidence.
If you are an AV safety engineer: Study the safety monitor intervention failure. A 6.25 percent intervention rate across 16 crashes suggests a systemic design problem, not individual negligence. Whether the monitors were undertrained, fatigued, distracted, or operating under policies that set an intervention threshold too high, the outcome is the same: the backup system did not back up. Tesla's Autopilot and FSD research has consistently shown that humans are poor monitors of automated systems because the automation itself degrades their vigilance. If the monitor is not reliable, the system needs a different backup.
If you invest in the AV sector: Separate platform risk from technology risk. Uber's stock price depends partly on autonomous ride growth, which grew tenfold year-over-year last quarter. But every partner's crash is now Uber's brand incident. The platform model creates a portfolio diversification benefit for technology risk, since one partner's failure does not invalidate the others, but it concentrates brand risk, since one partner's headline lands on Uber's name.
Bottom Line
Uber built a $200 billion company on the principle that it does not need to own the cars to own the rides. The platform model now extends to autonomous vehicles: Uber does not build the self-driving technology, does not operate the fleets, and does not employ the safety monitors. What it does own is the passenger relationship. When a rider opens the Uber app in Dallas and gets matched with an Avride robotaxi that then changes lanes into a van, the experience is an Uber experience. Avride's 16 crashes in four months and its safety monitors' 93.75 percent failure-to-intervene rate are Avride's technical problem and Uber's brand problem. NHTSA described Avride's system as showing "excessive assertiveness and insufficient capability." That phrase is an engineering diagnosis, but it is also a description of a business model: deploy fast, learn on the road, and put the Uber brand on every ride while you figure it out. Every AV company's first miles are its worst. The question is who sits in the back seat while the worst miles happen.