Urban Autonomy Lab
Chapter 2 · EV97

NHTSA Opens PE26-003 Investigation into Avride ADS After 16 Crashes in Dallas and Austin — First US Pre-Investigation of a Robotaxi Operator's ADS in 2026 — 8 May 2026

Date collected: 11 May 2026
Summary

First 2026 NHTSA preliminary evaluation of a robotaxi operator's ADS; safety drivers present but rarely intervened; "excessive assertiveness and insufficient capability" diagnoses behavioural-policy defect

Headline Finding

On 8 May 2026 the National Highway Traffic Safety Administration’s Office of Defects Investigation (ODI) opened Preliminary Evaluation PE26-003 into the Avride Automated Driving System (ADS) following sixteen crashes in Dallas and Austin between December 2025 and March 2026. Avride is the rebranded autonomous-vehicle unit of Yandex’s spin-off, and a partner of Uber, which has carried Avride passenger services on its platform in Dallas since December 2025. NHTSA’s notice frames the pattern as “excessive assertiveness and insufficient capability,” and observes that “the vehicles’ behavior may also constitute traffic safety violations.” A preliminary review of crash videos identified vehicles changing lanes into the path of others, failing to slow or stop for traffic ahead, and striking stationary objects partially blocking lanes. Material to the paper’s failure-mode framing: each crash occurred with an in-vehicle safety operator on board; in only one case did the safety monitor attempt to intervene. This is the first preliminary evaluation NHTSA has opened in 2026 of a robotaxi operator’s ADS in commercial passenger service, and the first to land on an Uber partner’s stack.

Detail

  • Investigation reference: ODI Preliminary Evaluation PE26-003 (INOA-PE26003-14280), opened 8 May 2026.
  • Crash count and geography: 16 crashes. Many in Dallas (where Avride has run a passenger service through the Uber app since December 2025); some in Austin (where Avride conducts on-public-road testing). Property damage in multiple cases; one “alleged minor injury.”
  • Observed failure modes (NHTSA preliminary review):
    • Unsafe lane changes into the path of adjacent vehicles (e.g. December 2025 Dallas: Avride vehicle attempted to lane-change around a parked pickup and struck a van in the adjacent lane).
    • Failure to slow or stop for traffic ahead.
    • Strikes against stationary objects partially blocking lanes (at least one collision with a dumpster).
  • Safety operator behaviour: Every crash occurred while an in-vehicle safety operator was present and monitoring. Of the sixteen crashes, only one describes the safety monitor attempting an intervention. NHTSA does not name this as the central concern but the regulatory pattern is unmistakable: the safety-driver-as-safety-net was not, in practice, intervening to prevent the system’s “excessive assertiveness.”
  • NHTSA characterisation: “Excessive assertiveness and insufficient capability” — language new to NHTSA’s ADS investigation lexicon. This formulation diagnoses a behavioural-policy defect at the ADS level rather than a sensor or perception failure. It also explicitly suggests the behaviour may constitute traffic violations — the second time in twelve months NHTSA has used statutory-traffic-violation framing in an ADS preliminary evaluation (the first being Tesla’s Robotaxi Austin notice).
  • Avride response: The company stated all sixteen crashes had been reported to NHTSA in accordance with the Standing General Order on automated-driving crash reporting (2021) and that it had “implemented targeted technical and operational mitigations to address findings from each reported incident.”
  • Uber operational exposure: Uber operates Avride passenger services in Dallas. Uber’s Q1 2026 earnings call (6 May 2026) reported 10x year-over-year growth in AV mobility trips and 10 new or expanded AV partnerships. The opening of PE26-003 lands 48 hours after Uber’s quarterly disclosure that did not name Avride risk.

Governance Implication

Four points relevant to the paper:

  1. Safety-driver-as-safety-net is empirically insufficient where the ADS is “assertive.” NHTSA’s preliminary finding that crashes occurred with safety monitors present but rarely intervening establishes — under federal investigation authority, not academic conjecture — that the supervised-deployment governance model assumed by US AV regulation (the safety driver as the operative safety guarantor) breaks down where the ADS’s behavioural policy is the proximate cause. The paper’s Ch.5 Allocation treatment, and Ch.7 Toolkit chapter on supervised-test conditions, should treat NHTSA’s PE26-003 framing as a foundational citation for the limit of the safety-driver model. Note the contrast with the earlier (pre-AVride) Cruise dragging incident, in which the failure mode was disclosure and remote-operator action, not in-vehicle supervision.

  2. “Excessive assertiveness” as a regulatory concept — the behavioural-policy axis of ADS oversight. Prior NHTSA ADS investigations have centred on perception failures (Tesla AP, Cruise) or simultaneous fleet failures (Cruise dragging, the regulatory questions following Apollo Go Wuhan). PE26-003 is the first preliminary evaluation in which NHTSA names the ADS’s behavioural policy — its tendency to act assertively in mixed-traffic conditions — as the problem. This is the regulatory phrasing the paper has been searching for to formalise its Allocation-failure-mode discussion: cities are not just managing where AVs may operate (the spatial ODD) but how they may act once operating (the behavioural ODD). The Ch.3 Institutional/Operational Environment chapter and the Ch.5 Failure Modes Allocation section should integrate this.

  3. The Uber partner-network multiplies exposure. Uber’s Q1 2026 disclosure of ten new or expanded AV partnerships, combined with its launch of “Uber Autonomous Solutions” — selling insurance, operations, maintenance and training data to AV partners — extends Uber’s reputational and operational exposure across an increasingly diverse partner stack of varying safety maturity. The Ch.4 Stakes treatment of fleet-operator-as-a-service (cross-reference EV94 Hertz/Oro Mobility) should now treat the Uber-partner-network exposure surface as an explicit governance design problem: a single high-volume rideshare platform now intermediates multiple operators of varying safety profiles, with PE26-003 as the first crystallisation of that exposure.

  4. Federal pre-emption tension in the SELF DRIVE Act becomes operationally testable. The House SELF DRIVE Act of 2026 (H.R. 7390, introduced 5 February 2026) provides that operators may preempt state and local AV-deployment regulations so long as they submit a federal safety case. PE26-003’s documentation of a US-resident, Uber-platform-carried operator’s behavioural-policy defect under existing state pilot frameworks makes the federal-preemption question — would a Self Drive Act federal certification have prevented this? — concretely answerable in legislative testimony.

Sources

Thread