Urban Autonomy Lab
Chapter 2 · EV104

Tesla Unredacts All 17 NHTSA Robotaxi Crash Narratives; Two Crashes Were Teleoperator-Caused, Not ADS-Caused — 15 May 2026

Date collected: 18 May 2026
Summary

First federally-documented case in which a remote assistance operator is named as the proximate cause of a robotaxi crash; teleoperator-caused crashes at 8 mph and 9 mph in Austin, below Tesla's promised 10 mph cap

Headline Finding

On 15 May 2026 Tesla quietly unredacted all 17 Robotaxi crash narratives previously filed to NHTSA under Standing General Order 2021-01 with the designation “[REDACTED, MAY CONTAIN CONFIDENTIAL BUSINESS INFORMATION].” The unredacted filings reveal that two of the 17 crashes — one in July 2025 and one in January 2026, both in Austin — occurred while a remote teleoperator, not the on-board Automated Driving System, was actively piloting the vehicle. In both cases the safety monitor in the driver’s seat had requested remote assistance, and the teleoperator’s manoeuvre directly caused the collision. The unredaction makes Tesla the last US robotaxi operator to comply with the de facto transparency norm that Waymo, Zoox, Avride, and May Mobility have followed since the SGO was issued. Tesla had previously argued the data should be withheld because public release would cause “financial harm.” The unredaction comes 7 days after NHTSA opened Preliminary Evaluation PE26-003 into Avride (EV97), in which the central federal-investigation framing was the failure of supervised-deployment safety models — and it directly contradicts the safety narrative Tesla offered to Senator Markey in 2025, which assured lawmakers that remote operators would not actively drive the vehicle except below 10 mph for short repositioning.

Detail

  • Filing: Tesla amended response to NHTSA Standing General Order 2021-01, unredacted narratives now visible across all 17 ADS-equipped incidents reported between July 2025 and March 2026.
  • Sample size: 17 unique ADS incidents, all 2026 Tesla Model Y, all in Austin, all with safety monitor present in driver’s seat, all without paying passengers on board.
  • Injury distribution: 13 property-damage only; 2 no injury; 1 minor injury without hospitalization; 1 minor injury with hospitalization.
  • Two teleoperator-caused crashes (Tesla’s own narrative language):
    • July 2025: Tesla ADS stopped on Austin street, unable to proceed. Safety monitor requested remote assistance. Teleoperator “took over vehicle control and gradually increased vehicle speed and turned the Tesla ADS left toward the left side of the street,” then drove “up the curb and made contact with a metal fence” at approximately 8 mph. This crash produced the only minor-injury-without-hospitalization outcome in the dataset.
    • January 2026: Safety monitor requested remote assistance. “The teleoperator took over vehicle control when the ADS was stopped and proceeded straight on the street. The Tesla vehicle made contact with a temporary barricade for a construction site at approximately 9MPH, scraping the front-left fender and tire.”
  • Tesla’s prior statement to Senator Markey (2025): remote operators may pilot the vehicle “as long as they stay under 10 miles per hour” and “this capability enables Tesla to promptly move a vehicle that may be in a compromising position, thereby mitigating the need to wait for a first responder or Tesla field representative to manually recover the vehicle.” Both teleoperator-caused crashes were at speeds (8 mph, 9 mph) below the stated 10 mph cap.
  • Other ADS-fault incidents disclosed in the unredaction: drove into metal chain across an unprotected left turn (September 2025); clipped the gooseneck hitch of a parked dump trailer at 18 mph (October 2025); reversed into a wooden electrical pole at 1 mph (January 2026); hit a curb in reverse at 2 mph (January 2026).
  • Most serious incident in dataset (the single hospitalization): July 2025, Tesla ADS creeping at 2 mph in a right-turn slip lane in Austin, rear-ended by following SUV — ADS at low speed in a position where a competent human driver would not have been operating.
  • Comparative benchmark: Tesla’s Austin Robotaxi fleet has logged approximately one crash per 57,000 miles across roughly 800,000 total miles driven. Tesla’s own published human-driver benchmark is one crash per 229,000 miles — implying Tesla’s supervised Robotaxi operation is currently producing approximately 4x the crash frequency of human drivers, before adjusting for the lower-speed urban-arterial operating environment.
  • Coverage: Electrek 15 May 2026 (unredacted narrative source); TechCrunch 15 May 2026 (Sean O’Kane); Interesting Engineering 15 May 2026; Yahoo Autos 15 May 2026.

Governance Implication

This is the first federally-documented case in which a remote-assistance operator (RAO) is named as the proximate cause of a robotaxi crash. The CAM paper’s Ch.5 Failure Modes treatment of Legitimacy and Allocation should incorporate three structural points:

  1. The Remote Assistance Operator is itself a failure surface — not only a failsafe. Throughout 2024–2025 the regulatory and operator narrative positioned RAOs as a safety guarantor, the human-in-the-loop element that compensated for ADS limitations at the edges of the operating design domain. Both Tesla teleoperator-caused crashes occurred precisely in the contexts the RAO is supposed to recover (ADS unable to proceed; safety monitor requested help). The RAO’s intervention then caused the crash. The paper’s Ch.5 treatment of Allocation failure should explicitly name teleoperator-caused operational defects as a sub-mode — distinct from in-vehicle safety operator failure to intervene (the failure mode named in NHTSA PE26-003 for Avride, EV97).

  2. Behavioural-policy regulation must extend to remote-assistance behavioural policy. NHTSA’s PE26-003 framing introduced behavioural-policy-as-defect into US ADS regulation. The Tesla unredaction extends the surface: behavioural policy is operative not only when the ADS is driving but when the RAO takes manual control. The CAM paper’s Ch.7 Toolkit Rules framing of Institutional ODD should distinguish (a) ADS behavioural ODD, (b) safety-monitor escalation protocol, and (c) RAO operational policy as three distinct behavioural surfaces, each requiring separate municipal/state regulatory visibility. The 10 mph cap Tesla cited to Markey is the operative example: an operator-set behavioural cap that was technically observed (both crashes were under 10 mph) but did not prevent harm, because the cap was the wrong instrument.

  3. The SGO transparency norm is now operationally complete — and the case for codifying it is strengthened. Tesla’s voluntary unredaction means that for the first time since the SGO was issued, all major US robotaxi operators report crash narratives to NHTSA in publicly readable form. Tesla used the “confidential business information” designation to withhold this data for approximately 12 months. The Markey-promised Remote Assistance Operator legislation (watching brief from EV45 / EV54 / EV97) has acquired a second federal-investigation foundation: PE26-003 (in-vehicle safety operator failed to intervene) and the Tesla unredaction (RAO took over and caused the crash) together name the full lifecycle of human-in-the-loop deployment as empirically insufficient under the current regulatory regime. The CAM paper’s Ch.9 What To Do Now section should treat the Tesla unredaction as the trigger for a statutory crash-reporting regime that cannot be designated confidential, and for separate regulatory standards governing RAO certification, training, and behavioural policy.

The Advocates for Highway and Auto Safety March 2026 poll (EV65) found 78% public support for mandatory federal crash-reporting and 79% support for RAO safety standards. The Tesla unredaction is the empirical case those public majorities were anticipating.

Sources

Thread