Uber automated test vehicle kills a pedestrian in Tempe

On March 18, 2018, an Uber developmental automated test vehicle struck and killed Elaine Herzberg, 49, as she walked her bicycle across North Mill Avenue in Tempe, Arizona. The vehicle was driving under computer control with a human operator responsible for monitoring the road and intervening. The National Transportation Safety Board (NTSB) investigation found failures in perception, emergency response, operator attention and the management of public-road testing.^[1]

Detection did not produce a safe response

The modified Volvo XC90 approached at about 45 mph. Its automated driving system first detected Herzberg 5.6 seconds before impact, but alternated among vehicle, unknown-object and bicycle classifications; it never classified her as a pedestrian. Classification changes discarded the object’s tracking history when predicting its path. Unknown objects were treated as stationary, while a newly classified bicycle could be assigned a path in the direction of traffic. Those assumptions prevented the system from reliably anticipating her crossing.^[1]

At 1.2 seconds before impact, the system recognized an imminent collision. Its design then imposed one second of “action suppression”: it withheld braking while checking the hazard and possible alternative paths or awaiting human intervention. Uber’s Advanced Technologies Group (ATG) said this addressed false alarms. No operator alert accompanied the start of that interval.^[1]

After suppression ended, avoiding the collision required braking beyond the automated system’s configured limits. The system issued an auditory alert and initiated gradual slowdown; it was not configured to apply maximum braking merely to reduce the severity of an unavoidable crash. This was a control-policy restriction, distinct from the vehicle’s physical braking capacity. Volvo’s factory collision-warning and automatic emergency-braking functions were also deactivated during Uber automated operation because the two systems’ radars and braking commands were considered incompatible.^[1]

The operator had been looking down toward her phone and returned her gaze to the road about one second before impact. She took steering control 0.02 seconds before the collision and braked afterward. Impact speed was about 39 mph. The NTSB concluded that an attentive operator would likely have had sufficient time to avoid the crash or lessen its impact; the late return of attention left too little time.^[1]

Findings about responsibility and oversight

The report, adopted November 19, 2019, identifies the operator’s failure to monitor because of visual distraction by her personal phone as the probable cause. It separately identifies contributing factors: Uber ATG’s inadequate safety risk assessment, ineffective operator oversight and insufficient measures against automation complacency, all resulting from an inadequate safety culture. Herzberg’s impaired crossing outside a crosswalk and Arizona’s insufficient oversight of automated-vehicle testing also contributed. These are the investigation’s formal findings, rather than a claim that one classification error alone caused the death.^[1]

The organizational findings had concrete bases. Uber had replaced two-person crews with single operators in 2017. Supervisors rarely reviewed inward-facing camera footage, despite its availability. The system depended on an operator to notice emergencies without an immediate alert, yet the company had not adequately managed the risk of that operator disengaging from supervision.^[1]

The NTSB also distinguished institutional roles: federal vehicle-safety regulator NHTSA set motor-vehicle standards, while states governed road operation. It recommended mandatory submission and ongoing federal evaluation of developers’ safety self-assessments, stronger Arizona testing-permit review, and completion of Uber ATG’s safety management system. Recommendations describe proposed safeguards; they do not establish their later implementation.^[1]

Consequences and later proceedings

Uber ATG stopped public-road automated testing on March 19. Arizona suspended its automated-testing privileges on March 26. Testing resumed in Pittsburgh on December 20, 2018, under a 25 mph limit with two operators. The NTSB documented removal of action suppression, braking for crash mitigation, retention of tracking history across classification changes and reactivation of Volvo’s collision-warning and braking functions. These changes address specific failures; their existence is not evidence of general safety.^[1]

On March 29, 2018, Reuters reported that Herzberg’s daughter and husband had reached a settlement with Uber, citing their attorney. Terms were undisclosed. That reported resolution is distinct from the NTSB’s safety findings and a judicial determination of civil liability.^[2]

On July 28, 2023, operator Rafaela Vasquez pleaded guilty to endangerment and received three years of supervised probation. The amended sentencing order records the guilty plea and probation beginning that day. This criminal disposition concerns the operator; it does not replace the investigation’s separate findings about system design and institutional oversight.^[3]

The human-control question is concrete here: could the person assigned responsibility notice the hazard and change the outcome in time? The crash does not establish a hidden learned objective or Deceptive Alignment. As with Tay public chatbot incident, understanding the failure requires examining the deployed service and its safeguards alongside its automated behavior. The incidents guide compares what these different records can establish.