Written by Abhirup Roy
SAN FRANCISCO (Reuters) – The General Motors Cruises robotaxi that struck and dragged a pedestrian 20 feet (6 meters) in an October crash made a number of technical mistakes. The failure worsened after the crash, an analysis commissioned by GM on Thursday revealed.
The College of Engineering's report said that failure to detect the woman's location, inaccurate detection of which part of the car hit her, and errors regarding the location of the car itself kept the vehicle in emergency mode even after hitting the woman. The vehicle continued to drive without stopping. Consultancy company Exponent.
The woman, who has not been identified, was injured but survived.
U.S. automaker GM's self-driving division, Cruise, has ceased operations and is facing multiple investigations, including one by the Justice Department, since the crash that rocked the autonomous vehicle (AV) industry late last year.
A technical report by Exponent, part of a report by law firm Quinn Emanuel, provides a second-by-second account of how the accident unfolded and details the technical errors.
The accident occurred after the pedestrian collided with another vehicle traveling one car length in front of him in an adjacent lane, sending him into the path of the Cruise AV, named Panini.
According to the report, the cruise vehicle could not have predicted the accident. It detected that the other car had hit a person and slowed down in just a few seconds, hitting the woman herself.
But after that, I couldn't find her. The woman fell and most of her body, except for her legs, was out of the field of view of the lidar object detection sensors. The cruise car struck her with its left wheel before coming to a stop, Exponent said.
At this point the vehicle could have made an emergency stop on the spot, but did not do so. Mistaking the collision to be a side impact rather than a head-on collision, the car proceeded approximately 20 feet at 7.7 mph (12.4 kph), pulling the pedestrian under and achieving its stated goal of pulling the car to the curb. safety.
In fact, the review found that the car was already in the lane next to the curb but did not realize it due to a location error.
The pedestrian's feet and lower legs were visible on the wide-angle left side camera from the moment of impact until the vehicle stopped, but neither the pedestrian nor her feet were classified or tracked by the vehicle, despite momentarily detecting the foot. , the defendant said. .
He added that there were no problems with sensor or vehicle maintenance.
Cruz said in a blog post Thursday that he updated the software to address the underlying issue.
The report itself makes clear that humans could have done better.
“After an AV makes contact with a pedestrian, an alert and attentive human driver would not notice that some kind of impact had occurred and continue driving without further investigating the situation,” the expo said. Nent said.
(Reporting by Abhirup Roy in San Francisco; Editing by Peter Henderson and Lincoln Feast.)