Uber software failed to identify jaywalking woman

The heavily lauded software inside the Uber self-driving prototype that killed a woman in Arizona last year could not understand what she was, even though it detected her five full seconds before striking her.

Uncovered in a raft of new documents from the US federal investigation into the first known self-driving fatality was the fact the software wasn’t programmed to recognise people outside of pedestrian crossings.

The National Transportation Safety Board (NTSB), which usually covers major issues like plane crashes, found Uber had also switched off the standard Volvo autonomous emergency braking system that would have saved 49-year-old Elaine Herzberg’s life.

The car’s detection data proved it could have avoided Herzberg with better software. The NTSB records showed it detected her a full 5.6 seconds before it hit her at 70km/h.

The software records show it classified her as a ‘vehicle’ and then reclassified her as ‘other’, then ‘vehicle’ again, ‘other’ again, then ‘bicycle’ and then ‘other’ one more time and back to ‘bicycle’.

“The system design did not include a consideration for jaywalking pedestrians,” stated the NTSB’s preliminary report.

“Uber had equipped the test vehicle with a developmental self-driving system. The system consisted of forward- and side-facing cameras, radars, LIDAR, navigation sensors, and a computing and data storage unit integrated into the vehicle.

“Uber had also equipped the vehicle with an aftermarket camera system that was mounted in the windshield and rear window and that provided additional front and rear videos, along with an inward-facing view of the vehicle operator.

“In total, 10 camera views were recorded over the course of the entire trip.”

Each time it reclassified Herzberg, it reset the system that predicted where she was going.

By the time it realised it was out of time to stop or steer around her and that an impact was inevitable, it had just 1.2 seconds left before impact.

But Uber’s software had other failings up its sleeve.

Fed up with false alerts, its programmers had built a full second of lag into the system, giving its 44-year-old ‘driver’ Rafaela Vasquez an audible warning just 0.2 seconds before impact instead of 1.2 seconds.

Vasquez didn’t even begin emergency braking until a full second after the car hit the woman.

This was partly, the NTSB said, because Uber’s self-driving department didn’t have a safety department of any kind, much less a safety manager.

It follows last week’s report that Uber’s research prototypes had been involved in 37 autonomous driving crashes in the 18 months leading up to the fatal crash last year.

While the Herzberg family settled a lawsuit with Uber 11 days after the crash, the NTSB’s investigators will release a comprehensive report on November 19, detailing who was at fault and why, along with recommendations that will reverberate through the entire self-driving research industry.

Uber has: "Adopted critical program improvements to further prioritize safety,” said Uber spokeswoman Sarah Abboud in a statement.

“We deeply value the thoroughness of the NTSB’s investigation into the crash and look forward to reviewing their recommendations."