March 30, 2017 8:00 AM, EDT

Uber Crash Shows Human Traits in Self-Driving Software

Company's Self-Driving SUV 'Just Wanted to Beat the Light,' Eyewitness Tells Police

A recent crash involving an Uber Technologies Inc. driverless car suggests autonomous software sometimes takes the same risks as the humans it may one day replace.

The accident March 24 in Tempe, Arizona, caused no major injuries. Another human-driven car turning left failed to yield, hit the Uber car and flipped it on its side. After a short pause, the company's self-driving test fleet was back on public roads in Tempe, Pittsburgh and San Francisco early in the week of March 27.

But the Tempe Police Department report, released March 29, recounts a complex story.

The Uber Volvo SUV, outfitted with autonomous driving sensors, was heading south on a wide boulevard with a 40 mph speed limit. It had two of the company's test drivers in front and no paying passengers. The light turned yellow as the vehicle entered an intersection. A green Honda on the other side of the road was trying to make a left at the light. The driver thought it was clear and turned into the oncoming Uber SUV, according to the police report. 

In a statement to police, Patrick Murphy, an Uber employee in the car, said the Volvo SUV was traveling 38 mph, a notch below the speed limit. He said the traffic signal turned yellow as the Uber vehicle entered the intersection. He then saw the Honda turning left, but "there was no time to react as there was a blind spot" created by traffic. The Honda hit Uber's car, pushing it into a traffic pole and causing it to turn on its side.

During the event, the Uber vehicle was in autonomous mode, a spokeswoman for the company and the Tempe police said.

Others involved in the accident, though, didn't imagine a robot behind the wheel. Alexandra Cole, the driver of the Honda, told police that she could not see any cars coming when she decided to make the left turn. "Right as I got to the middle lane about to cross," she wrote, "I saw a car flying through the intersection."

Another witness told police that Cole was not at fault. "It was the other driver's fault for trying to beat the light and hitting the gas so hard," Brayan Torres told police in a statement. "The other person just wanted to beat the light and kept going."

Eyewitness accounts often can be unreliable, and other witnesses in the police report did not say that the Uber car was at fault — something the police agreed with. Still, Torres' account raises the question of whether Uber's self-driving sensors spotted the light turning yellow and, if so, whether it decided it could safely continue through the intersection.

One of Uber's self-driving SUVs ran a red light in San Francisco in 2016, and on five other occasions, the company's mapping system for its cars failed to recognize traffic lights in the area, the New York Times reported in February.

Uber's problems show the potential hurdles to winning approval for autonomous vehicles from the public and regulators. The company, and such rivals as Alphabet Inc.'s Waymo and major automakers, are working to tweak software to handle "edge cases," such as unusual driving conditions.

Self-driving cars have been criticized more often for driving too cautiously, slowing or stopping when human drivers would be more aggressive. Autonomous vehicles operated by Waymo have been rear-ended due to such issues and the company has been working to make its system more human.

There is a potential upside for Uber from the Tempe crash: It now has rich, unique data to use for its self-driving program. Last year, after a Waymo car car bumped into a bus, the company said it used the incident, and "thousands of variations on it," to refine its software. "This is a classic example of the negotiation that’s a normal part of driving —we’re all trying to predict each other’s movements," Waymo added.