Now that video footage has been released of the moments preceding a fatal crash between an Uber self-driving vehicle and a pedestrian, it appears that more than just a flaw in Uber autonomous technology may have contributed to the crash.
A human safety driver, who did not act as the required fail-safe in the event something went wrong, was distracted moments before the crash. Yet that driver may not have been solely to blame, either. Investigators are still probing the incident, but it was likely a confluence of both factors.
Video footage of the crash, released Wednesday by the Tempe Police Department, sheds those insights on the circumstances which led to the death of Elaine Herzberg, 49, the first person killed in an incident involving a self-driving system on U.S. public roads.
Contrary to early news reports that described Herzberg as “abruptly” and “suddenly” entering the roadway, front-facing video shows she had already crossed at least one lane of Mill Avenue, possibly more, and was methodically pushing her bicycle in the center of a lane at the time she was struck by the Volvo XC90 SUV that Uber uses for testing. Weather conditions were clear.
It was dark at 10 p.m., the time of the crash, which affected what the front-facing cameras could capture. But the performance of the lidar and radar sensors in the self-driving system is not affected by lighting conditions, according to industry experts. Video of the immediate pre-crash circumstances underscores questions about why these sensors either didn’t detect Herzberg or why the system didn’t initiate evasive actions based on her presence.
The video doesn’t answer those questions, but it helps to crystallize key aspects of the investigation now underway, according to Anuj Pradhan, a research scientist at the University of Michigan Transportation Research Institute, who studies human factors related to vehicles, automated vehicles, and the broader traffic environment.
“Nighttime or daytime should not matter for systems using lidar or radar to pick up objects,” he told C/D. “And for all intents and purposes, the path and trajectory of the pedestrian seems fairly obvious. It was lateral to the vehicle’s direction of travel and carries a relatively big signature. I don’t want to say it’s straightforward, but it’s not one of those difficult edge cases. Given all these facts, the incomplete conclusion one could come to is that either a sensor didn’t work or the pedestrian was detected but the algorithms for perception didn’t work. There was a break in the chain there.”
If the inaction of the machine was one revealing aspect of the videos released, another was the actions of Uber’s human “safety driver,” who is supposed to monitor the vehicle’s operations and intervene should something go awry. Instead, inward-facing cameras show the driver, Rafaela Vasquez, was distracted and not paying attention to the road in the moments preceding the crash.
Analysis of the video shows Vasquez glanced downward for nine of the 12 seconds immediately preceding the crash. She glanced upward one second before the crash, and the look of panic and horror that overcame her face is one of several wrenching aspects of the videos, which, as released, stop at impact. Neither the Tempe Police Department nor Uber have said precisely what distracted Vasquez, who is listed in police records as Rafael but, according to Uber, is named Rafaela and identifies as female.
“The video is disturbing and heartbreaking to watch, and our thoughts continue to be with Elaine’s loved ones,” an Uber spokesperson said. “Our cars remain grounded, and we’re assisting local, state, and federal authorities in any way we can.”
“The video is disturbing and heartbreaking to watch, and our thoughts continue to be with Elaine’s loved ones.”
– Uber statement
The Tempe Police Department said its Vehicular Crimes Unit is investigating the crash and that the video will be among the evidence turned over to the Maricopa County Attorney’s Office, which is to make a determination on whether Vasquez or Uber will face charges related to the crash or Herzberg’s death. The National Transportation Safety Board (NTSB) and National Highway Traffic Safety Administration (NHTSA) both have teams on the ground in Tempe examining the crash.
Further, the videos counter earlier reports that indicated there was little the vehicle or its human safeguard could have done to avoid the crash. A San Francisco Chronicle report published Monday said Herzberg “abruptly walked from a center median into a lane of traffic.” The report quoted Tempe’s police chief, Sylvia Moir saying, “Preliminarily, it appears that the Uber would likely not be at fault in this accident, either.” Moir had seen the video at the time she made the statement. On Thursday, the Tempe Police Department disputed the context of her comments.
“While we recognize two media outlets have attributed exclusive interviews with the chief, we respectfully disagree with how their claim of interviews have been characterized,” said Detective Liliana Duran, a police spokesperson. “One was completely taken out of context from a brief interview on an entirely unrelated matter.”
This March 19, 2018, still image, taken from video provided by ABC-15, shows investigators at the scene of a fatal accident involving a self-driving Uber vehicle on the street in Tempe, Arizona.
The Chronicle did not return a request for comment Thursday.
In regard to the distracted state of the driver, the crash is similar to thousands of distraction-related crashes that occur every day on U.S. roads. Driver distraction accounted for 3450 traffic fatalities in 2016, according to federal statistics, though safety organizations like the Insurance Institute for Highway Safety believe distraction-involved crashes are likely undercounted because drivers often self-report and are reluctant to incriminate themselves, and because crash databases don’t consistently record information across state and federal bureaucracies.
Even safety drivers for automated vehicles are not immune. If anything, the task of overseeing a system that is performing as designed almost all of the time can be one of the most difficult tasks for humans to maintain. It’s a pivotal issue for manufacturers developing automated-driving systems that leave human drivers involved in at least some portion of the driving task, and one that likely rises to the forefront in the wake of the Uber crash.
“It’s this whole issue of human beings [being] very, very bad at monitoring a task if they’re only asked to monitor without being engaged,” Pradhan said. “It’s like you’re a quality-control person on an assembly line looking for one anomaly. It’s a boring task that humans are not good at doing.”
Traffic deaths have precipitously risen over the past two years for which full statistics are available, climbing at their fastest pace since a two-year period from 1962 to 1964, a period that spurred President Lyndon Johnson to push for the creation of the Department of Transportation. Pedestrian deaths, in the current span, have grown at a faster pace than overall fatalities and accounted for 5987 deaths in 2016, 15.9 percent of the total.
Amid such a backdrop of rising carnage, self-driving vehicles have been pitched as a compelling means to dramatically increase safety and reduce deaths. Any number of companies developing autonomous technology have reminded the public that 94 percent of all crashes, according to NHTSA, are caused by human error or behavior.
In this sense, the Uber crash represents the worst of both worlds—encapsulating the shortcomings of both human and machine.