Uber “software bug” led self-driving car to hit pedestrian

Uber researched the March 2018 incident in which an Uber self-driving vehicle struck and killed a pedestrian in Tempe, Arizona. Uber found that the vehicle’s onboard sensors did in fact detect the woman in the vehicle’s path, but the vehicle’s software incorrectly decided shew as a “false positive” and that the car did not need to stop for her.

Experts on self-driving car technologies said a tradeoff is required, between smooth rides and jerky stop-and-start to pause for potential objects.

Uber backup drivers fell short in safety functions

CityLab reported widespread shortcomings of the backup drivers who were responsible for supervising Uber’s self-driving cars. One, it is unclear whether humans can do a good job supervising machines that work well most of the time — requiring intense concentration to identify the occasional error, when most of the time, there are tempting distractions. Uber’s 8 to 10-hour shifts, with one 30 minute lunch break, were grueling — and drivers were often assigned to repeat the same driving “loops” which likely made the task particularly dull for drivers. Additional challenges included working entirely alone (without other humans) (after Uber removed a second staff person from each vehicle), and, CityLab reported, the vehicles’ frequent hard braking.

Meanwhile, CityLab spoke with multiple drivers who were dismissed from Uber for safety infractions, including using a phone while a vehicle was in motion — undermining any suggestion that all safety drivers do as instructed.

Relationship with Arizona governor questioned

In its efforts to court Arizona governor Doug Ducey, Uber built a relationship with Ducey that was unusually close. The Guardian obtained emails showing that the relationship included joint press conferences, Uber service on the governor’s policy committees, Uber providing meeting space to the governor when he visited San Francisco, and even the governor potentially wearing an Uber shirt.

Ducey enacted policies favorable to Uber. In Phoenix, city staff reported “pressure placed on us by the governor” to enact policies that Uber requested. In one episode, Uber asked that the governor promote Uber Eats via a tweet, which he did the next day. Ducey’s Uber dealings were particularly close on the subject of self-driving cars. After California revoked DMV registration of Uber vehicles that had not obtained the permits California said were needed, Uber sought to bring those vehicles to adjacent Arizona, which the governor permitted. Moreover, prior to Uber’s announcement of its self-driving vehicles on the road in Arizona, Ducey had allowed the vehicles to operate unannounced.

The public benefit of Ducey’s pro-Uber policies was not always apparent. The governor touted collaboration between Uber and Arizona’s College of Optical Sciences, but that school’s dean commented that “Our dialog with Uber has not led to any significant ongoing research engagement.” The governor allowed Uber to test self-driving vehicles on Arizona roads, only to backtrack when an Uber self-driving vehicle struck and killed a pedestrian in Tempe, Arizona. The governor touted economic benefits expected to result from Uber’s activities in Arizona, but while Uber brought self-driving cars to the state, its engineering teams largely remained elsewhere.

Removed second staff person from autonomous cars

Historically, Uber’s autonomous cars had two staff members onboard: One to take over driving in case of problems, and another to monitor onboard systems to track performance and label data. But Uber later moved to a single operator. Reviewing 100 pages of internal company documents, the New York Times reported that some employees expressed safety concerns about the change. Among other concerns, they noted that solo work would make it harder to remain alert during monotonous driving.

Broadly, problems seemed to have unfolded as internal critics worried. One Uber autonomous car safety driver was fired after being seen asleep at the wheel. When an Uber vehicle struck and killed a pedestrian in Tempe, Arizona, early review of the onboard video shows the staff person looking down or sideways, perhaps at a phone or onboard systems, but not at the road.

Self-driving cars fell short of expectations

Reviewing 100 pages of internal company documents, the New York Times reported that Uber vehicles were falling short of company objectives. For example, Google cars could drive an average of nearly 5,600 miles before a driver had to take control from the computer, whereas Uber vehicles struggled to meet the company’s target of one intervention every 13 miles.

Self-driving vehicle struck and killed pedestrian

An Uber self-driving vehicle struck and killed a pedestrian in Tempe, Arizona.

Early reports indicated that the pedestrian was crossing a roadway after dark, outside a crosswalk, and that Uber would probably be deemed not at fault in this incident.

But reviewing the crash video, multiple concerns arose. For one, Uber’s onboard driver — responsible for taking over in case of system problems — was looking down or sideways, hence unable to see the pedestrian. If her hands were on the steering wheel, ready to take over driving from the computer, that is not apparent from the video. Two, the pedestrian was making steady progress across the roadway. Three, some experts said a standard automatic emergency braking system, even on ordinary commercial vehicles, would have been able to detect the pedestrian and at least apply the brakes.

Velodyne, which makes the LIDAR sensors used on Uber’s autonomous cars, expressed surprise that the Uber vehicle hit the pedestrian. A Velodyne spokesperson explained in an email: “We are as baffled as anyone else. … Certainly, our Lidar is capable of clearly imaging Elaine and her bicycle in this situation.” Velodyne suggested that Uber’s software might be at fault, explaining that “[o]ur Lidar doesn’t make the decision to put on the brakes or get out of her way” and that Uber’s systems would need to make those decisions.

Salle Yoo questioned Kalanick’s handling of Levandowski

In addition to insisting that Uber hire outside investigators to check what improper information Otto held, then-General Counsel Salle Yoo questioned other aspects of Kalanick’s handling of that acquisition. Business Insider explains that she said she wanted Uber to fire Levandowski long before the company did so, and also that she was excluded from critical discussions about Levandowski. Business Insider reports that these disagreements led to Yoo’s departure from Uber.