A self-driving Uber’s killing of an Arizona pedestrian this month may have been related to the company’s decision two years ago to scale back safety sensors used to detect objects in the road, according to a report.
Uber decided in 2016 to use only one roof-mounted lidar sensor on a new fleet of Volvo SUVs — replacing a fleet of self-driving Ford Fusion cars that had been using seven lidar units, according to diagrams obtained by Reuters.
That decision “resulted in a self-driving vehicle with more blind spots than its own earlier generation of autonomous cars, as well as those of its rivals,” Reuters reported late Tuesday.
Specifically, scaling back to a single lidar — a radar-like technology that uses laser light pulses to detect road hazards — “introduced a blind zone around the perimeter of the SUV that cannot fully detect pedestrians,” according to the report, which cited interviews with former employees and Raj Rajkumar, the head of Carnegie Mellon University’s transportation center who has been working on self-driving technology for over a decade.
The lidar system made by Velodyne — one of the top suppliers of sensors for self-driving vehicles — sees objects in a 360-degree circle around the car, but has a narrow vertical range that prevents it from detecting obstacles low to the ground, according to information on Velodyne’s website as well as former employees who operated the Uber SUVs.
Autonomous vehicles operated by rivals Waymo, Alphabet’s self-driving vehicle unit, have six lidar sensors, while General Motors’ vehicle contains five, according to information from the companies.
Uber declined to comment on its decision to reduce its lidar count and referred questions on the blind spot to Velodyne. Velodyne acknowledged that with the rooftop lidar there is a roughly three-meter blind spot around a vehicle, saying that more sensors are necessary.
“If you’re going to avoid pedestrians, you’re going to need to have a side lidar to see those pedestrians and avoid them, especially at night,” Marta Hall, president and chief business development officer at Velodyne, told Reuters.
The safety of Uber’s self-driving car program is under intense scrutiny since Elaine Herzberg, 49, was killed last week after an Uber Volvo XC90 SUV operating in autonomous mode struck and killed her while she was jaywalking with her bicycle in Tempe, Ariz.
The precise causes of the Arizona accident are not yet known, and it is unclear how the vehicle’s sensors functioned that night or whether the lidar’s blind spot played a role. The incident is under investigation by local police and federal safety officials who have offered few details, including whether Uber’s decision to scale back its sensors is under review.
Uber has said it is cooperating in the investigation and has pulled all of its autonomous cars off the road, but has provided no further details about the crash.
The earlier Fusion test cars used seven lidars, seven radars and 20 cameras. The newer Volvo test vehicles use a single lidar, 10 radars and seven cameras, Uber said.
At Uber’s September 2016 unveiling of its Pittsburgh self-driving car operation, it was still using the Fusions, but had a Volvo on display. Uber staff pointed to the sleekness of the SUV and the relatively small roof mount with only one lidar system, a more attractive upgrade from the Fusion, which had a bulkier look with more sensors attached to the exterior.
A former employee said Uber justified the decision to slim down to one lidar by saying they “overdid it” with the additional sensors on the Fusions, suggesting the multiple lidars were unnecessary as Uber continued to refine its self-driving system.
Uber’s decision to move from the Fusion to a much taller vehicle exacerbated the issue of a blind spot from a single lidar unit, said former employees, because the lidar now sits up higher on top of an SUV, further reducing its ability to see low-lying objects — from squirrels to the wheels of a bicycle or a person’s legs.
One former Uber employee involved in testing both the Fusions and Volvo SUVs said that during a test run in late 2016, the Volvo failed to see a delivery truck’s tailgate lift that extended into the street, and the car nearly hit it going 35 miles per hour.
Since Uber launched a self-driving car program in early 2015, it has hustled to catch up with Waymo, which began working on the technology in 2009. Uber management moved swiftly and confidently even as some car engineers voiced caution, according to former employees, in a rush to get more cars driving more miles.
Seven experts who have reviewed the crash agree that a self-driving system should have seen Herzberg and braked. She had crossed nearly the entire four-lane, empty road before being struck by the front right side of the vehicle. The night was clear and streetlights were lit.
“Radar is supposed to compensate for (the lidar’s) blind spot,” said Rajkumar.
Uber declined to comment on its radar system. Volvo Car Group, owned by China’s Geely, declined to comment. A Ford spokesman said the company was not involved in Uber’s use of the Fusion or the self-driving technology employed on the cars.
To be sure, there are many possible causes of the crash other than the lidar blind spot. There could have been a software failure in the Uber car, said Richard Murray, an engineering professor at California’s Institute of Technology and the former head of Caltech’s student self-driving team.
“But this would be quite surprising since there was nothing else on the road,” he said.
With Reuters
