Honing Interactions Between Cyclists, Pedestrians And Self-Driving Cars A Work In Progress

Sep 11, 2017

Between the nose of one of Uber’s shiny self-driving Volvo XC90s and one of Pittsburgh Tour Company’s hop-on-hop-off double decker buses was the cyclist. Through the windshield, his confusion was plain to see: He craned his neck, first left, then right, trying to see around the bus to figure out why it had come to a stop at a green light on Penn Avenue in the Strip District.

The car was working on the same question.

Driving is a succession of choices made within the parameters of traffic laws and informed by experience. And that’s where self-driving cars have an edge over humans, said Clark Haynes, an autonomy engineer at Uber Advanced Technologies Group (ATG) in Pittsburgh.

“When one self-driving Uber experiences an event, sees something interesting with a cyclist, it’s able to report that back. We can analyze that, and we can teach the algorithms to learn more about that event, handle it better,” he said. “We deploy it out to the fleet and instantly, a hundred plus cars already know how to handle that [situation].”

The technology company has been testing its cars on Pittsburgh streets for nearly a year. A survey issued by cyclist and pedestrian advocacy group Bike Pittsburgh this spring asked how people felt about sharing the road with autonomous vehicles; a lot of people liked them because robots don’t get road rage.

“In the U.S., 94 percent of all accidents are of human cause. That’s a huge opportunity,” said Haynes. As a bike commuter, one of the reasons he was motivated to work at Uber was the opportunity to make the roads safer.

But it’s going to take time.

At the moment, Uber’s self-driving cars can’t read cyclist turn signals, and might have trouble seeing a cyclist if he or she passes too close to the side of the vehicle, said Haynes. But that’s why they employ a vast range of sensing equipment—such as lidar, radar and cameras—and have their human vehicle operators annotate the data as it's collected.

Uber is not the only autonomous vehicle company racing to make a clearer picture. Delphi Automotive has a Pittsburgh lab currently testing a handful of self-driving cars, and expects to increase to at least 50 by early 2018. Delphi's chief technology officer Glen De Vos said each individual sensor has its weaknesses, but together can gather the level of information needed to operate in an urban environment.

“Work continues, though, to improve the basic sensing capability,” he wrote in an email. “Higher resolution cameras, higher angular discrimination and longer range radars, higher density point cloud information from lidars that will permit the algorithms that have been developed to better estimate the likely future movement and intention of pedestrians and cyclists.”

Besides Uber and Delphi, other companies are ramping up operations and testing in the region. Aurora Innovation, a company begun by veteran engineers from Google, Tesla and Uber, is looking for a self-driving vehicle operator in Pittsburgh, as is Argo AI.

Under state law, a licensed driver must be behind the wheel of autonomous vehicles, so it’s not clear to pedestrians, cyclists and other drivers when a car is driving itself versus being driven by a human. When respondents to Bike Pittsburgh’s survey wrote that they’d been passed on their bikes with fewer than 4 feet—a violation of Pennsylvania law—or been cut off, it could have been human error or an algorithm.

“We, as the public, need to trust that the cars are ready for prime time and ready to be on the streets,” said Bike Pittsburgh’s advocacy director Eric Boerer. “I don’t know if anybody can answer that other than [the companies].”

Public trust is crucial and built in part by just being out in the community, agreed Haynes of Uber and De Vos of Delphi. De Vos said early self-driving markets “will help socialize the technology in those communities.”

In August, Uber ATG opened its doors to Bike Pittsburgh members for a presentation and question-and-answer session during BikeFest, Bike Pittsburgh’s annual celebration of all things bicycle-related. The event lifted the veil a bit, said executive director Scott Bricker, as Uber engineers made themselves available for questioning by Bike Pittsburgh members.

But the fact remains that no one outside the company can access the vast amounts of data Uber collects to prove—and improve—the safety and efficacy of its cars.

“We want there to be community input into these technologies,” said Bricker. “And an acknowledgement that these companies are using our streets to perfect an imperfect technology. There’s an inherent risk that comes along with that.” 

Haynes said the cars are actively under development.

"If they were perfect this would already be a product," he said. 

In the interest of gathering feedback from the public, Uber has launched an e-mail address: aroundtown@uber.com. People can send observations to inform engineers working to improve the technology. Depending on the nature of the email, Uber may or may not send a response.

Back on Penn Avenue, the vehicle operator hovered his hands over the wheel. The car remained in self-driving mode, but the road scenarios were piling up: a car struggling to parallel park, the bus, the cyclist, pedestrians appearing from between parked cars. The human torqued the wheel, and was back in control.