Canada’s leading developer of self-driving systems has unveiled its new strategy for teaching autonomous vehicles to perform safely: It’s basically a driving school for robots run by another robot.
More precisely, Waabi Innovations Inc. has developed an artificial intelligence-powered simulator that not only recreates the driving experience but challenges an automated driver’s weaknesses so it learns faster. On Wednesday, the Toronto-based company announced its simulator will provide a level of experience that would otherwise take thousands of autonomous vehicles driving millions of miles on the road to acquire.
“We’ve flipped the equation for how much simulation versus real world you need‚” said Raquel Urtasun, the company’s chief executive officer and a professor of computer science at the University of Toronto. “It’s really the key for unlocking self-driving technology.”
Waabi is one of several companies developing computer systems that can safely control a car or truck under a full range of real-world settings. People in the industry say there are commercial incentives to having machines minimize the need for human drivers, as well as safety benefits if those systems can reduce traffic accidents caused by human error and inattention. In Canada alone, vehicle collisions account for 1,500 to 2,000 fatalities each year. A 2015 study by the U.S. National Highway Traffic Safety Administration found human error contributed to 19 of every 20 collisions.
But humans also bring years of experience and intuition about what happens on the road, including what pedestrians and other drivers may be thinking at any given moment. This cognitive advantage allows people to compensate when driving under challenging conditions, especially when something unexpected occurs.
Self-driving systems lack this life experience but can be taught how to react in different driving scenarios using an approach known as machine learning. However, this requires an environment in which the training can take place. Doing so with a real car on a test track is expensive and time-consuming. It is also unlikely to provide sufficient exposure to situations that are rare or unforeseen. For this reason, simulators that are sophisticated enough to mimic the real world are seen as crucial to the industry’s progress.
“This is where the field is moving,” said Krzysztof Czarnecki, a professor of computer engineering at the University of Waterloo who is not involved with Waabi. “You need it – basically at every stage of development.”
He added that Waabi and other developers have come to recognize that conventional simulators, such as those found in video games, do not guarantee that a self-driving system will encounter everything it needs to learn to perform safely in the real world.
To counter this gap, Waabi’s simulator, Waabi World, is guided by machine learning just like the self-driving system it hosts. It generates novel traffic situations, then observes how the virtual driver responds to them. If it sees a problem with the driver’s performance, it adjusts to provide more cases that may cause the same problem to reoccur.
“Waabi World plays adversary against the Waabi driver,” Dr. Urtasun said. “It can learn to understand its vulnerabilities … and create the scenarios automatically that will make the system fail.”
Once a failure has been identified, the simulator can provide the automated driver with the information it needs to improve.
Dr. Urtasun described the simulator as a foundational proprietary tool that Waabi won’t license out or share externally – except with regulators, so they can test other self-driving systems. “This is our first big milestone,” she said.
Waabi’s progress is being closely watched. Dr. Urtasun is a global star in autonomous vehicle systems development. She joined U of T in 2014 after serving as an assistant professor with the Toyota Technological Institute in Chicago and became the chief scientist of Uber’s autonomous vehicle unit in 2017. Uber sold the unit to Aurora Innovation Inc. in late 2020, and Dr. Urtasun left with many members of her team to start Waabi.
Within months Waabi completed one of the largest early stage financings ever by a Canadian AI startup, raising US$83.5-million, led by Silicon Valley financier Khosla Ventures and backed by several Canadian venture capital firms and leading AI academics.
“She has definitely been on the leading edge,” Dr. Czarnecki said.
Shawn Chance, a partner with OMERS Ventures, one of Waabi’s venture investors, said the simulator “is the promised land here and what makes Waabi’s technology very valuable” because it will save time and money turning machines into real world-ready drivers.
Despite falling short on earlier promises, the auto industry continues to tout the arrival of autonomous vehicles later this decade, including several announcements last month from General Motors, Volvo and Intel-owned Mobileye in partnership with Chinese automaker Geely.
But even if self-driving vehicles can overcome the technical hurdles in order to share the road with humans, they are likely to encounter non-technical obstacles. A 2018 accident involving an Uber driverless vehicle that killed a pedestrian in Tempe, Ariz., was seen as a setback for the industry, though the situation was complicated by the fact that a human backup driver did not prevent the accident. Regardless, any further casualties caused by vehicles without a human at the wheel would likely spark fierce political and regulatory blowback. A broader issue is that of job displacement, particularly for those who drive for a living, including about 300,000 people in the Canadian trucking industry.
IVAN SEMENIUK, SCIENCE REPORTER
SEAN SILCOFF, TECHNOLOGY REPORTER
The Globe and Mail, February 9, 2022