Computer scientist Raquel Urtasun, Waabi's founder and CEO, says her company has created next-generation simulation for training self-driving cars.
WaabiWaabi, a Toronto-based AI startup that came out of stealth last year, says it’s developed an advanced simulator that can train autonomous vehicles to handle nearly limitless types of driving conditions–in a virtual world–and do so faster and more thoroughly than self-driving rivals that prioritize road tests.
The Waabi World platform is more comprehensive than any used by competitors as it can more accurately mimic real-world scenarios and create the types of rare, challenging “edge cases” that occur on the road only rarely, company founder and CEO Raquel Urtasun tells Forbes. Learning in this elaborate virtual world is happening constantly, preparing the software to eventually drive a range of vehicles from robotaxis to semi-trucks.
It’s the “most scalable, high-fidelity, closed-loop simulator that ever existed and, we believe, the key to unlocking self-driving technology at scale,” says Urtasun, who is also professor of computer science at the University of Toronto and a past chief scientist for Uber’s autonomous vehicle team. “It's an immersive and reactive environment that can automatically design tests for our self-driving brain, which we call the Waabi Driver, and also automatically assess the skills of the Waabi Driver. Ultimately, it can also teach the Waabi Driver to learn the skills of driving.”
The company isn’t planning to license the Waabi World platform to other AV developers. Urtasun said the focus is on intense training of its software for now, declining to say when Waabi’s first autonomous vehicles would be on the road. We’re advancing really well, really, really fast,” she said, without elaborating.
The pace of development for the leading autonomous vehicle companies, including Alphabet Inc.’s Waymo, General Motors-backed Cruise, Argo Ai, which is funded by Ford and Volkswagen, Amazon’s Zoox and robotic truck developer TuSimple, has been somewhat slower than anticipated though all the companies are making progress toward broader commercialization. Those companies, which have raised billions of dollars, are doing rigorous on-road testing to verify software and sensors, as well as racking up countless miles in simulation, yet it’s unclear when robotaxis and autonomous trucks will be ubiquitous.
Waabi–a name inspired by an Ojibwe word meaning “she has vision”–emerged from stealth mode in June 2021, announcing an initial $83.5 million funding round. Urtasun declined to say whether she’s raised additional funds in subsequent rounds. The initial focus will be applying the Waabi Driver software to self-driving trucks, but Urtasun envisions broad usage across all forms of vehicles.
“We talk a lot about how you need a different brain for the self-driving vehicle for it to generalize to all the different things that might happen, but there is one more piece that you need which is a different way of testing, as well as training the system to really handle all situations,” she says. “The industry mainly does this by testing and driving millions of miles in the real world. This is very costly and it doesn't scale. ... We need a totally different approach and this is what Waabi World is.”
If Waabi does indeed have a more comprehensive, robust simulation system than any that currently exists, it could prove to be a critical asset for the young company. The current limitations of simulation-based training software is one reason that Zoox is increasing its real-world testing, CTO and cofounder Jesse Levinson told reporters this week.
“In many ways the simulator is actually more useful than real driving but of course you can’t just not drive in real life either because simulations are not perfect, yet, We have not built the matrix for self-driving cars such that there are absolutely no differences between the simulated world and the real world,” he said. “It's getting better and better so there are fewer and fewer things that you learn in real life that you can't learn in simulation. But I do think it's going to be a long time before there's nothing you can learn in real life that you can't learn in simulation.”
