The problem is that in order for an AI to learn to deal with the chaos of real streets, it needs to be exposed to the full range of events it might encounter. That’s why driverless car manufacturers have spent the last decade driving millions of miles on roads around the world. A select few, like Cruise and Waymo, have begun testing vehicles without human drivers in a handful of quiet urban environments across the US. But progress is still slow. “Why haven’t we seen an extension of these little pilots? Why aren’t these vehicles everywhere?” asks Urtasun.
Urtasun makes bold claims for the head of a company that not only hasn’t tested its technology on the road, but doesn’t even have real vehicles. But by avoiding most of the expense of testing the software on the road in real vehicles, it hopes to build an AI driver faster and cheaper than its competitors and give the entire industry a much-needed boost.
Waabi isn’t the first company to develop realistic virtual worlds to test self-driving software. In recent years, simulation has become a mainstay for manufacturers of driverless cars. The question, however, is whether simulation alone will be enough to help the industry overcome the last technical hurdles that have prevented it from becoming a viable proposition. “Nobody has built the matrix for self-driving cars yet,” says Jesse Levinson, co-founder and CTO of Zoox, an autonomous vehicle startup that Amazon bought in 2020.
In fact, almost all manufacturers of autonomous vehicles now use simulation in some form. It speeds up testing, exposes the AI to a wider range of scenarios than it would see on real roads, and lowers costs. But most companies combine simulation with real-world testing, typically switching back and forth between real and virtual roads.
Waabi World plans to take the use of simulations to a new level. The world itself is generated and controlled by the AI, which acts as both driving instructor and stage manager – identifying the AI driver’s weaknesses and then rearranging the virtual environment to test them. Waabi World simultaneously teaches multiple AI drivers different skills before combining them into a single skillset. All of this happens non-stop and without human intervention, says Urtasun.
Driverless car companies use simulations to test how the neural networks that control vehicles deal with rare events — a bike messenger blocking their path, a truck the color of the sky blocking their path, or a Chicken crossing the road – and then tweaking them accordingly.
“If you have an event that’s rare, it takes thousands of miles of road to test it properly,” says Sid Gandhi, who works on the simulation at Cruise, a company that has started rolling out fully autonomous vehicles on a limited number of Testing Streets in San Francisco. That’s because rare — or long-tail — events may only occur once in a thousand. “As we work to solve the long tail, we will rely less and less on real-world testing,” he says.