This ultra-realistic virtual world is a driving school for self-driving AI.

The problem is that in order for the AI ​​to learn how to deal with the chaos on real roads, it must be exposed to the full range of events that it can encounter. That’s why self-driving car companies have spent the last decade driving millions of miles on the streets of the world. Some, such as Cruise and Waymo, have begun testing driverless vehicles in several quiet urban settings in the US. But progress is still slow. “Why don’t we see the expansion of these little pilots? Why aren’t these cars everywhere?” Urtasun asks.

Urtasun boldly claims to be the head of a company that not only hasn’t road-tested its technology, but doesn’t even own a car. But by avoiding much of the cost of testing software on real streets, it hopes to build an AI driver faster and cheaper than its competitors, giving the entire industry a much-needed boost.

Virtual drivers

Waabi isn’t the first company to develop realistic virtual worlds for testing self-driving software. In the past few years, simulation has become a mainstay for self-driving car companies. But the question is whether one simulation will be enough to help the industry overcome the latest technical hurdles that have kept it from becoming a viable offering. “No one has built Matrix for self-driving cars yet,” says Jesse Levinson, co-founder and CTO of Zoox, an autonomous vehicle startup bought by Amazon in 2020.

In fact, almost all autonomous vehicle companies are now using simulation in one form or another. This speeds up testing by exposing the AI ​​to a wider range of scenarios than on real roads and reduces costs. But most companies combine simulation with real-world testing, typically cycling between real and virtual roads.

Waabi World takes the use of simulation to the next level. The world itself is created and controlled by artificial intelligence, which acts as a driving instructor and stage director, identifying the weaknesses of the artificial intelligence driver, and then rearranging the virtual environment to test them. Waabi World trains multiple AI drivers with different abilities at the same time before combining them into a single skill set. All this happens without stopping and without human intervention, says Urtasun.

Rare Events

Self-driving car companies use simulation to test how the neural networks that control vehicles handle rare events — a bicycle messenger driving ahead, a sky-colored truck blocking the road, or a chicken crossing the road — and then fine-tuning them. respectively.

“When you have an event that happens infrequently, it takes thousands of miles to properly test it,” says Sid Gandy, who works on simulations at Cruise, a company that has begun testing fully autonomous vehicles on a limited number of roads in San Francisco. . This is because rare – or long-tailed – events can occur only once in a thousand. “As we work to solve the long tail problem, we will rely less and less on real-world testing,” he says.