The problem is that for an AI to learn to handle real-world chaos, it needs to be exposed to a whole range of possible events. That’s why driverless car companies have driven millions of miles on the streets of the world over the past decade. Some, such as Cruise and Vemo, have begun testing unmanned vehicles in a handful of quiet urban environments in the U.S. But progress is still slow. “Why haven’t we seen the expansion of these little pilots?” Why aren’t those vehicles everywhere? “Asks Urtasun.
Urtasun makes bold claims for the head of a company that has not only road-tested its tech, but also has no real vehicles. But by avoiding the bulk of the cost of road-testing software in real vehicles, she hopes to build an AI driver faster and cheaper than her competitors, giving much-needed impetus to the industry as a whole.
Waabi is not the first company to develop a virtual world to test self-driving software. In the last few years, simulation has become the mainstay for driverless car companies. But the question is whether the simulation alone will be enough to help the industry overcome the final technical hurdles that have prevented it from becoming a viable proposition. Jesse Levinson, co-founder and CTO of Zoox, an autonomous-vehicle startup purchased by Amazon in 2020, says, “No one has yet created a metric for self-driving cars.
In fact, almost all autonomous-vehicle companies now use simulation in some form or another. It speeds up testing, exposes AI to a wider range of views than you would see on real roads, and it reduces costs. But most companies combine simulation with real-world testing, usually moving back and forth between real and virtual paths.
Wabi World plans to take the use of simulation to another level. The world itself is generated and controlled by AI, which acts as both a driving instructor and a stage manager – AI rearranges the virtual environment to identify the driver’s vulnerabilities and then test them. Wabi World teaches multiple AI drivers different capabilities at once before combining them into a single skill set. All of this happens nonstop and without human input, says Urtasun.
Driverless-car companies use simulations to help test how neural networks that handle vehicles handle rare events – a bike courier cutting ahead, a truck blocking the sky color road, or a chicken crossing the road – And then tweak it. Accordingly
“When you rarely have an event, it takes thousands of road miles to properly test it,” says Sid Gandhi, who works on a simulation at Cruise, a company that has a completely autonomous number of roads in San Francisco. Testing of vehicles has started. . That’s because rare-or-long-tail-events can only happen once in a thousand. “As we work on resolving the long tail, we will rely less and less on real-world testing,” he says.