A new way to teach robots to drive
An autonomous truck startup called Waabi has developed what it calls "the ultimate school for self-driving vehicles" — one that requires humans to spend less time tutoring them on actual roads.
Why it matters: The more that autonomous vehicles learn to drive in the virtual world, the less time they need to practice on physical streets, which could be a safer, faster way to bring them to market.
How it works: Today's method for educating self-driving cars is a mix of simulation and real-world driving — one that's very labor-intensive for humans.
- AVs can't possibly log enough real-world test miles to prepare for every potential scenario — like an airplane landing on the road ahead, for example.
- So instead, developers drive cars around collecting sensor data, which they use to create a digital representation of the journey — with pools of people manually labeling objects like trees, curbs and pedestrians.
- AV companies use that data to train their AV systems and then put their cars back on the road for more testing.
- It's laborious and time-consuming, which is why companies developing self-driving technology are limited to a handful of neighborhoods in a few cities.
But Waabi — whose CEO, Raquel Urtasun, used to be Uber's top AV scientist — takes a fresh approach to simulation, one that minimizes human intervention:
- "Waabi World" uses AI to create a digital twin of the real world, and to simulate the many ways an AV might see it.
- The system enables the Waabi driver to learn on its own, designing tests to challenge the robot and continuously assessing its skills.
Like a driver's ed teacher who won't let you pass until you master parallel parking, Waabi World uses AI to pinpoint the Waabi driver’s weaknesses and automatically create "adversarial scenarios" it will have difficulty handling.
- "One way to think about this is that Waabi World is deliberately playing against the Waabi Driver, identifying and exploiting its weaknesses while the Driver simultaneously learns its skills," the company writes in a blog post.
- "It’s a battle of scenarios and driving skills — one AI system versus another."
The intrigue: In Waabi World, the robot can also learn to master different skills simultaneously.
- For example, the Waabi driver might be learning how to drive down a quiet suburban street, on a 5-lane freeway and downtown during rush hour — all at the same time.
Between the lines: Self-driving vehicles need to develop the same sort of intuition that human drivers do.
- A 16-year-old with a newly minted driver's license can't possibly be prepared for every situation they're going to encounter, but instinct and experience help them improve.
The big picture: Waabi's emphasis on simulation could help it compete with self-driving truck companies that are further along.
- TuSimple, Embark and Waymo Via are already hauling freight for customers in real-world tests and plan to commercialize their technology within a couple of years.
- A simulator like Waabi's could also be used one day by government regulators to validate whether a company's self-driving technology is safe for deployment.
Editor's note: This story has been updated with new wording to clarify how AV companies typically use training data.