Generative AI could power the next wave of self-driving cars

• By tapping the power of generative AI, some companies see a faster, more capital-efficient way to teach robots to drive.

Driving the news: As AI begins to leap from the digital realm into the physical world, investors are pouring fresh capital into this next phase of AV development — "AV 2.0."

• Driverless truck company Waabi said today it had raised $200 million in an oversubscribed Series B round led by Uber and Khosla Ventures.
• That follows a huge deal last month by Wayve, a U.K.-born startup, which raised $1.05 billion in a Series C round led by SoftBank Group.
• Nvidia, one of the biggest names in AI, participated in both deals.
• Softbank also invested a reported $1 billion last September in Stack AV, an automated trucking company founded by former Argo AI execs.

The big picture: These AV 2.0 companies are developing self-learning systems for autonomous driving.

• Using AI, their goal is to teach a virtual driver to reason like a human so it can make snap decisions correctly and safely — even in novel situations.

Catch up quick: The original approach to AV development involved racking up millions of miles in test cars, collecting data that could be used to program self-driving algorithms with step-by-step instructions.

• Companies drove around 24/7 hunting for "edge cases" — rare events like the woman in a wheelchair chasing a duck that a Waymo test vehicle encountered in 2018.
• But it can take tens or even hundreds of years to log enough miles to prove that self-driving cars are safe, one report found.
• Alternative testing methods such as simulation are useful, though it's still time-consuming and capital-intensive to write an AV instruction manual.

The latest approach, driven by extraordinary leaps in generative AI, is based on intuitive learning, enabling AV companies to accelerate tech development.

• Waabi, for example, was founded in 2021 — well behind industry rivals — but says it has made speedy progress and is aiming to launch fully driverless trucks by 2025, roughly the same timeline as its rivals.

Zoom in: Waabi's Spanish-born founder and CEO, Raquel Urtasun, is an AI pioneer who helped lead Uber's now-defunct autonomy project.

• The Toronto-based startup invented what's called an end-to-end AI system capable of human-like reasoning.

• It requires significantly less training data and compute resources than other approaches — and can be tested in "Waabi World," an advanced virtual simulator.

Yes, but: Waabi still faces the same bar for safety.

• "Learning-based AVs can be taught driving skills more quickly, but the black-box nature of machine learning-based behavior makes it more difficult to validate safety," AV safety expert Philip Koopman, an associate professor at Carnegie Mellon University, tells Axios.
• "Even if you drive a billion miles in simulation, you still need to make sure the simulation wasn't missing edge cases that will cause crashes in the real world."

What they're saying: Waabi's system is "provably safe," Urtasun tells Axios, because its decisions can be interpreted and traced, unlike other "black box" AV systems.

• "The industry needs a big step forward in terms of proving the safety of their systems," says Urtasan, whose claims will be outlined in an upcoming white paper.
• "Incumbents ... will say, 'If I have driven enough miles, I should be safe.' That's not a sign of whether you are safe. It's only a sign of how much cash you spent driving."

Reality check: Generative AI, as fascinating as it is, remains far from perfect, as anyone who has played around with ChatGPT has discovered.

• No one wants a self-driving car to experience the kind of AI hallucination that urges you to put glue on a pizza.

The bottom line: When artificial intelligence comes to the physical world, the stakes become a lot higher.