UK-based self-driving startup Wayve takes its tech on the road

The electric Ford Mustang Mach-E rolled to a stop at an intersection in construction-riddled downtown Detroit and paused.

Across the street, a construction worker gestured at the vehicle, attempting to coordinate how the vehicle should proceed.

Up to that point, the vehicle had been making decisions on its own. But in this case, it required a little human intervention. Carlos, the safety operator behind the wheel of the vehicle equipped with self-driving tech startup Wayve’s software, gestured back at the construction worker, took over the controls, and safely navigated the vehicle along its route.

It was one example of the tricky edge cases that companies like Wayve are attempting to solve as they use some combination of AI, sensors, and cameras to enable autonomous driving.

The Wayve vehicle was in Detroit as part of a road show the London-based startup has undertaken as it prepares to launch its technology, initially as part of advanced driver assistance systems (ADAS) that require human engagement—and eventually, it hopes, in fully self-driving vehicles.

Tech Brew got a demo ride in a second-generation prototype vehicle equipped with six cameras, a radar unit on the front of the car, and Wayve’s signature AI software. The system enabled hands-off, eyes-on driving.

Central to Wayve’s approach to autonomy is what it calls its “Embodied AI” foundation model, which was designed for applications like driving and which the company says allows vehicles to understand the environment it’s in and learn in real time.

“What this model is trained to do is really understand how the world moves,” Rudi Rankin, Wayve’s marketing director, told us.

The model can be customized to automakers’ specifications. Wayve has “no intention of building our own robot fleet,” Rankin said. Instead, the startup—which raised over $1 billion last year—looks to integrate its software into automakers’ assisted-driving tech stacks.

“To do it in an urban environment has been quite tricky, so this AI software in its current format, when integrated with an automaker’s existing base software, can provide the capabilities to do point-to-point navigation in an urban environment, in addition to highway,” Rankin said.

Working together: On Sept. 21, Nissan announced that it was launching demonstrations of its next-gen ProPilot ADAS, which uses Wayve’s software. The system is slated to launch in Japan in fiscal year 2027, per a news release. The Japanese automaker reported that a fleet of prototype vehicles equipped with Wayve’s AI Driver software and Nissan’s perception tech successfully navigated the streets of central Tokyo.

The next-gen version of ProPilot is slated to offer capabilities on urban streets, taking it beyond highway driving.

Wayve’s “embodied AI has the ability to grasp its surroundings and anticipate what will happen next and how its own actions will affect the environment, allowing it to drive safely in harmony with its surroundings like a skilled human driver,” per the release.

And in other partnership news, on Sept. 18 Wayve announced it had signed a letter of intent with Nvidia for a potential $500 million investment in Wayve’s next funding round. Nvidia participated in the startup’s Series C funding round, and the companies have been working together since 2018. Wayve is building its third-generation robot platform on Nvidia’s Drive AGX Thor AI computer as it works to reach higher levels of autonomy.

Wayve, in partnership with Uber, recently got the OK to conduct advanced commercial trials on public roads in the UK.

Ridealong: The demo in Detroit took place along a 20-minute preplanned route. Wayve is focused on point-to-point navigation.

“One of the unique things about this end-to-end AI approach is…the feeling of the ride doesn’t feel programmed,” Rankin said. “It feels a lot more natural to the way that you and I might drive. And that’s resulting from the fact that we train on driving data.”

At one point during the drive, while approaching a right-hand turn, the vehicle encountered a cyclist who was preparing to cross the street. The interaction didn’t require any intervention from the safety operator; the vehicle waited for the cyclist to cross before turning.

“Those are the kinds of things that are really difficult to program. The opposite approach…is to use coding. You program in all the rules of the road,” Rankin said. “And what ends up happening is you have this very binary logic. I see a stop sign; I stop. Or, I see a red light; I stop. You can quickly get to a driving demonstration using that approach.”

“Where it becomes really difficult is when things start to blur lines,” she added. “When you don’t have a binary choice of do you go or do you stop. We solve that through millions and millions of examples of pedestrians. And so the model starts to understand, when I see a pedestrian, I can accurately predict what might happen.”