Waabi: the company betting miles don't matter
Raquel Urtasun argues you cannot drive your way to safety, and has raised $1.28 billion to prove it in simulation instead. Five years on, the thesis is compelling, the claims are unverified, and the public-road scoreboard reads zero.
Raquel Urtasun's argument is that everyone else is counting the wrong thing. Life-threatening situations arrive about once every ten million miles, so a fleet of a thousand trucks would drive a full year to meet one by chance — which makes real-world mileage, the industry's favourite scoreboard, a poor proxy for safety. Waabi's answer is to build the world instead: a generative simulator, Waabi World, that trains and tests an end-to-end virtual driver against billions of scenarios, and a mixed-reality rig that injects simulated traffic into a real truck's live sensor feed as it drives a test track. In January the thesis raised $750 million, took the company to roughly $3 billion, and added a deal to put 25,000 Waabi-driven robotaxis on Uber.
It is the most intellectually interesting position in autonomous trucking, and the least demonstrated. Five years and $1.28 billion in, Waabi has never run a driverless truck on a public road. The only driver-out milestone is a closed course in Phoenix in October 2025. The 2025 launch promised at the Series B slipped to "the next few quarters," then to a 2027 commercialisation target — and Waabi now says the gate is somebody else's hardware: it will not remove the driver until an OEM ships a fully redundant truck. Which is fair. It is also awkward, because on the Volvo VNL Autonomous that Waabi helped develop, the virtual driver going driverless first, in early 2027, is Aurora's.
Waabi's published "realism score" for its simulator — the claim that a simulated truck ends up within about ten centimetres of where the real one did. No paper, no data, no independent audit; a rival's chief scientist called its significance "unclear."
The autonomy
Waabi's boldest claim landed in June 2026. A driver trained entirely on a Peterbilt was moved to the Volvo VNL — different sensors, different controls, different shape — and, Urtasun says, drove highways and surface streets competently from its first mile with no retraining, no new data and no fine-tuning. If "zero-shot" transfer across vehicle bodies really works, it collapses the integration cost that makes every new platform a year-long engineering project, and it underwrites the one-brain-many-vehicles logic behind the robotaxi deal. The evidence offered publicly is a video and a photograph. Not one outlet that covered it tested the claim, and Volvo, which endorsed it, is a co-developer and an investor. Urtasun has told the industry "show me your evidence; as a scientist I want to see the proof." It is a standard worth applying evenly.
"Waabi may well be right that simulation beats mileage. The trouble with that argument is that the only way to prove it is with the miles it says don't matter."
What we're watching
- A public-road driver-out date — there has not been a Waabi announcement of any kind since June.
- Any independent verification, or a peer-reviewed paper, behind the realism and zero-shot claims.
- The robotaxi programme: 25,000 vehicles promised, no OEM, no cities and no launch date named.
Fact sheet — Waabi (2026)
The ai.cars verdict
- The most interesting technical thesis in the sector: build the world, don't drive it.
- Waabi World and mixed-reality testing are genuinely novel, and the zero-shot transfer claim would be a big deal if verified.
- Five years and $1.28 billion in, no driverless truck has run on a public road — the only driver-out was a closed course.
- Its headline claims rest on blog posts, video and a photograph, from a founder who demands proof from everyone else.
- Verdict: brilliant people, a real argument, and a scoreboard that is still empty. 2027 is when it stops being a thesis.
How does it stack up? Compare the Waabi Driver with the Aurora Driver, or build your own comparison.
