Wayve is a British autonomous-driving company built around a single contrarian idea it calls AV2.0: that a self-driving car should be one large neural network that learns to drive from experience, rather than a stack of hand-engineered modules leaning on detailed maps. The company describes AV2.0 as “a single, learned AI driver trained to understand the world, anticipate risk, and adapt to new environments,” converting camera and radar input into driving outputs through one end-to-end model.
Wayve frames this in explicit contrast to the established approach, which it labels AV1.0. Traditional systems, it argues, “rely on hand-engineered stacks, HD maps and a rule-based approach,” which “limits scalability and the ability to rapidly generalize to new environments.” By learning to drive the way a human does, from watching the road, Wayve aims for “autonomy that can travel beyond a single city, route, or operating domain” instead of being painstakingly mapped into one city at a time.
This end-to-end, mapless philosophy is the high-risk, high-reward edge of the field. If it works, it sidesteps the enormous cost of building and maintaining HD maps and could generalize to new places far faster than the city-by-city rollout of robotaxi incumbents. The open question is whether a learned system can be made safe and verifiable enough for full deployment.
For a general reader, Wayve embodies the central live debate in self-driving: whether autonomy is an engineering problem to be decomposed and mapped, or a learning problem to be solved end-to-end, the same tension between hand-built systems and large learned models playing out across all of AI.