End-to-end driving is an approach in which a single neural network learns to map raw sensor input directly to driving commands, rather than splitting the task into separate hand-engineered stages for perception, prediction, planning, and control. The traditional “modular” stack, sometimes labeled AV1.0, leans on detailed HD maps and explicit rules; end-to-end systems instead try to learn the whole behavior from data, the way a person learns to drive by driving.
Wayve, which builds its business around this idea, describes it as “a single, learned AI driver trained to understand the world, anticipate risk, and adapt to new environments,” turning camera and radar input into driving outputs through one model. The argued payoff is generalization: a learned driver could travel “beyond a single city, route, or operating domain” without the painstaking per-city mapping that modular systems require, and it sidesteps the cost of building and maintaining HD maps. comma.ai’s openpilot likewise runs driving through “a large neural network” rather than a tower of hand-written logic.
The objection is safety and verifiability. A modular system can be inspected stage by stage, while a single end-to-end network is harder to interrogate and certify when something goes wrong, which matters enormously for a machine that can kill. Whether learned driving can be made trustworthy enough for full deployment is one of the open questions of the field.
For a general reader, end-to-end driving is the autonomy version of a debate running through all of modern AI: whether to engineer behavior explicitly or to learn it from data, trading transparency for the ability to generalize.