End-to-end driving policies can achieve strong task performance, but they offer limited control over driving style — the way a vehicle accelerates, brakes, and maintains distance. The new arXiv paper YOUDrive takes up this problem directly.
The authors propose treating driving style as a "composable direction" for end-to-end autonomous driving. Rather than training a policy that implicitly settles on one style, the approach appears to make style an adjustable component of the driving policy.
Because the abstract is brief, the technical details of how the composition works are not yet available. But the framing suggests a move toward giving users and developers a clearer way to specify how an autonomous vehicle should behave on the road.