What's this about?
DriveZero separates perception learning from action learning, trains a privileged driving policy in interactive closed-loop worlds, then distils it into a camera-only planner. It reports state of the art on NAVSIMv1, NAVSIMv2 and HUGSIM without human trajectory supervision.
What exactly is “closed-loop training”?
The model's action changes the simulated world, which produces the next observation and consequence. Human logs are fixed recordings; closed-loop worlds let the policy experience and recover from situations humans did not demonstrate.
Why is this worth noting?
Imitation caps behaviour at the coverage of recorded drives. Interactive training can explore more, though benchmark gains still need real-road safety validation.