DriveZero: End-to-End Driving Beyond Human Demonstrations
DriveZero trains a camera-only driving planner without human trajectory supervision.
The paper pairs a perception model distilled from frozen vision foundation models with an action model trained through closed-loop reinforcement learning. Its DriveRL framework turns real driving logs into interactive worlds and trains a privileged teacher policy with PPO rollouts. That teacher then supervises DriveZero through generated trajectories, including augmented driving intents beyond logged behavior. On the cited benchmarks, the authors report state-of-the-art results on NAVSIMv1, NAVSIMv2 and HUGSIM, plus a 93.57 mean nuPlan score for DriveRL. HF Daily Papers' note
The paper pairs a perception model distilled from frozen vision foundation models with an action model trained through closed-loop reinforcement learning. Its DriveRL framework turns real driving logs into interactive worlds and trains a privileged teacher policy with PPO rollouts. That teacher then supervises DriveZero through generated trajectories, including augmented driving intents beyond logged behavior. On the cited benchmarks, the authors report state-of-the-art results on NAVSIMv1, NAVSIMv2 and HUGSIM, plus a 93.57 mean nuPlan score for DriveRL. HF Daily Papers' note
score 5