DriveZero beats nuPlan's log-replay expert without human demos

The teacher policy inside DriveZero, an end-to-end driving system trained without any human trajectory demonstrations, scored a mean of 93.57 across nuPlan's three community splits (Val14, Test14-hard and Test14-random), beating the log-replay expert on each, according to a preprint posted to arXiv on 5 September.

The 93.57 belongs to DriveRL, the reinforcement-learning teacher inside the stack, evaluated with what the authors call "value-guided test-time action search" in both reactive and non-reactive modes. DriveRL is trained with PPO in a closed-loop simulator built by converting real driving logs into interactive worlds, a way of using recorded data without imitating any specific trajectory.

Perception is handled separately by DriveVFM, which the paper describes as consolidating "multiple frozen vision foundation models, including DINOv3, SigLIP2, SAM and Depth Anything V2, into a single backbone from raw images alone, requiring no task-specific annotations." The final DriveZero planner is camera-only, distilled from that RL teacher. Because the teacher is goal-conditioned, the authors write that it can be "queried under augmented driving intents, yielding diverse, goal-consistent supervision that logged data cannot provide."

On the other benchmarks, the abstract says DriveZero "achieves state-of-the-art performance on NAVSIMv1, NAVSIMv2 and the closed-loop HUGSIM benchmark without any human trajectory supervision." The abstract publishes no per-benchmark numbers for those runs and does not list author affiliations.