Glonce

DriveZero skips human driving logs, hits SOTA on NAVSIM and HUGSIM

researchThe Glonce editorsHugging Face🔥71Glonce rating71Impact65Novelty75Relevance90Surprise45Players55Credibility80
DriveZero skips human driving logs, hits SOTA on NAVSIM and HUGSIM

Most end-to-end autonomous-driving systems learn by copying human driving logs, so what they can do is capped by the quality and range of behavior those recordings happen to contain. DriveZero is built to get past that ceiling. It splits the driving problem into a perception model and an action model, pretrains each in the regime that suits it, then combines both into a single end-to-end planner. The two halves need different training recipes: perception has to understand the world and benefits from large, varied visual data; action has to interact with the world and needs closed-loop feedback.

On the action side, the authors introduce DriveRL, a mixed-agent closed-loop reinforcement-learning framework. It converts real driving logs into interactive worlds, then trains a privileged teacher policy inside them with PPO through closed-loop rollouts. Because this teacher is goal-conditioned, it can also be queried under driving intents that were never in the original logs, producing diverse, goal-consistent supervision that recorded data alone cannot provide.

On the perception side, DriveVFM consolidates several frozen vision foundation models, DINOv3, SigLIP2, SAM and Depth Anything V2, into a single backbone built from raw images alone, with no task-specific annotations required.

DriveZero unifies the two: a camera-only planner that distills the frozen DriveRL teacher through the trajectories it rolls out. On nuPlan, DriveRL with value-guided test-time action search reaches a mean score of 93.57 across the Val14, Test14-hard and Test14-random community splits, in both non-reactive and reactive modes, exceeding the Log-Replay expert on all three. The paper further states that DriveZero achieves state-of-the-art performance on NAVSIMv1, NAVSIMv2 and the closed-loop HUGSIM benchmark, without any human trajectory supervision; no numeric scores for these three results are given in the source, only the 93.57 nuPlan figure, which belongs to DriveRL rather than the combined DriveZero planner.

Key facts

  • DriveZero splits driving into a perception model (DriveVFM) and an action model (DriveRL), pretrains each separately, then combines them into one end-to-end, camera-only planner.
  • DriveRL is a mixed-agent closed-loop reinforcement-learning framework that converts real driving logs into interactive worlds and trains a privileged teacher policy with PPO through closed-loop rollouts.
  • DriveVFM merges several frozen vision foundation models, including DINOv3, SigLIP2, SAM and Depth Anything V2, into a single backbone from raw images alone, with no task-specific annotations.
  • On nuPlan, DriveRL with value-guided test-time action search scores a mean of 93.57 across the Val14, Test14-hard and Test14-random splits, beating the Log-Replay expert on all three.
  • DriveZero, distilled from the DriveRL teacher, is reported as state-of-the-art on NAVSIMv1, NAVSIMv2 and the closed-loop HUGSIM benchmark without human trajectory supervision, though no numeric scores are given for those three results.

Why it matters

End-to-end autonomous-driving systems that only imitate human logs inherit the limits of what got recorded: they cannot learn a maneuver or a recovery that never appears in the training data. DriveZero's split design, perception pretrained on broad visual data, action trained in a closed-loop simulated world with reinforcement learning, is an attempt to remove that ceiling and let the driving policy learn behavior beyond what any human driver happened to log.

Who it affects

The paper speaks to researchers and teams building end-to-end, camera-only driving planners, and to anyone benchmarking against the nuPlan, NAVSIM or HUGSIM evaluation suites that the results are measured on.

How to use it

The source gives no code release, model weights, pricing or license terms; it describes a research result to be read and benchmarked against, not a system available to run.

How solid is it

The concrete, numeric result is DriveRL's mean score of 93.57 on nuPlan across the Val14, Test14-hard and Test14-random splits, stated to exceed the Log-Replay expert on all three. The claim that the combined DriveZero planner is state-of-the-art on NAVSIMv1, NAVSIMv2 and HUGSIM is qualitative in the source text: no numeric scores are given for those three benchmarks, and no comparison figures appear for any competing method beyond the single Log-Replay mention. The text also names no authors, institutions or training compute and dataset details.

Risks and caveats

DriveRL's teacher policy is trained inside a simulated, closed-loop world built by converting real driving logs, not on public roads, so nothing in the source speaks to whether behavior learned beyond human demonstrations holds up under real-world conditions. Combined with the missing numeric scores for the NAVSIM and HUGSIM results and the absence of author or institutional attribution, the state-of-the-art claim for DriveZero itself is harder to independently check than the nuPlan figure for DriveRL.