AIAny
Icon for item

DriveZero: End-to-End Driving Beyond Human Demonstrations

Trains end-to-end driving without human trajectory supervision by decoupling perception and action: DriveVFM distills multiple frozen vision foundation models into a single camera backbone, and DriveRL trains a privileged closed-loop RL teacher whose rollouts supervise a camera-only planner, yielding state-of-the-art closed-loop benchmark results.

Introduction

Driving systems that learn from human logs inherit human limitations and miss long-tail interactive scenarios. DriveZero answers this by separating "seeing" and "acting": perception is pretrained from massive, label-free visual models, while action is learned via closed-loop reinforcement learning in interactive worlds derived from real logs. The core insight is that perception should be trained to form general, reusable visual representations, whereas action requires closed-loop feedback and interaction to exceed demonstrated behavior.

Key Findings
  • DriveRL (action): converts driving logs into mixed-agent interactive worlds and trains a privileged 5.7M-parameter policy with PPO through closed-loop rollouts; the teacher can be queried under augmented goals to produce supervision not present in the original logs. This yields a planner that surpasses log-replay experts on nuPlan closed-loop splits.

  • DriveVFM (perception): consolidates multiple frozen vision foundation models (e.g., DINOv3, SigLIP2, SAM, Depth Anything V2) into a single image backbone without task-specific labels, producing representations suitable for downstream camera-only planning.

  • DriveZero (unification): distills the DriveRL teacher rollouts into a camera-only planner that learns from privileged, goal-conditioned rollouts rather than human trajectories. The resulting system achieves state-of-the-art performance on NAVSIMv1, NAVSIMv2 and the closed-loop HUGSIM benchmark and outperforms human log-replay on multiple nuPlan settings.

Who it's for and tradeoffs

Great fit if you need a research or engineering recipe to move beyond imitation learning for driving — especially when human logs are sparse in interactive, safety-critical, or long-tail scenarios. The approach reduces dependence on curated human demonstrations and leverages large frozen vision models to avoid per-task labeling.

Look elsewhere if you require end-to-end training purely from raw human logs with minimal simulation or if you cannot afford the infrastructure for mixed-agent closed-loop rollouts and RL training; DriveZero’s gains depend on building interactive rollouts and training a privileged teacher, which requires simulation/rollout tooling and RL compute.

Where It Fits

DriveZero sits between imitation-driven end-to-end pipelines and pure self-play/vectorized RL methods: it reuses real logs to construct interactive scenes (preserving realism) while using closed-loop RL and goal augmentation to cover behaviors absent from the logs. For teams with existing log data plus rollout/simulation tooling, it offers a pragmatic path to closed-loop robustness without manual labeling.

Method snapshot
  • Perception: distill multiple frozen foundation models into a driving backbone from raw images (no task labels).
  • Action: create mixed-agent interactive worlds from logs and train a privileged teacher policy with PPO; include value-guided test-time action search for improved closed-loop performance.
  • Distillation: train a camera-only planner to imitate teacher rollouts, including rollouts under augmented goals to broaden behavioral coverage.

This design emphasizes task-appropriate pretraining regimes (massive visual data for perception; closed-loop interaction for action) and uses teacher-generated rollouts to escape imitation limits while retaining camera-only deployability.

Information

  • Websitearxiv.org
  • AuthorsHao He, Chengcheng Hu, Zirun Su, Heng Zhang, Haisong Liu, Jinke Li, Haochen Tian, Zhenwei Shen, Hongyang Li, Zhichao Li
  • Published date2026/09/05

More Items

Trains a world-action model that predicts future visual states to guide zero-shot robotic manipulation; introduces CoAE, SVP, IDM and KASO to pretrain generative and action components from scratch on manipulation data, scaling up to 30,000 hours and improving zero-shot success.

Evaluates a weak teacher's RL-induced policy shift on the student's own rollouts and amplifies verifier-supported updates so stronger models can learn from weaker supervisors and surpass them. It rescales only verifier-supported policy-gradient components to preserve optimization fixed points while accelerating learning, reducing student updates versus standard RL or distillation.

Combines sparse verifier outcomes with dense privileged‑hindsight token scoring to learn an outcome‑calibrated, normalized distribution over complete responses for on‑policy self‑improvement. Key features: sign‑gated guidance (retain/reverse/disable per verifier advantage), profiled trajectory balance with one log‑partition per rollout group, and explicit correction against false‑positive self‑guidance.