Driving · 2026-11-10

DreamVLA: From Generated Human Motions to Humanoid Vision-Language-Action Policies

Kevin Park, Izat Rasyad, Dvij Kalaria, Shankar Sastry

arXiv:2611.10872PDF

AI-assisted summary based on the paper. Consult the linked source for methods and limitations. Benchmark numbers appear after editor verification.

TL;DR

DreamVLA trains humanoid vision-language-action policies in simulation without new task-specific teleoperation data. It generates human motion trajectories (via DreamControl) and turns them into a residual RL expert on a SONIC backbone that produces physically executed interaction demonstrations. A reference-conditioned DAgger distillation collects only successful expert recovery segments from learner failures using the same reference suffix, then fine-tunes GR00T on the aggregated dataset.

Why it matters

The authors target a bottleneck in whole-body humanoid VLA training: collecting large amounts of teleoperation demonstrations. DreamVLA replaces teleoperation with generated motion trajectories plus residual RL to produce physically valid interaction data, and addresses distribution shift by aggregating successful corrective recoveries conditioned on the original generated reference suffix.

Method

  • Generate task-conditioned motion references with DreamControl/OmniControl (with spatiotemporal constraints), retarget to Unitree G1, and augment contact points with HumanoidMimicGen.
  • Train a residual RL expert on a SONIC motion-token backbone to co-track humanoid references and object outcomes under randomized physics; filter rollouts by task success and export successful episodes to LeRobot format.
  • Distill into a VLA by reference-conditioned recovery DAgger: after learner failure, restart the RL expert from a sampled learner state with the same reference suffix and aggregate only successful recovery segments to fine-tune GR00T action-chunk training.

Limitation

The evaluation is limited to four simulated interaction tasks on a single humanoid embodiment. Generated motions can be incompatible with robot kinematics or contact requirements, and task-specific guidance, filtering, object references, and reward weights remain necessary parts of the pipeline. Successful-rollout filtering may underrepresent difficult configurations.

Abstract (from arXiv)

Training vision-language-action (VLA) policies typically depends on substantial robot demonstration data, making human teleoperation a costly and cumbersome bottleneck, particularly for whole-body humanoid interaction. Reinforcement learning in simulation offers an alternative for scaling this data, but designing task-specific rewards remains challenging. To address this, we propose DreamVLA, which uses generated motion trajectories to guide reinforcement learning (RL) in simulation, converting them into reliable task execution trajectories. These trajectories train VLA models in simulation without new task-specific teleoperation data. Our central contribution is a reference-conditioned DAgger distillation procedure: after a failed learner rollout, we restart the RL expert from a sampled learner state with the same reference suffix and aggregate successful corrective trajectories, addressing the distribution shift left by offline imitation. We evaluate picking from a platform, opening a drawer, pressing a button, and walking and picking from a platform, with ablations of DAgger iterations, residual RL, and generated dataset scale. Project website: ember-center-berkeley.github.io/DreamVLA.

Related papers

Publication record

Bibliographic evidence describes where and when a record appeared. It does not validate scientific claims.

First announced by arXiv: 2026-11-10 · Indexed by Pub-AI: 2026-11-10 UTC

preprint

DreamVLA: From Generated Human Motions to Humanoid Vision-Language-Action Policies

Kevin Park, Izat Rasyad, Dvij Kalaria, Shankar Sastry

First announced by arXiv: 2026-11-10 · Original record ↗

arxiv
2611.10872
Explicit source versions (1)
  • v1 · Date not supplied
  1. Preprint available

    Source date: 2026-11-10 · First detected: 2026-11-10 UTC · Historical baseline

    Source evidence (1)
    • Official evidence ↗ · RSS guid=oai:arXiv.org:2611.10872v1; arxiv:announce_type=new; item pubDate (official metadata)
Abstract · arxiv

Training vision-language-action (VLA) policies typically depends on substantial robot demonstration data, making human teleoperation a costly and cumbersome bottleneck, particularly for whole-body humanoid interaction. Reinforcement learning in simulation offers an alternative for scaling this data, but designing task-specific rewards remains challenging. To address this, we propose DreamVLA, which uses generated motion trajectories to guide reinforcement learning (RL) in simulation, converting them into reliable task execution trajectories. These trajectories train VLA models in simulation without new task-specific teleoperation data. Our central contribution is a reference-conditioned DAgger distillation procedure: after a failed learner rollout, we restart the RL expert from a sampled learner state with the same reference suffix and aggregate successful corrective trajectories, addressing the distribution shift left by offline imitation. We evaluate picking from a platform, opening a drawer, pressing a button, and walking and picking from a platform, with ablations of DAgger iterations, residual RL, and generated dataset scale. Project website: ember-center-berkeley.github.io/DreamVLA.