Generated motion. Learned interaction.

DreamVLAFrom Generated Human Motions to
Humanoid Vision-Language-Action Policies

Kevin Park1,* Izat Rasyad2,* Dvij Kalaria2 Shankar Sastry2

1 Princeton University2 University of California, Berkeley

* Equal contribution

Read the paper Explore results

The idea

Abstract

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.

From motion to action

A look at DreamVLA

Overview video coming soon
DreamVLA pipeline: generate task-conditioned motions, train a residual RL expert, distill with corrective DAgger, and validate a humanoid VLA in simulation.
Pipeline overview

Generated human motions guide a residual RL expert. Successful demonstrations and corrective recoveries train a vision-language-action policy.

Whole-body interaction

Four tasks. One framework.

Clean third-person GR00T rollouts.
Drawer and button episodes available.

Drawer and button videos show selected successful pre-DAgger GR00T rollouts from the 200-reference models, without expert takeover, overlays, or action smoothing. The third-person camera is for visualization only; each policy uses its robot-mounted camera. These clips illustrate behavior, not success rates. Picking videos remain pending. Clip provenance ↗

Learning from more experience

What improves task success?

Reported results across all four tasks.
100 evaluation rollouts per task.

01 / Corrective learning

More DAgger rounds, stronger policies

Task success at DAgger rounds 0, 1, 2, 3: picking 0, 17, 30, 54 percent; drawer 55, 57, 54, 62; button 50, 55, 61, 72; walking and picking 0, 1, 7, 18.
Picking reaches 54% after three rounds. Starting with 200 trajectories, each round adds 200 successful corrective trajectories. The matched offline control with 800 trajectories achieves 0% on picking. Paper, Table 2 ↗
View reported values
DAgger task success (%)
Task / rounds0123
Pick from platform0173054
Open drawer55575462
Press button50556172
Walk and pick01718
02 / Data scaling

More generated data helps interaction

Task success with 10, 50, and 200 training episodes: drawer 4, 15, 55 percent; button 0, 15, 50 percent; both picking tasks remain at 0 percent.
Drawer and button success rise with more data. Generated-reference budgets of 10, 50, and 200 have matching episode counts. Both picking curves overlap at 0% throughout this study. Paper, Table 4 ↗
View reported values
Training-data task success (%)
Task / episodes1050200
Pick from platform000
Open drawer41555
Press button01550
Walk and pick000

Build on this work

BibTeX

@misc{park2026dreamvla,
  title  = {DreamVLA: From Generated Human Motions to
            Humanoid Vision-Language-Action Policies},
  author = {Park, Kevin and Rasyad, Izat and
            Kalaria, Dvij and Sastry, Shankar},
  year   = {2026}
}