DreamVLA: From Generated Human Motions to Humanoid Vision-Language-Action Policies
DreamVLA: From Generated Human Motions to Humanoid Vision-Language-Action Policies
This paper introduces DreamVLA, a framework that leverages generated motion trajectories to guide reinforcement learning in simulation for training vision-language-action (VLA) policies on humanoid robots. The method addresses the data bottleneck of human teleoperation by using reference-conditioned DAgger distillation to aggregate successful corrective trajectories after failed learner rollouts. Evaluated on tasks including picking from a platform, opening a drawer, pressing a button, and walking while picking, the approach trains VLA models in simulation without requiring new task-specific teleoperation data.
The authors propose DreamVLA, a method that uses generated motion trajectories to guide reinforcement learning in simulation for training vision-language-action policies on humanoid robots, eliminating the need for new task-specific teleoperation data.
Source: arXiv Robotics — research abstracts · Read original article ↗
Loading article text…
What the source reports
Publisher-reported claims, with original evidence. These results have not been independently verified by RoboSignal.
- Environment
- SimulationOpen source S3
- paper: not_reported
- code: not_reported
- weights: not_reported
- dataset: not_reported
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
s, 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 s
Open source S3
ampled 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.
Open source S4
Source：arXiv Robotics — research abstracts · arxiv.org