DreamVLA trains humanoid policies on generated human motion instead of teleoperation

Researchers led by Kevin Park have presented DreamVLA, a way to train vision-language-action policies for humanoid robots in simulation without new teleoperation data for each task, in a paper accepted to two CoRL 2026 workshops.

Collecting robot demonstrations by teleoperation is costly, especially for whole-body humanoid tasks, and reinforcement learning needs rewards written for each task, the authors write. DreamVLA steers reinforcement learning in simulation with generated motion trajectories and turns them into task executions that succeed, which then train the policies.

Its main step is a DAgger-style distillation: after a learner fails, the reinforcement learning expert restarts from a sampled state of the learner with the same reference and adds the successful corrections to the data. The team tested four tasks: opening a drawer, pressing a button, and picking from a platform while standing or after walking.