Event

PlayPolicy: Learning Dexterous Manipulation from Human Play Data

1 reports · 1 sources

Botao He and other authors released the PlayPolicy dexterous manipulation learning framework: the robot first learns a play policy via behavior cloning from several minutes of human play data, then autonomously deploys to collect hours of interaction data to train a world model; at inference time, this policy proposes candidate action sequences for model predictive control. The authors claim that, in simulation and real-world environments, this method achieves a 69.6% higher success rate than hindsight behavior cloning, and reaches success rates comparable to reinforcement learning with 490 times less interaction data, without any task-specific demonstrations or sim-to-real transfer. The above results are all claims made by the authors in the paper.