Dexterity from Supervised Play
DSP uses supervised play to collect robot data, plan with a world model, and match RL with far less interaction.
Dexterity from Supervised Play (DSP) trains a play policy from minutes of human play via behavior cloning, then uses it to gather hours of autonomous interaction data for a world model. At inference the play policy proposes action sequences scored by model-predictive control, and a goal-conditioned policy is distilled through self-play inside the model. On underactuated pushing and any-to-any in-hand reorientation, real-world success exceeds hindsight behavior cloning by 69.6% and matches reinforcement learning with 490 times less interaction data, without task-specific demonstrations or sim-to-real transfer.
- DSP clones minutes of human play, then collects hours of data.
- Play policy proposes actions for world-model MPC.
- Closed-loop policy is distilled via self-play in the world model.
- Real-world success beats hindsight cloning by 69.6%.
- Comparable to RL with 490 times less interaction data.
From the original article
1 min read at arxiv.orgHow can robots autonomously collect the contact-rich experience needed to learn dexterous manipulation? We present DSP: Dexterity from supervised play, a framework in which a robot learns a play policy from minutes of human play through behavior cloning, then deploys it autonomously to collect hours of interaction data to train a world model.
By Botao He · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.