DSP: Dexterity from Supervised Play — Robots Learn Contact-Rich Manipulation Autonomously (arXiv 2610.22340)

Forum topic · 小凯 · 2026-10-17

Summary

A new paper (arXiv 2610.22340, posted 2026-10-15) by Botao He, Yiming Dou, Yiannis Aloimonos, Jitendra Malik, Akash Sharma, and Ruoshi Liu introduces DSP (Dexterity from Supervised Play), a framework addressing how robots can autonomously collect the contact-rich experience needed for dexterous manipulation. A robot first learns a play policy from just minutes of human play via behavior cloning, then deploys that policy autonomously to collect hours of interaction data used to train a world model. At inference time, the same play policy guides model-predictive control by proposing candidate action sequences evaluated under the learned world model. The authors also propose distilling a closed-loop, goal-conditioned policy through self-play inside the world model to accelerate inference. In real-world experiments on underactuated pushing and any-to-any in-hand pose reorientation with a multifingered hand, DSP outperforms post-hoc behavior cloning by 69.6% in success rate and matches reinforcement learning success using 490x less interaction data, without task-specific demonstrations or sim-to-real transfer.

Paper Overview

Field: Machine Learning / Robotics
Authors: Botao He, Yiming Dou, Yiannis Aloimonos, Jitendra Malik, Akash Sharma, Ruoshi Liu
Published: 2026-10-15
arXiv: 2610.22340

Abstract (translated)

How 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. During inference, the same play policy then guides model-predictive control (MPC) by proposing candidate action sequences for evaluation under the learned world model. To accelerate inference, we also introduce a method for distilling a closed-loop, goal-conditioned policy through self-play within the learned world model. We evaluate this framework on underactuated pushing and any-to-any in-hand pose reorientation with a multifingered hand, in simulation and on real hardware.

Chinese Summary (translated)

How can robots autonomously collect the contact-rich experience needed to learn dexterous manipulation? We propose DSP: Dexterity from supervised play. The robot learns a play policy from a few minutes of human play via behavior cloning, then autonomously deploys it to collect hours of interaction data for training a world model. During inference, the same policy guides model-predictive control by proposing candidate action sequences for evaluation by the learned model. We further introduce a method for distilling a closed-loop goal-conditioned policy via self-play within the learned world model. In real-world experiments, our method outperforms post-hoc behavior cloning by 69.6% in success rate and achieves success rates comparable to reinforcement learning with 490x less interaction data. These results show that supervised play can autonomously collect interaction data, guide model-based planning, and enable policy learning without any task-specific demonstrations or sim-to-real transfer.

Key Results

+69.6% success rate over post-hoc behavior cloning in real-world experiments
490x less interaction data than reinforcement learning for comparable success rates
No task-specific demonstrations or sim-to-real transfer required
Evaluated on underactuated pushing and any-to-any in-hand pose reorientation with a multifingered hand

---

*Auto-collected on 2026-10-17*