arXiv Preprint: DSP Framework Uses Human Play Demos for Efficient Dexterous Manipulation Planning
New research introduces DSP, a framework that mimics human goal-free play to autonomously collect data and guide model-predictive control, significantly outperforming traditional baselines in simulated cube reorientation tasks.
ImportanceMaterialEvidenceE2 unreplicatedWrite-upQuick
Researchers have proposed a new framework called DSP (Dexterity from Supervised Play) to address how robots can autonomously collect contact-rich experience for learning dexterous manipulation. The method first learns a general policy from minutes of human goal-free 'play' demonstrations via behavior cloning, then deploys it autonomously to collect hours of interaction data for training a world model. During inference, this play policy acts as an action prior, guiding Model-Predictive Control (MPC) to evaluate candidate action sequences.
In simulated multifingered hand cube reorientation tasks, DSP achieved a 91.4% success rate, surpassing Hindsight Behavior Cloning (53.9%) trained on the same data by 37.5 percentage points. Additionally, DSP required only 701K interaction frames to match the performance of PPO reinforcement learning trained on 344M frames, representing approximately 490 times less interaction data. These comparisons are based on counted simulation frames, not elapsed time or hardware data.
The paper notes that hardware experiment results are currently considered 'provisional,' requiring a frozen protocol and uncertainty estimates, making the simulated comparison the decisive quantitative evidence. The core contribution lies in transforming human play into autonomous data collection and planning priors, avoiding the need for task-specific demonstrations or sim-to-real transfer.