Dexterity from Supervised Play

arXiv 2026 34.8 method

TLDR

DSP uses human play to train a play policy that collects data, trains a world model, and guides MPC for dexterous manipulation, outperforming baselines in real-world experiments.

Reasoning

Strong empirical evaluation in simulation and real world with clear gains over hindsight behavior cloning and data-efficiency gains over RL. Weakness: limited detail in abstract about world model architecture and generalization; no discussion of failure cases.

Read-first score

Read-first score 34.8, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 39.

Recency 5%

100

Uses a gentle age decay so recent papers surface without erasing older foundations. 2026

Methodology quality 15%

Screens visible abstract and analysis fields for experiment, dataset, baseline, metric, and limitation evidence. markers=evaluation,experiment,result

Topical relevance 25%

55.7

Uses existing LLM keyword relevance scores normalized to 0-100. world model,world simulator,generative world model,interactive world model,video world model,world dynamics prediction,model-based reinforcement learning world model

Reproducibility 15%

Screens links and visible text for paper, code, dataset, artifact, and repository signals. pdf=True; code=False; dataset=False; markers=none

Graph prestige 15%

5.7

PageRank-style prestige over the local citation graph built from available OpenAlex references. local_pagerank=0.0016

Citation impact 15%

Uses OpenAlex-shaped citation metadata as a bibliometric attention signal, separate from paper quality. cited_by_count=0

Citation velocity 10%

Citation velocity estimates citations per publication-year to reduce old-paper bias. velocity=0.00

Field roles

FrontierMethodology anchor

Rank sensitivity

Stability: volatile; rank range: 292.

Keyword Scores

world model

world dynamics prediction

world simulator

model-based reinforcement learning world model

interactive world model

generative world model

video world model

Deep Analysis

Innovations

DSP framework learns a play policy from minutes of human play via behavior cloning and uses it to autonomously collect hours of interaction data for world model training.
The learned play policy guides model-predictive control by proposing candidate action sequences for evaluation under the learned world model.
A method for distilling a closed-loop, goal-conditioned policy through self-play within the learned world model.
Demonstrates supervised play enabling dexterous manipulation without task-specific demonstrations or sim-to-real transfer.

Methodology

DSP trains a play policy from minutes of human play via behavior cloning, then deploys it autonomously to collect hours of interaction data to train a world model. At inference, the play policy proposes candidate action sequences for model-predictive control under the learned world model, and a goal-conditioned policy is distilled through self-play inside the world model. The framework is evaluated on underactuated pushing and any-to-any in-hand pose reorientation with a multifingered hand in simulation and real world.

Key Results

In real-world experiments, DSP outperformed hindsight behavior cloning by 69.6% in success rate and achieved success comparable to reinforcement learning while using 490 times less interaction data.

Resources

Paper