Reproduce “DexFlyWheel: A Scalable and Self-improving Data Generation Framework for Dexterous Manipulation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Kefei Zhu, Fengshuo Bai, YuanHao Xiang, Yishuai Cai, Xinglin Chen, Ruochong Li, Xingtao Wang, Hao Dong and 3 more
DexFlyWheel introduces a scalable, self-improving data generation framework for dexterous manipulation, leveraging minimal human demonstrations to iteratively produce diverse, high-quality datasets. The framework achieved an 81.9% average success rate on multi-factor generalization tasks, outperforming baselines and successfully transferring learned policies to real-world dual-arm robots.
Key insights
- The 'minor adjustment observation' — that different objects often require only small trajectory changes — enables effective object generalization through a residual reinforcement learning policy.
- An iterative 'data flywheel' mechanism allows for continuous self-improvement, where progressively better policies generate more diverse and higher-quality data for subsequent training.
- Combining strong behavioral priors from imitation learning with the adaptive capabilities of residual reinforcement learning is crucial for simultaneously achieving human-like performance and robust generalization.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.