Reproduce “Object-Focus Actor for Data-efficient Robot Generalization 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

Yihang Li, Tianle Zhang, Xuelong Wei, Jiayi Li, Lin Zhao, Dongchi Huang, Zhirui Fang, Minhua Zheng and 2 more

Researchers from JD Explore Academy and Beijing Jiaotong University developed Object-Focus Actor (OFA), a hierarchical framework enabling dexterous robots to generalize manipulation skills to arbitrary object placements and diverse backgrounds using as few as 10 human demonstrations. The system achieves high success rates in real-world tasks by integrating object-centric perception with relative, object-focus control policies.

Key insights

  • The final phase of dexterous manipulation trajectories converges to a consistent "object-focus end trajectory" regardless of the object's initial position.
  • This consistency can be leveraged to simplify the learning problem by planning to a standardized "pre-manipulation pose" relative to the object.
  • Learning an object-centric policy using localized "hand-focus images" and relative proprioceptive states and actions enhances data efficiency and generalization.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.