Reproduce “A System for General In-Hand Object Re-Orientation” 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

Tao Chen, Jie Xu, Pulkit Agrawal

Researchers at the Massachusetts Institute of Technology developed a system for general in-hand object reorientation using a multi-finger robotic hand, demonstrating success across diverse objects and challenging scenarios including downward-facing manipulation without explicit object shape knowledge. The system achieved over 90% success on the EGAD dataset for upward-facing reorientation and improved downward-facing success rates from 0% to over 70% with a gravity curriculum.

Key insights

  • High success rates in complex in-hand reorientation can be achieved without explicit knowledge of an object's precise 3D shape, suggesting that manipulation can be more shape-agnostic than previously assumed.
  • The learned policies exhibited emergent dexterous skills, such as throwing and catching objects for reorientation, highlighting the potential of RL for discovering novel manipulation strategies.
  • Curriculum learning and stable initialization are effective strategies for overcoming hard exploration problems in RL for tasks involving intrinsic instability, particularly when manipulating objects against gravity.

Summary by alphaXiv.

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