Reproduce “Learning Multi-Step Manipulation Tasks from A Single Human Demonstration” 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

Dingkun Guo

A robot system learns complex, multi-step manipulation tasks from a single human demonstration by interpreting human actions into generalizable robot primitives based on hand-object and object-object contact dynamics. It demonstrated 20-40% overall task success rates in a dishwashing task across varied objects and environments, with individual steps achieving up to 100% success.

Key insights

  • Abstracting human demonstrations through contact dynamics (hand-object and object-object) enables data-efficient learning of manipulation tasks for robots.
  • Leveraging generalizable vision models combined with object-centric pose estimation allows for adaptability to different objects and task environments.
  • Explicitly modeling and compensating for human-robot kinematic and collision geometry differences is vital for a robot to achieve task goals without precise human motion mimicry.

Summary by alphaXiv.

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