Reproduce “From a Single Demonstration to a General Policy for Contact-Rich 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
Xing Li, Oliver Brock
The authors present a robot policy that follows surfaces, joints, and changes in contact rather than replaying a demonstrated path. The system identifies environmental constraints through autonomous exploration and can incorporate targeted human corrections. On an unseen green puzzle, it completed the internal-pin sequence in 10/10 trials, while one-shot Trajectory Transfer completed 0/10, but this does not establish general manipulation.
Key insights
- Environmental Constraints (ECs) serve as an effective inductive bias in LfD, allowing robots to separate general task structure from instance-specific details and achieve strong generalization in contact-rich tasks from minimal demonstrations.
- Online adaptation of EC models through continuous sensory feedback and compliant control is critical for robustness in real-world physical interaction, enabling adaptation to imprecise initial models and unexpected variations during execution.
- A modular, interpretable policy structure based on EC primitives facilitates straightforward failure detection and targeted human corrections for out-of-distribution variations, significantly reducing the effort for policy refinement.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.