Effective and Incremental Learning from Demonstrations for Peg-in-hole Assembly

Develop data-efficient and incremental learning-from-demonstrations methods for peg-in-hole assembly that improve adaptation to environmental changes and generalization to new situations, including principled strategies for feature extraction and demonstration selection.

Background

The paper reviews LFD techniques (e.g., DMPs, GMR/GMM, HMM) for encoding and reproducing assembly skills, noting limitations in trajectory-level generalization and robustness under uncertainty.

The open question emphasizes improving data efficiency and incremental learning capabilities—such as selecting high-quality demonstrations and extracting robust features—to achieve better adaptation and generalization in real-world assembly.