Robust Learning and Bayesian Inference of Riemannian Probabilistic Movement Primitives for Physical Human-Robot Synergic Interaction
Jiajun Wu, Yihui Li, Haifei Zhu, Yisheng Guan
Abstract
Although Bayesian interaction primitives exhibit strong capabilities in skill learning and reproduction for physical human–robot interactions, they require extensive demonstrations and fail to adequately address the inherent geometric constraints. In this study, we propose the iRie-ProMPs method, based on Riemannian probabilistic movement primitives, for synergic interactions. Our method incorporates a modified Expectation-Maximization algorithm for learning and a factor graph-based Bayesian inference algorithm for skill reproduction. Experiment results demonstrated that the modified Expectation-Maximization algorithm significantly improved the learning robustness, even with as few as five demonstrations. In human–robot interaction tasks, the iRie-ProMPs approach outperformed existing methods, improving trajectory inference accuracy by 32.4% and reducing pose deviation by 70.0%. The proposed method offers a promising framework for skill learning and reproduction in physical human–robot synergic interactions, particularly under limited demonstrations and geometric constraints.
BibTeX
@inproceedings{ral2025_robustlearningan,
title = {Robust Learning and Bayesian Inference of Riemannian Probabilistic Movement Primitives for Physical Human-Robot Synergic Interaction},
author = {Jiajun Wu and Yihui Li and Haifei Zhu and Yisheng Guan},
booktitle = {RA-L 2025},
year = {2025}
}