Abstract:
Statistical Learning of Robotic Demo Trajectories Based on Multicriteria Segmentation and Multi-Demonstration Alignment (HSMM) addresses complex tasks in human-robot interaction and intelligent manufacturing. The research goal of this study is to automatically extract generalized key segments from multiple robotic demonstration trajectories in the absence of prior annotations and establish statistical and parametric models for universal trajectory reproduction across diverse tasks and conditions. To achieve this, the research tasks include multicriteria segmentation (speed, curvature, acceleration, direction change), trajectory alignment using Hidden Semi-Markov Models (HSMM), and subsequent implementation of statistical representations (ProMP, GMM/GMR, DMP). The proposed methodology begins with the smoothing of raw data and the identification of key points via topological simplification and non-maximum suppression, then, using HSMM, it ensures consistent segmentation of multiple demonstrations into characteristic segments. The conducted experiments confirm the results of the approach, demonstrating low reconstruction error while simultaneously improving data compression and preserving key actions, indicating the high efficiency of the method. Finally, the novelty and practical significance of this study can be highlighted by the potential industrial applications (such as welding, painting, etc.), as well as the future prospective expansions of the method to more dynamic and non-stationary scenarios, requiring adaptive and statistically grounded trajectory planning.
Keywords:robot learning from demonstrations, trajectory segmentation, probabilistic motion primitives, multicriteria analysis, Hidden Semi-Markov Model (HSMM).
Fulltext:
PDF file (1218 kB)