LAMPS: A Novel Robot Generalization Framework for Learning Adaptive Multi-Periodic Skills

Abstract

Learning from Demonstrations (LfD) methods are applied to transfer human skills to robots from expert demonstrations, enabling them to perform complex tasks. However, existing methods often struggle to handle such long-horizon human skills as cleaning or wiping stains on the surface, which involve multiple periodic and transitional movement primitives. To address this limitation, this paper proposes a novel framework for segmenting, learning, and generalizing multi-periodic human skills, enabling robots to effectively learn different movement primitives and execute these skills in new environments. Specifically, the framework introduces an unsupervised learning method to segment long-horizon human demonstrations into periodic and discrete movement primitives. Further, a novel type of discrete dynamical movement primitives, namely transitional movement primitives, is employed to enhance the fluidity of combining different periodic movement primitives in skills. These primitives collectively form a lightweight state machine during task execution, where state transitions are governed by visual perception, thereby enabling generalization to long-horizon tasks composed of arbitrary numbers of periodic subtasks. To validate the effectiveness of the proposed approach, we conduct extensive experimental evaluations, including step-by-step validation of each method in simulation and the implementation of the entire presented framework in the real world. The results confirm that the proposed framework accurately learns and generalizes multi-periodic human skills, providing a feasible solution for transferring complex multi-periodic demonstrations to robots in practical applications. The project website can be found at: https://nkrobotlab.github.io/LAMPS/