A skill learning approach based on dynamic movement primitives and quadratic-neural energy functions

In this paper, we propose a skill learning method based on the combination of energy change and trajectory optimization. First, we propose a novel quadratic-neural energy function (QNEF) to achieve a unified characterization of multiple skill features from demonstrations. Second, the trajectories are segmented using QNEF and its gradient to generate multi-layer energy sequences, which enables accurate segmentation of non-specific trajectories and supports spatio-temporal alignment through Global Time Warping (GTW). In addition, inspired by natural energy systems, we formulate the energy function as a coupling term and integrate it into dynamic movement primitives (DMPs) to construct quadratic-neural energy function dynamic movement primitives (QNEF-DMPs). The proposed method autonomously adjusts trajectories based on energy levels while preserving trajectory features, enabling continuous obstacle avoidance. Moreover, the visualization of the energy field enhances both intuitiveness and physical interpretability. Finally, the effectiveness of the method is demonstrated through practical experiments on the ROKAE robot platform.