Time Regulation and Stage Concatenation for Optimized Probabilistic Dynamic Movement Primitives

This article enhances the multitrajectory imitation learning algorithm through three aspects: data sources, model performance, and generalization expansion. To address the challenge of temporal misalignment in demonstrated trajectories from multiple sources, a hybrid approach integrating dynamic time warping and template matching algorithms is adopted to guarantee congruent temporal traits. Subsequently, a modified regular system is incorporated into the probabilistic dynamic movement primitives framework, enhancing trajectory fitting within specified time intervals. Additionally, a concatenation mechanism is devised to mitigate discontinuities stemming from sequential strategy implementation, enhancing the smoothness and stability of trajectories at junction points. Experimental evaluations conducted on robotic systems demonstrate the effectiveness of the proposed methods in improving motion fidelity, minimizing fitting errors, and maintaining computational efficiency. These findings contribute to advancing imitation learning strategies for complex robotic motion tasks.