Probabilized Dynamic Movement Primitives and Model Predictive Planning for Enhanced Trajectory Imitation Learning

摘要

Recent studies of mimic learning with control methods have been predominantly confined to the single-trajectory learning framework of dynamic movement primitives. In contrast, the probabilistic movement primitives method offers notable flexibility but sacrifices the convergence of the state equation. In this article, probabilized dynamic movement primitives method in state equation representation is proposed for multitrajectory learning and challenging constraint conditions. By combining incremental learning for expanding the original trajectory distribution, while considering the inherent error of the actuator in model predictive planning to adjust the state set under the obstacle environment, this new algorithm performs well in adapting far-deviated via point and obstacle constraints. The simulations and robot experiments of trajectory tasks show that the algorithm is feasible with arduous constraint conditions, acquires great conformability with the demonstrated trajectory, while reducing the loss of free space under the end effector position error.