Research and Implementation of Complex Task Based on DMP

Abstract

Aiming at the problem of poor task versatility in the strategy learning of complex tasks and cannot be generalized to obstacle environments, a method for robots learning from demonstration based on Dynamic Movement Primitive (DMP) is proposed. This method divides the demonstration task into several subtasks, and uses DMP to model each subtask. We utilize the local weighted regression algorithm to learn the corresponding weight of each subtask. Then the task trajectory is obtained by sequentially combining the generated subtask trajectory. The obstacle avoidance coupling item is added to the DMP framework to generate the obstacle avoidance task trajectory. The UR5 robotic arm is used to perform grasp and place tasks on the simulation platform as a demonstration task to verify the method in this article.