Learning and Executing Primitive Skills Based on Adaptive Control

Abstract

In this paper, we focus on the tasks learning and executing problem for manipulators. Dynamic movement primitives (DMPs) are employed as the basic motion models, which enable the robot to learn skills from human kinesthetic teaching. A stage teaching strategy is proposed to improve the generality of the framework, such that complex tasks for multi-joint manipulators can be learned. A DMPs joining method is integrated to concatenate complex movement sequences with smooth and accurate transitions in position and velocity space. Besides, motions can be generalized to different goals and durations. Furthermore, an adaptive controller with neural networks is introduced to improve the learning performance and to ensure the performance of motion execution.