Robot Learning from Human Demonstration of Peg-In-Hole Task

Abstract

Human beings can adapt to new complex tasks with high accuracy in less time. In this sense, if a robot can learn from demonstrations of human task strategies, then it will greatly improve the level of adaptation of the robot at work. To achieve this, we propose a framework for the users to teach the robot task skills from the demonstrations, and the behavior of the robot is encoded by probabilistic model, impedance system and stiffness estimate. Impedance control is widely used in complex tasks to obtain the desired dynamics between tools and environments. We estimate the stiffness from the demonstrations to avoid the manual adjustments of impedance parameters, which allows the robot to use the optimal impedance parameters for each task. Specifically, the proposed method learns the impedance behavior and the trajectory following skills simultaneously. A series of peg-in-hole assembly experiments on Barrett WAM robot are provided to verify the effectiveness of the proposed learning method.