Robot Programming by Demonstration with Local Human Correction for Assembly

Abstract

For the sake of boosting the usability of robots in assembly applications, Programming by Demonstration (PbD) has been regarded as a feasible solution to transfer human worker’s skills to robots. With one or more human assembly demonstrations, an assembly policy, which serves to generate a robot motion trajectory given a specific target, can be learned and then used to generate a smooth robot assembly trajectory. However, when it comes to complex and subtle assembly tasks that are widely spreading in industries, the success rate of the learned policy is limited by physical constraints introduced by the human demonstrations, the robot and the task. Motivated by this problem, this work proposes a robot PbD method of the human-in-the-loop framework that draws support from a human operator to reinforce the learned assembly policy. The Gaussian Mixture Model is utilized to pre-structure the policy and it is adjusted by requesting the human operator to provide corrections at critical positions, where unexpected collisions occur. A kinesthetic teaching experiment has demonstrated the effectiveness of the proposed method.