Skill Transfer Between Industrial Robots by Learning from Demonstration

Abstract

Industrial robots are the key role in modern manufacture. In order to allow different robots to perform the same task, they must be programmed manually which is time-consuming and cumbersome. The ability to transfer robotic skills between heterogeneous robots would provide significances to industry. This paper presents a prototype control architecture for industrial robots which allows multiple heterogeneous robots of different morphologies to perform the same task without manually programming each robot separately. We selected three articulated industrial manipulators with different degree of freedoms (Yaskawa Motoman HP3JC: 6, Universal Robot UR5: 6 and Rethink Robotics Baxter: 7) to test our approach. Two simple tasks were chosen for implementation and analysis. Results indicate that our approach works with some errors. Maximum and average position mismatch errors for UR5 are 4.9380 cm and 4.4761 cm. For Baxter, maximum and average position mismatch errors are 1.7770 cm and 1.6068 cm. Baxter has a different range of motion therefore some places could not be reached with it.