Imitation Learning in Industrial Robots: A Kinematics based Trajectory Generation Framework

This paper presents a simplified approach of imitation learning for an industrial robot. The approach utilizes a teleoperation based trajectory planner to generate an end-effector trajectory through direct imitation of the human motion. The adapted planner exploits the features of the human arm kinematic model and the motion tracking system to achieve real time imitation for trajectory generation. In addition, a trajectory generalization framework, based on clustering and the closest point search is also proposed. This generic framework retrieves an optimal trajectory by utilizing all the demonstrations of the task. The approach is verified experimentally on five degrees of freedom industrial robot for a manufacturing application, where a precise trajectory is desired for execution. The experimental results reflect that the proposed approach provides an effective way to teach robots from human task demonstrations.