Optimal contact force of Robots in Unknown Environments using Reinforcement Learning and Model-free controllers

摘要

Classical robot interaction control requires knowledge of the robot and environment dynamics to design the position and force controllers and to obtain the optimal contact force (optimal desired force) off-line. When the robot and environment dynamics are unknown, model-free controllers are needed to overcome this issue, however optimal contact force is not guarantee and require an identification method that adds more complexity to the control problem. This paper uses reinforcement learning methods to learn online the optimal contact force and use model-free controllers to guarantee position and force tracking. Also it is modified the classical robot interaction control to avoid the transformation from joint space to task space using Jacobian properties. Experiment results are presented to verify our approach.