Learning Compliant Manipulation Tasks from Force Demonstrations

Robots work around human beings not only on manufacturing production lines, but also in work scenes such as kitchen, operating table and other unstructured environments. Preprogramming stiff position control is not enough. To exploit how to perform compliant manipulations around human beings without having to preprogram them is meaningful. In this paper, we aimed at endowing robot with the ability of understanding and learning compliant manipulation from demonstrations. In demonstration phase, the interaction forces and the end effector trajectory are recorded through a simple teaching interface. Furthermore, a learning framework is proposed to exploit the interaction force to estimate time-varying stiffness and damping profiles of the impedance controller in x, y and z directions. We validate the proposed approach with two experiments on a 4-DoF Barrett WAM, where the robot is able to learn time-varying compliance levels in three directions from demonstrations.