Learning Motion and Impedance Behaviors from Human Demonstrations

Human-robot skill transfer has been deeply investigated from a kinematic point of view, generating various approaches to increase the robot knowledge in a simple and compact way. Nevertheless, social robotics applications require a close and active interaction with humans in a safe and natural manner. Torque controlled robots, with their variable impedance capabilities, seem a viable option toward a safe and profitable human-robot interaction. In this paper, an approach is proposed to simultaneously learn motion and impedance behaviors from tasks demonstrations. Kinematic aspects of the task are represented in a statistical way, while the variability along the demonstrations is used to define a variable impedance behavior. The effectiveness of our approach is validated with simulations on real and synthetic data.