A Framework for Modelling Local Human-Robot Interactions Based on Unsupervised Learning

This paper addresses the problem of teaching a robot interaction behaviors using the imitation learning paradigm. Particularly, the approach makes use of Gaussian Mixture Models (GMMs) to model the physical interaction of the robot and the person when the robot is teleoperated or guided by an expert. The learned models are integrated into a sample-based planner, an RRT*, at two levels: as a cost function in order to plan trajectories considering behavior constraints, and as configuration space sampling bias to discard samples with low cost according to the behaviors. The algorithm is successfully tested in the laboratory using an actual robot and real trajectories examples provided by an expert.