Tackling the Correspondence Problem Closed-Form Solution for Gesture Imitation by a Humanoid's Upper Body

摘要

Learning from demonstrations (LfD) is receiving more attention recently as an important modality for teaching robots and other agents new skills by untrained users. A successful LfD system must tackle several problems including the decision about what and whom to imitate but, ultimately, it needs to reproduce the skill it learned solving the how to imitate problem. One promising approach to solving this problem is using Gaussian Mixture Modeling and Gaussian Mixture Regression for reproduction. Most available systems that utilize this approach rely on kinesthetic teaching or require the attachment of special markers to measure joint angles of the demonstrator. This bypasses the correspondence problem which is accounting for the difference in the kinematic model of the demonstrator and the learner. This paper presents a closed-form analytic solution to the correspondence problem for an upper-body of a humanoid robot that is general enough to be applicable to many available humanoid robots and reports the application of the method to a pose copying task executed by a NAO robot using Kinect recorded data of human demonstrations.