Kinesthetic learning of behaviors in a humanoid robot

Abstract

This work presents an approach for learning of behaviors by kinesthetic teaching in a humanoid robot. The approach enables the robot to improve and reproduce a specific behavior incrementally every time a new teaching trial is provided, and therefore, it is more suitable for real-world human-robot interaction. The algorithm consists of projection of motion data to a latent space and description of motion data in a Gaussian Mixture Model (GMM). The latent space and GMM can be refined incrementally after each kinesthetic teaching. The number of components in the GMM is adjusted accordingly in a real-time manner. Experiments with a Nao humanoid robot show the feasibility of the approach. We demonstrate that the robot can reproduce learned behaviors well through continuous kinesthetic trials.