Reward-Weighted GMM and Its Application to Action-Selection in Robotized Shoe Dressing

Abstract

In the context of assistive robotics, robots need to make multiple decisions. We explore the problem where a robot has multiple choices to perform a task and must select the action that maximizes success probability among a repertoire of pre-trained actions. We investigate the case in which sensory data is only available before making the decision, but not while the action is being performed. In this paper we propose to use a Gaussian Mixture Model (GMM) as decision-making system. Our adaptation permits the initialization of the model using only one sample per component. We also propose an algorithm to use the result of each execution to update the model, thus adapting the robot behavior to the user and evaluating the effectiveness of each pre-trained action. The proposed algorithm is applied to a robotic shoe-dressing task. Simulated and real experiments show the validity of our approach.