The human-in-the-loop learning system considered in the research, where the behavior of the human influences the behavior of the robot and, simultaneously, the behavior of the robot influences the behavior of the human. After a number of interactions, the learning system queries the memory for a new set of training data. The data are then projected onto a low-dimensional manifold using dimensional reduction techniques. Once the state vectors are projected onto a low dimensional manifold, we group the resulting points into sets according to the action performed in that state. Thus, we obtain for each possible action a set of states in which the corresponding action should be triggered. For each action, a GMM is learned. The model encodes a probability density function of the learned state vectors. In the next round of interaction, the action corresponding to the GMM with the highest likelihood is executed by the robot.
© Copyrights Shuhei Ikemoto. All Rights Reserved