Exploration in structured space of robot movements for autonomous augmentation of action knowledge

Imitation learning has been proposed as the basis for fast and efficient acquisition of new sensorimotor behaviors. Movement representations such as dynamic movement primitives were designed to enable the reproduction of the demonstrated behaviors and their modulation with respect to unexpected external perturbations. Various statistical methods were developed to generalize the acquired sensorimotor knowledge to new configurations of the robot's workspace. However, statistical methods can only be successful if enough training data are available. If this is not the case, usually the teacher must provide additional demonstrations to augment the database, thereby improving the performance of generalization. In this paper we propose an approach that enables robots to expand their knowledge database autonomously. Efficient exploration becomes possible by exploiting the structure of the search space defined by the previously acquired example movements. We show in real-world experiments that this way the robot can expand its database and improve the performance of generalization without the help of the teacher.