Efficient sensorimotor learning from multiple demonstrations

Abstract

Abstract In this paper, we present a new approach to the problem of learning motor primitives, which combines ideas from statistical generalization and error learning. The learning procedure is formulated in two stages. The first stage is based on the generalization of previously trained movements associated with a specific task configuration, which results in a first approximation of a suitable control policy in a new situation. The second stage applies learning in the subspace defined by the previously acquired training data, which results in a learning problem in constrained domain. We show that reinforcement learning in constrained domain can be interpreted as an error-learning algorithm. Furthermore, we propose modifications to speed up the learning process. The proposed approach was tested both in simulation and experimentally on two challenging tasks: learning of matchbox flip-up and pouring.