Task adaptation through exploration and action sequencing

Abstract

General-purpose autonomous robots need to have the ability to sequence and adapt the available sensorimotor knowledge, which is often given in the form of movement primitives. In order to solve a given task in situations that were not considered during the initial learning, it is necessary to adapt trajectories contained in the library of primitive motions to new situations. In this paper we explore how to apply reinforcement learning to modify the subgoals of primitive movements involved in the given task. As the underlying sensorimotor representation we selected nonlinear dynamic systems, which provide a powerful machinery for the modification of motion trajectories. We propose a new formulation for dynamic systems, which ensures that consecutive primitive movements can be splined together in a continuous way (up to second order derivatives)