Non-parametric Models and Contextual Policy Search for More Efficient Robot Skill Generalization

Abstract

Robot skill learning is a promising alternative to hand designing controllers that encode the skill. However, most learning algorithms are highly data-inefficient and cannot generalize to unseen situations efficiently. In order to scale robot skill learning algorithms to more complex tasks we address the above challenges in this thesis. Firstly, we evaluate the contextual extension of the Relative Entropy Policy Search (REPS) algorithm on complex robot learning tasks. Secondly, we propose a model-based extension of the contextual REPS algorithm to address data-efficiency. We investigate how Gaussian Process (GP) models can be used for model learning and trajectory prediction, and how we can incorporate models in the REPS learning framework. We present a novel probabilistic model learning framework with a special focus on learning models for long-term trajectory prediction. Finally, we investigate how model learning and trajectory prediction can be evaluated in feature spaces with kernel embedding of conditional distributions.