A Structured Prediction Approach to Robot Imitation Learning

This thesis is primarily focused on movement primitives-based imitation learn- ing, within the context of robot programming by demonstration. Specifically, the imitation problem is tackled from a supervised-learning perspective. Therefore, it allows us to resort to theoretical tools from structured prediction, which can handle data-sets with complex structures. The first part of the thesis provides an overall background, in which we overview state-of-the-art imitation learning algorithms as well as discuss relevant technical tools. We formally introduce our contribution in part II. Our algorithm is not only capable of learning usual Euclidean trajectories (Chapter 7), but also trajectories lying on some manifold (Chapter 8). The capability of adapting manifold trajectories distinguishes our approach from other imitation learning algorithms. Subsequently, we provide a few extensions to augment our approach, including trajectory refinement by policy search (Chapter 10), imitation learning with constraints (Chapter 11), and probabilistic trajectory transfer (Chapter 12). We then conclude the thesis in the epilogue.