Learning control policies from constrained motion

Many everyday human skills can be framed in terms of performi ng some task subject to constraints imposed by the task or the environment. Co nstraints are usually unobservable and frequently change between contexts. In this thesis, we explore the problem of learning control po licies from data containing variable, dynamic and non-linear constraints on mo tion. We show that an effective approach for doing this is to learn the unconstraine d policy in a way that is consistent with the constraints. We propose several novel algorithms for extracting these po licies from movement data, where observations are recorded under different cons trai t . Furthermore, we show that, by doing so, we are able to learn representations o f movement that generalise over constraints and can predict behaviour under new c onstraints. In our experiments, we test the algorithms on systems of vary ing size and complexity, and show that the novel approaches give significant impr ovements in performance compared with standard policy learning approaches that are naiv to the effect of constraints. Finally, we illustrate the utility of the approac hes for learning from human motion capture data and transferring behaviour to several r obotic platforms.