Derivative-Free Semiparametric Bayesian Models for Robot Learning

Model-Based Reinforcement Learning (MBRL) is gaining much interest in the robot learning community; in MBRL, the model serves as a representation which is largely task-invariant, and thus can facilitate transfer of knowledge across multiple tasks in the same domain. Learning reliable models for physical systems, however, remains a challenging problem. This paper summarizes recent semiparametric and derivative-free modelling techniques, and presents some key points for a new methodology to formulate derivative-free semiparametric Bayesian models with applications to robot learning. The modeling technique is demonstrated using a real robotic system, and is shown to consistently perform better than other state-ofthe-art techniques.