On-Robot Bayesian Reinforcement Learning for POMDPs