The trials and tribulations of training a physical robot with reinforcement learning