Using Reward-Weighted Imitation for Robot Reinforcement Learning