Robot Policy Transfer with Online Demonstrations: An Active Reinforcement Learning Approach