Leveraging Demonstrations for Deep Reinforcement Learning on Robotics Problems with Sparse Rewards