Continuous Control with Coarse-to-fine Reinforcement Learning