Bridging Imagination and Reality for Model-Based Deep Reinforcement Learning