Reinforcement Learning with Latent Flow