Learning Future Representation with Synthetic Observations for Sample-efficient Reinforcement Learning