Keypoints into the Future

Self-Supervised Correspondence in Model-Based RL