ROLL: Visual Self-Supervised Reinforcement Learning with Object Reasoning. https://arxiv.org/abs/2011.06777