Learning Long-term Visual Dynamics with Region Proposal Interaction\n Networks

Learning long-term dynamics models is the key to understanding physical\ncommon sense. Most existing approaches on learning dynamics from visual input\nsidestep long-term predictions by resorting to rapid re-planning with\nshort-term models. This not only requires such models to be super accurate but\nalso limits them only to tasks where an agent can continuously obtain feedback\nand take action at each step until completion. In this paper, we aim to\nleverage the ideas from success stories in visual recognition tasks to build\nobject representations that can capture inter-object and object-environment\ninteractions over a long-range. To this end, we propose Region Proposal\nInteraction Networks (RPIN), which reason about each object's trajectory in a\nlatent region-proposal feature space. Thanks to the simple yet effective object\nrepresentation, our approach outperforms prior methods by a significant margin\nboth in terms of prediction quality and their ability to plan for downstream\ntasks, and also generalize well to novel environments. Code, pre-trained\nmodels, and more visualization results are available at https://haozhi.io/RPIN.\n