Solving New Tasks by Adapting Internet Video Knowledge

Video generative models, beyond enabling the production of astounding visual creations, offer a promising pathway for unlocking novel, text-conditioned robotic behaviors, whether utilized as a video planner or as a policy supervisor. When pretrained on internet-scale datasets, such video models intimately understand alignment with natural language, and can thus facilitate novel text-conditioned behavior generalization. At the same time, however, they may not be sensitive to the specificities of the particular environment in which a policy of interest is to be learned. On the other hand, video modeling over in-domain examples of robotic behavior naturally encodes environment-specific intricacies, but the scale of available demonstrations may not be sufficient to support generalization to unseen tasks via natural language specification. In this work, we investigate different adaptation techniques that integrate in-domain information into large-scale pretrained video models, and explore the extent to which they enable novel text-conditioned generalization for robotic tasks. Furthermore, we highlight the individual data and training requirements of each approach, which range from utilizing only a few still frames illustrating the subject of interest, to direct finetuning over videos labeled with text descriptions. We successfully demonstrate across robotic environments that adapting powerful video models with small scales of example data can successfully facilitate generalization to novel behaviors, both when utilized as policy supervisors, and as visual planners.