This paper presents a method for imitation learning that decouples learning an image-based goal from the controller. This decomposition in itself is not novel, but the authors incorporate the latest neural network methods to generate the goal images and learn the controller. There was discussion among the reviewers about how to properly validate and compare the proposed method; everyone appreciated the fact that the work used a physical instantiation but there was concern about the lack of baseline comparisons. However, in the end, the PC concluded that the manuscript along with the latest revisions should be presented at NeurIPS.