A Appendix

We demonstrate the ability of DGRL to learn factorial representations on real world robot data, where the robot arm moves in presence of background video distractors [28]. Further details on the robot arm data collection are provided below. The data contains rich temporal background noise. We first learn a representation with a simple auto-encoder, following by the discretization bottleneck of DGRL, and then reconstruct the image with different discrete factors. Figure 11 demonstrates factorization in the learnt representation tweaking the different factors used in the discretization bottleneck. In particular, we observe some form of "compositionality" emerging as the decoder was never trained on some of the combinations of factors, for instance (person in the background + orange lamp) and (person in the background + arm to the left).