Reviews: Interaction Networks for Learning about Objects, Relations and Physics

The paper is very well-written, experimentally and theoretically sound, and generally easy to comprehend and nicely self-contained. The proposed model is (as far as I can tell) genuinely novel and fulfills the exciting promise of a general-purpose, learnable physics engine with impressive results. I have only a few remarks, mostly concerning the paper's presentation: 1) In my regard, there are some issues with the introduction of the paper, which seems a bit unfocused. While the model is excellently described in its definition of page 3, the introduction appears too abstract in order to provide a foundation for the technical details and referencing Figure 1 before the model definition appears rather confusing. To fix this, a short general description of the data stream through the network might be helpful, as well as a short remark pointing towards section 2. This could also help to convey how the model is a combination of structured models, simulation and deep learning (line 25).