Reviews: Flexible neural representation for physics prediction

The authors propose a novel hierarchical object representation based on particles to cover both rigid geometrical shapes and deformable materials. Each scene is represented as a graph, with disconnected components corresponding to the objects and the support of the scene. Each graph has a tree-like structure, where higher levels correspond to coarser scales, and the leaves correspond to the original particles placed in the object. They also propose an adapted neural network architecture, called Hierarchical Relation Network, that learns to predict physical dynamics for this representation. This multiscale approach is end to end differentiable, allowing this propagation mechanism to be learned.