Manipulation of granular materials by learning particle interactions

Manipulation of granular materials such as sand or rice remains an unsolved problem due to challenges the difficulty defining their configuration modeling and particles interactions. Current approaches tend simplify material dynamics omit interactions between particles. In this paper, we propose use a graph-based representation model interaction rigid bodies manipulating it. This allows planning manipulation trajectories reach desired material. We graph neural network (GNN) particle via message-passing. To plan trajectories, minimise Wasserstein distance predicted distribution configuration. demonstrate that proposed method is able pour into both in simulated real scenarios.