MeshGraphNetRP: Improving Generalization of GNN-based Cloth Simulation

Abstract

Deep learning-based cloth simulation approaches have potential in achieving real-time simulation of complex cloth by directly learning a mapping from control input to resulting cloth movement, bypassing the need for time-consuming dynamic solving and collision processing. Recent advancements have demonstrated the effectiveness of Graph Neural Networks (GNN) in learning cloth dynamics. However, existing GNN-based models have limitations in predicting scenarios involving complex cloth movement. To overcome this limitation, we propose a novel GNN-based model that incorporates several components, including RNN-based state encoding and physics-informed features. Our model significantly improves the accuracy of cloth dynamics prediction in various scenarios, including those with complex cloth movement driven by control handles. Furthermore, our model demonstrates generalization capabilities for cloth mesh topology and control handle configurations. We validate the effectiveness of our approach through ablation studies and comparisons with a baseline model.