← ArchivePaper2023

MeshGraphNetRP: Improving Generalization of GNN-based Cloth Simulation

Emmanuel Ian Libao, Myeongjin Lee, Sumin Kim, Sung-Hee Lee

Improved graph neural network approach for cloth simulation with better generalization to unseen garment shapes and body poses.

Open the paper ↗Trace in the lineage graph

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.

How to read this

Category: Method: a GNN-based learned cloth simulator
Contributions: A GNN cloth model adding RNN-based state encoding and physics-informed features to improve prediction of complex cloth movement
Improved accuracy on scenarios with control-handle-driven cloth motion versus a baseline
Demonstrated generalization across cloth mesh topology and control-handle configurations, supported by ablations
Context: Builds on graph-neural-network cloth dynamics and neural cloth simulation, targeting the weak generalization of prior GNN models on complex motion.Builds on: Neural Cloth Simulation
Correctness: Generalization claims rest on the added recurrent encoding and physics-informed inputs and are evaluated against a baseline via ablation; coverage of garment shapes, poses, and handle setups outside the tested range remains the key uncertainty.
Clarity: Reasonably accessible; a first pass conveys which components were added and why, a second pass is needed for the message-passing and state-encoding details.
How to read it: First pass for the added components and the generalization framing; second pass on the RNN state encoding and physics-informed features if you plan to extend the GNN.

Builds on

Neural Cloth Simulation 2022

Built upon by

Nothing yet.

Related work

D-Cloth: Skinning-based Cloth Dynamic Prediction with a Three-stage Network 2023 / CGF
Dynamic Deformables: Implementation and Production Practicalities 2020 / SIGGRAPH Courses
Cloth and Skin Deformation with a Triangle Mesh Based Convolutional Neural Network 2020 / CGF
Detail-Aware Deep Clothing Animations Infused with Multi-Source Attributes 2023 / CGF

Keywords

cloth simulation graph neural networks meshgraphnets data-driven simulation physics-informed features rnn encoder gru encode-process-decode message passing scheduled sampling rollout control handles mesh topology generalization kinetic energy loss bending arcsim rotation invariance

This page summarises the entry and links to its original source. The archive never hosts or redistributes the publication itself.Show it in the full archive list →