Learning Simulatable Models of Cloth with Spatially-varying Constitutive Properties for Robotics

Guanxiong Chen, Shashwat Suri, Yuhao Wu, Ganidhu K. Abeysirigoonawardena, Etienne Vouga, David Levin I.W., Dinesh K. Pai

Keywords: differentiable simulation, surrogate modelling, fabric modelling

Abstract: Real fabrics exhibit complex spatial variation from stitching, hemming, and other processes. Simulating these with finite element methods is computationally demanding, and suffers from membrane locking artifacts that make cloth artificially stiff. We introduce Mass-Spring Net, a learned, simulatable model that can model complex, spatially-varying fabric behavior from motion observations alone. Our approach accurately models spatially varying properties, is robust to membrane locking, and can potentially enable fast fabric manipulation in robotics. Compared to prior work our method achieves much-faster training time, resists membrane locking that exist in synthetic training data, and early results show that it maintains high accuracy and good generalization to novel scenarios.

Serve As Reviewer: ~Guanxiong_Chen1

Submission Number: 19