Executive Summary : | Differentiable simulation of deformable objects like cloth is increasingly important in various applications, including fabrication, computer-aided design, computer vision, and robotic manipulation. Differentiable simulators provide gradients of simulated dynamics, allowing for optimization, inference, and learning. However, cloth simulation in computer graphics has built upon sophisticated numerical techniques such as adaptive remeshing, yarn-level homogenization, implicit frictional contact, and multiscale modeling. This project aims to design a differentiable cloth simulation framework that incorporates state-of-the-art advances in cloth simulation, focusing on adaptive remeshing and multiscale elasticity models. The project will undertake a systematic study of smoothness and differentiability issues arising from these techniques in cloth simulation. Recent advances in approaches for related issues, such as differentiable remeshing and smoothed contact models, can provide insights for adding differentiability to advanced cloth simulation techniques. The project will build upon existing cloth simulation libraries developed by the PI and collaborators to rapidly develop the differentiable simulator. The resulting simulator will be validated through synthetic experiments modeling various tasks in computational fabrication, computer vision, and robotics.
If successful, the project will advance fundamental understanding in differentiable simulation and close the gap between modern cloth simulation techniques used in computer graphics and existing differentiable simulators used in vision and robotics applications. It will also provide an efficient and accurate tool for gradient-based optimization of cloth dynamics, potentially enabling new applications in various domains. |