Simulating object deformations is a critical challenge in many scientific domains, with applications ranging from robotics to materials science. Learned Graph Network Simulators (GNSs) are an efficient alternative to traditional mesh-based physics simulators. Their speed and inherent differentiability make them particularly well-suited for inverse design problems such as process optimization. However, these applications typically offer limited available data, making GNSs difficult to use in real-world scenarios. We frame mesh-based simulation as a meta-learning problem and apply conditional Neural Processes to adapt to new simulation scenarios with little data. In addition, we address the problem of error accumulation common in previous step-based methods by combining this approach with movement primitives, allowing efficient predictions of full trajectories. We validate the effectiveness of our approach, called Movement-primitive Meta-MeshGraphNet (M3GN), through a variety of experiments, outperforming state-of-the-art step-based baseline GNSs and step-based meta-learning methods.
Keywords: Graph Network Simulators, Graph Neural Networks, Meta-Learning, Neural Processes, Deformable Object Simulation, MeshGraphNets
TL;DR: We introduce a latent task-specific Graph Network Simulator, which improves over existing learned simulators by framing mesh-based simulation as a meta-learning problem.
Abstract:
Supplementary Material: zip
Primary Area: learning on graphs and other geometries & topologies
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 6180
Loading