Reproduce “MPMWorlds: Material-Point-Method Simulations for Inferring and Extrapolating Physical Dynamics” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Žiga Kovačič, Kevin Ellis
MPMWorlds introduces a dataset of 2D Material Point Method (MPM) simulations, pairing executable code with videos to enable rigorous comparison of generative models. The study found Vision-Language Models (VLMs) generally achieve stronger long-horizon temporal stability, while Video Diffusion Models (VDMs) excel in short-term geometric preservation but often suffer from physical inconsistencies.
Key insights
- VLMs demonstrate superior long-horizon physical stability and consistency due to their reliance on an explicit, executable physics engine, but require substantial explicit scene information for accurate inference.
- VDMs achieve better short-term geometric fidelity and local visual plausibility but struggle with long-horizon physical consistency, often showing limited utilization of explicit structured physical information.
- A simple gating mechanism based on prefix reconstruction quality can effectively combine VLM's physical stability with VDM's geometric accuracy, suggesting future hybrid architectures could be beneficial.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.