Reproduce “Learning Physics-Grounded 4D Dynamics with Neural Gaussian Force Fields” 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

Shiqian Li, Ruihong Shen, Junfeng Ni, Chang Pan, Chi Zhang, Yixin Zhu

The Neural Gaussian Force Field (NGFF) framework from Peking University and Tsinghua University learns explicit force fields over 3D Gaussian representations directly from multi-view RGB inputs, enabling the prediction of physically realistic 4D dynamics. It achieves superior physical consistency and generalization for dynamic prediction and video generation while offering significantly faster inference speeds compared to traditional physics simulators.

Key insights

  • Explicitly modeling force fields over object-centric 3D Gaussian representations allows for robust physics-grounded dynamics prediction with significantly faster inference than traditional simulators.
  • Decomposing scenes into object-aware 3D Gaussians from raw multi-view RGB inputs provides an interpretable and effective representation for learning physical interactions.
  • A two-stage training approach, leveraging both real-world visual data for reconstruction and high-fidelity synthetic physics data for dynamics, enhances both visual fidelity and physical consistency.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.