Reproduce “Geometric Trajectory Diffusion Models” 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
Jiaqi Han, Minkai Xu, Aaron Lou, Haotian Ye, Stefano Ermon
Stanford University researchers developed GeoTDM, the first diffusion model specifically designed to generate 3D geometric trajectories while preserving physical symmetries. The model captures complex temporal distributions and achieves superior performance in generating realistic dynamics across N-body simulations, molecular dynamics, and pedestrian motion, for both unconditional and conditional tasks.
Key insights
- Preserving SE(3) physical symmetries throughout the diffusion process is critical for generating realistic and generalizable 3D geometric trajectories.
- Combining equivariant spatial convolutions with temporal attention in the Equivariant Geometric Trajectory Network (EGTN) effectively captures both local geometric properties and long-range temporal dependencies in dynamic systems.
- Introducing a learnable, generalized equivariant prior for conditional generation enables more flexible and accurate modeling of diverse dynamic behaviors than fixed priors.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.