Reproduce “UniPhysGen: Unified Physical Grounding for Simulation-Ready 3D Assets” 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
Xian Li, Rong Wei, Lujie Yang, Haolin Huang, Junyuan Fang, Siliang Tang, Jun Xiao, Rui Tang and 1 more
The UNIPHYSGEN framework enables the automatic transformation of diverse 3D meshes into simulation-ready assets by unifying articulation semantics and intrinsic physical properties. This approach allows for robust zero-shot deployment in physics simulations for embodied interactions, demonstrating superior performance on the UniPhys-Bench dataset with an 80.46% mIoU for articulation structure.
Key insights
- Jointly inferring articulation and intrinsic physical properties ensures greater consistency and physical plausibility compared to fragmented approaches.
- Combining semantic reasoning from Large Vision-Language Models with geometry-aware hypothesis generation and robust augmentation strategies (e.g., SO(3) rotation, spherical axis parameterization) is critical for accurate 3D articulation grounding.
- Simulation-driven verification effectively identifies and filters physically implausible annotations, ensuring the generated 3D assets are truly 'simulation-ready'.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.