Reproduce “Learning a Particle Dynamics Model with Real-world Videos” 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
Chanho Kim, Suhas V. Sumukh, Li Fuxin
Researchers at Oregon State University developed a framework that learns multi-object collision dynamics directly from unlabeled real-world videos. The approach combines a particle-based dynamics model with Gaussian Splatting and differentiable rendering to predict dense particle attribute changes, showing competitive performance on real-world bowling and falling cube stack scenarios.
Key insights
- Differentiable rendering, especially with Gaussian Splatting, enables direct 3D dynamics model supervision from 2D image observations, bypassing the need for explicit 3D ground-truth labels or physics simulators.
- Explicitly incorporating object identity and object-centric convolution layers significantly enhances the model's ability to capture complex multi-object collision dynamics compared to generic point cloud processing.
- A robust data pipeline, involving depth map generation, video object segmentation, 3D tracking, and ICP-based pose estimation, is critical for consistently extracting 3D Gaussian trajectories from noisy real-world video data.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.