Learning a Particle Dynamics Model with Real-world Videos
Researchers have developed a new method to train neural object dynamics models directly from unlabeled real-world videos, overcoming limitations of synthetic data. The framework uses a particle-based dynamics model integrated with Gaussian splatting to predict changes in particle position and rotation over time. This approach enables learning from real-world videos without needing explicit particle-level state labels, and includes a new dataset of approximately 500 videos showcasing diverse object interactions. AI
IMPACT Enables more realistic physics simulations by training directly on real-world data, potentially reducing the sim-to-real gap in AI.