Predicting physical dynamics from visual data remains a fundamental challenge in Artificial Intelligence (AI), as it requires both accurate scene understanding and robust physics reasoning. While recent video generation models achieve impressive visual quality, they lack explicit physics modeling and frequently violate fundamental laws like gravity and object permanence. Existing approaches combining 3D Gaussian splatting with traditional physics engines achieve physical consistency but suffer from prohibitive computational costs and struggle with complex real-world multi-object interactions. The key challenge lies in developing a unified framework that learns physics-grounded representations directly from visual observations while maintaining computational efficiency and generalization capability. Here we introduce Neural Gaussian Force Field (NGFF), an end-to-end neural framework that learns explicit force fields from 3D Gaussian representations to generate interactive, physically realistic 4D videos from multi-view RGB inputs, achieving two orders of magnitude speedup over prior Gaussian simulators. Through explicit force field modeling, NGFF demonstrates superior spatial, temporal, and compositional generalization compared to state-of-the-art (SOTA) methods, including Veo3 and NVIDIA Cosmos, while enabling robust sim-to-real transfer. Comprehensive evaluation on our GSCollision dataset—640k rendered physical videos (~4TB) spanning diverse materials and complex multi-object interactions—validates NGFF’s effectiveness across challenging scenarios. Our results demonstrate that NGFF provides an effective bridge between visual perception and physical understanding, advancing video prediction toward physics-grounded world models with interactive capabilities.
