Video generalization models can learn physics-based controls from synthetic data and generalize this control across diverse geometries, settings, and materials.