Current generative models produce visually convincing sequences that often remain geometrically inconsistent and physically implausible, limiting their use as predictive models for embodied agents. This project aims to build 4D world models grounded in the geometry, motion, and physical behavior of the real world rather than appearance alone. The research first establishes a strong foundation by investigating dense correspondence and inductive biases for scene reconstruction, specifically evaluating the integration of explicit 3D priors within latent or pixel spaces. These structural representations are then utilized for manipulation policies to improve geometry grounding and generalization over purely appearance-based features. Subsequently, these representations will provide structural guidance for generative world models to enable geometrically consistent, long-horizon predictions. Finally, to ensure true physical plausibility, the project will integrate real-world physical constraints such as contact, friction, and tactile feedback. By grounding the 4D representation in accurate physical interactions, the resulting framework will serve as a reliable simulation environment to support a robust real2sim2real pipeline for robotic deployment.