Embodied World Models

Embodied world models study how an agent and its surroundings change over time, especially when the agent acts, so it can anticipate outcomes before acting in the real world. Approaches range from video-generative and world-action models that jointly predict future observations and actions to latent-dynamics models and explicit 3D, geometric, or symbolic representations. These models are used for robot manipulation and navigation, policy training and evaluation, and generating synthetic experience; common methods include diffusion models, recurrent state-space models, and model-predictive control. A central tension is that visually plausible predictions may still violate physics or mislead decisions, while accurate, useful long-horizon predictions remain costly and difficult to evaluate in closed-loop and real-world settings.