Predicting frames as a form of world modeling
One of the most active research directions connected to world models is generative video: systems trained to predict or generate plausible future frames of a scene. At first glance this looks like a different problem from robot planning, but it rests on the same underlying idea, learning what tends to happen next given what's happening now. A model that can generate a convincing next frame of a rolling ball or a person opening a door has, in some sense, absorbed something about how that scene behaves over time.
This is part of why generative video research and world-model research increasingly overlap. A video-prediction model isn't automatically a full world model in the planning sense, since predicting a plausible-looking frame is not the same as predicting the physically correct outcome of a specific action. But the two lines of work share techniques, share the goal of capturing environment dynamics, and are often discussed together as a broader push toward machines that model how the world unfolds, not just what it looks like in a single moment.