Before a robot pushes an object, it helps to predict what will happen. A world model learns a representation of an environment and its possible changes. Given a current state, and sometimes a proposed action, it predicts a later state. That representation may involve images, video or internal features that people cannot directly view.
The central question is how events unfold over time. An agent can compare predicted outcomes before choosing an action. Google DeepMind’s Genie research, for example, explores generated visual environments that users can navigate and influence.
Consistency matters more than an impressive picture. An object should retain its properties behind an obstacle, and a predicted outcome should follow from the preceding situation. Convincing imagery can still contain incorrect physics. Training in a generated environment therefore needs validation in the real world. World models connect computer vision, AI agents and robotics through the practical problem of anticipating consequences.