Taxonomy Of World Models

World Labs proposes in an essay that the many systems called world models are best told apart by what they output. A renderer outputs observations, such as video frames meant for human eyes. A simulator outputs a structural description of the world's state (geometry, physics and dynamics) that people and programs can compute on. A planner takes observations and a goal and outputs actions. The essay says these functions can overlap but each is judged differently: visual fidelity for a renderer, structural faithfulness for a simulator.

The distinction helps explain why a convincing generated scene is not enough for training a robot: the essay argues renderers optimize for visual plausibility, while training needs a world whose geometry and responses to actions a program can rely on. It argues simulation is the bridge between visual generation and action planning, so physical and geometric consistency is central. Note that this is a company's framework, and World Labs' own product, Marble, is a 3D world generator in the simulator category it emphasizes. Functional taxonomy.