What is a world model in robotics?

The idea

Give a model enough video of things happening and it can learn to predict what happens next. Do that well enough, conditioned on an action, and you have something a robot can run forward in its head: if I push here, the glass tips.

The contrast with a physics engine is the important one. In a physics engine somebody wrote the physics — friction coefficients, contact solvers, integration steps. In a world model nobody did; it was inferred from watching. That is the appeal and the doubt in one sentence.

Two uses follow. A world model can be a planner's imagination, letting a policy evaluate options before acting. Or it can be a data factory, generating synthetic experience at a scale no teleoperation programme could reach.

Why the term is a mess

Because it is used for all of that and more. We have mapped the competing definitions in detail, and the short version is that "world model" currently spans video generators, action-conditioned simulators, latent predictive models and marketing.

Some builders have lost patience with it. Generalist AI's Pete Florence argues that VLA and world model are both temporary crutches, labels the field will drop once physical foundation models are simply what everyone builds. When a term means everything, it stops carrying information — and that is roughly where this one is.

Who is betting on it

1X has gone furthest publicly. It first showed an action-controllable world model as a high-fidelity training simulator, then stood up a dedicated World Model Lab and pivoted hard toward large-scale pretraining. Its evaluations director has described the approach as imagination-as-policy, using video generation to get zero-shot behaviour.

Outside humanoids, Waymo built a world model on DeepMind's Genie 3 to simulate the rare long-tail driving scenarios you cannot collect enough of by driving. AGIBOT's Genie Envisioner 2.0 moves from modelling the world to generating environments agents can evolve in. NVIDIA has put it at the centre of its robotics story, with Cosmos 3 pitched as moving the industry away from language-heavy VLAs toward world-action models.

The AGI argument behind it

The enthusiasm is not only practical. Yann LeCun has argued for years that language models are the wrong substrate for physical intelligence, and told an MIT audience that humanoid companies lack the world-model breakthroughs their roadmaps assume. Demis Hassabis makes a constructive version of the same case, describing world models and infinite training loops as keys to AGI and language models as insufficient for robotics.

So the term carries an argument, not just a technique: that copying demonstrations tops out, and understanding consequences does not.

Reading a world-model claim

Three questions cut through most announcements.

Is it action-conditioned? A model that generates plausible robot video is not the same as one that answers what happens if I do this. Only the second is useful for planning.

Is it being used, or shown? Generating footage is a demo. Training a policy inside it, or planning with it on hardware, is a result.

What replaced what? If a world model supplies synthetic data to a VLA, it is a component of the current stack, not a successor to it. Both framings are in circulation, and vendors are rarely explicit about which they mean.