What makes a world model different
A language model predicts the next word (see "How does a language model actually 'think'?"). A world model attempts something fundamentally different: predicting how the physical world changes in response to an action - for example, what happens when an arm grasps and moves an object. Research mostly trains such models on huge amounts of video footage, so they implicitly learn physical regularities like gravity, friction, and object permanence.
Why this matters for robotics
Classic robotics programming needs explicit rules for every new task. A good world model could instead predict on its own what a planned action would cause, before executing it - a kind of internal simulation. Early research prototypes already show this on individual, well-scoped tasks, such as grasping familiar object types.
The sim-to-real gap
The central unsolved problem: models trained in simulation or on video data often fail in the real, unpredictable physical world - friction, lighting, and material properties differ from the simulation. This "sim-to-real gap" has been one of the most stubborn robotics problems for years.
A realistic take
For narrowly defined, controlled environments - warehouses with standardized objects, for example - deployable systems already exist today. For open, unpredictable physical environments like private homes or construction sites, reliable maturity isn't expected within a few years - it depends on whether the sim-to-real gap gets fundamentally closed, not just on bigger models.
Why this matters for you as a decision-maker
Plain software automation (see "AI vs. automation: what fits where?") remains the relevant lever for most businesses. Physical automation pays off today mainly where the environment is already standardized and controlled - not as a general expectation of "robots that can do everything".