British startup Worldmodeldata is turning video game play into training material for “world models,” AI systems intended to understand how actions change an environment. The company pairs visual scenes with controller inputs, preserving the cause-and-effect information that ordinary internet video usually lacks.
That distinction matters for systems expected to steer a robot arm, pilot a vehicle, or navigate a 3D space. Text and video can show what exists, but physical control also depends on details such as direction, timing, force, and torque. Worldmodeldata plans to broker and organize gameplay records from studios, giving model developers an alternative to negotiating separately for many proprietary datasets.
Games are not the physical world, and even realistic simulations omit friction, sensor noise, safety constraints, and other conditions that robots face. The usefulness of a dataset will also depend on rights, game diversity, and whether player actions align with the target task. Still, the approach could provide a large source of synchronized observation-and-action examples at a time when world-model researchers have far less structured training material than language-model developers found in web text.