World models let embodied agents anticipate the consequences of their actions, but building them for robots that both move and manipulate objects is a hard problem. The difficulty grows when trying to cover many different robot designs, each with its own action space and sensory setup.
The paper, Ropedia W0, identifies two central challenges: heterogeneous action spaces across embodiments and limited multimodal data. To address them, the authors propose scaling loco-manipulation world models by incorporating human experience, rather than relying solely on robot-collected data.
The abstract outlines the problem and the proposed direction rather than presenting full results, so the work appears to be an early-stage research agenda. Still, it points toward a more general approach to robot world models that could transfer skills across very different bodies.