Friction is key to making better robot world models

World models are AI systems that help robots predict the consequences of their actions and generalize to new situations, but current models rely primarily on camera images and joint position data, which is insufficient for tasks involving physical contact with objects. The missing piece is the coefficient of static friction, the physical parameter that determines whether a grasped object will slip, which cannot be seen by cameras or inferred from joint currents alone. When robots condition their world models on friction as a direct input rather than trying to infer it from pressure maps or motor signals, they shift from statistical pattern-matching toward causal physical understanding, enabling reliable generalization across novel surfaces and objects. This approach, called VμA models (vision-μ-action), treats friction as a first-class conditioning input alongside vision and motor action.