Dense contact maps across the hand surface capture forces that cameras can't see. This is what closes the sim-to-real gap for dexterous manipulation.

Head-mounted and wrist-mounted cameras give the model the same viewpoint as the operator, not a detached third-person angle.

Every frame is labeled with 3D hand pose, contact state, object interactions, and spatial context. All streams time-aligned to a shared clock.

Detailed task descriptions, environmental context, and counterfactual annotations. What happened and what could have.

Physical priors distilled from our capture data. Geometry, contact, and persistence constraints that teach world models what's allowed to be true.