A compact learned representation encoding scene-specific physical properties like mass, friction, or damping for model adaptation.