Physical AI creates a different category of consequence
A system that generates text can be constrained by access rights and review. A system that moves a machine also has to be understood through physical safety, predictable control, intended use and conformity assessment. The EU Machinery Regulation is relevant because it brings the discussion back to a basic fact: capability becomes commercially meaningful only when it can enter a regulated physical environment.
Safety is a product architecture issue
Regulation should not be treated as a document produced at the end of development. In embodied systems, safety depends on the architecture of sensing, control, human interaction, emergency behaviour, software updates and limitation of use. As models become more adaptive, the relationship between learning behaviour and a documented safety case becomes a core design problem.
The signal to watch
Manfred is watching how builders translate general-purpose embodied models into defined operating domains. The strongest evidence will be systems with transparent limits, controlled updates, clear human responsibility and practical proof that safety engineering can keep pace with model capability. The question is not whether robots become more intelligent. It is whether organisations can make that intelligence deployable and accountable.