Action makes perception consequential.
A model can classify an image without affecting the world. An embodied system must decide what to do next: move, speak, adjust, wait, ask, warn or remain inactive.
The right action depends on more than the physical state.
A robot may know where a person is without knowing whether it should approach. An ambient system may detect activity without knowing whether interruption is welcome. Physical competence and social appropriateness are different problems.
Sensitivity can be expressed as restraint.
In human environments, the best response is not always the most proactive one. Confidence thresholds, reversible actions, permission and the ability to defer can be part of intelligent behavior.
Embodiment is broader than robotics.
The same loop can exist in a wearable, a room, a vehicle, a phone or a software agent connected to live context. What matters is that intelligence is continuously coupled to a changing situation.