1.1 · Physical AI as Causal Intervention
Physical AI begins when learned outputs can influence the evolution of a physical system. This lesson reframes the engineering objective from producing plausible content to making time-bounded, embodied decisions under uncertainty and constraints.
- Define Physical AI operationally.
- Trace the closed causal loop from sensing to action and back to sensing.
- Identify why embodiment, timing, uncertainty, and constraints change the ML contract.
Prediction is not intervention
A model prediction becomes Physical AI only when it affects planning, authorization, or control. Once an action changes the world, the next observation depends on the previous model output. Errors can therefore compound through the closed loop.
The complete causal loop
An observation is encoded into a representation, used to infer an action proposal, passed through authorization and control, executed by the plant, and sensed again. Network delay, friction, illumination, contact, and actuator response are part of the learning environment.
Operational definition
A Physical AI system couples learned perception, prediction, reasoning, or policy outputs to sensors and actuators. Its performance must be judged under embodiment, real-time deadlines, uncertainty, and physical limits.
The physical world closes the loop. Every model output must be evaluated by the state it can create next.
