Un sistema intelligente deve funzionare anche fuori dallo schermo.

A model on a screen can classify, predict or recommend. A physical system has to do something harder: observe an imperfect environment, decide under uncertainty and act without ignoring the consequences of that action.

That closed loop is the part I find most interesting. The quality of the system depends not only on the model, but also on sensor placement, timing, actuator limits, power, communication and the way failure is handled.

A strong prediction is not useful if the sensor is poorly calibrated, the control signal arrives too late or the mechanism cannot reproduce the requested motion. Physical intelligence therefore starts with the interfaces between electronics, software and the environment.

This changes the engineering question from “How capable is the model?” to “What complete sensing-to-action loop can be trusted for this task?”

When a system affects hardware, experiments or people, its uncertainty should be visible. Thresholds, assumptions and fallback behaviour need to be understandable enough that someone can decide whether the next action is justified.

For me, the goal is not intelligence that merely looks advanced. It is a system that makes the next physical step clearer, safer and easier to evaluate.

I evaluate an intelligent physical system by four questions: What does it sense? Which uncertainty changes the decision? What can it safely do? What evidence would cause it to stop or choose a different action?

Those questions connect my interests in electronics, automation, motor control and autonomous machines. They also keep the work close to reality, where every useful decision eventually meets a physical constraint.