Physical AI: What Does it Mean for Embedded Engineers?

• Share on X
• Share on LinkedIn

August 12, 2026

Blog

I don’t like it because it’s redundant, marketing-speak for what embedded engineers have been doing since the beginning of Machine to Machine (M2M), the precursor to IoT (now Sensor Fusion, maybe?), and Machine Learning (ML, what’s now called AI). The combination of M2M and ML has moved through many names, and while the tech has gotten faster, smoother, more powerful, and (thankfully) more secure, calling it Physical AI now doesn’t fundamentally change what’s going on under the hood.

Rant aside, is there anything different that you engineers and developers need to know as you are handed mandates from above to ensure the next product launch can have that magic term on the box?

Actually, yes.

As Rich Nass defines it, Physical AI is the application of AI and IoT tools to systems that directly interact with the real world. In this column, Nass points out that motor control is more important than ever in a world that can be and will be manipulated by machines operating independent of direct human control. I can’t agree more with Rich here (not always the case, as he well knows). Motor control has become critical infrastructure, if it wasn’t already, thanks to a wide variety of factors unique to Physical AI architecture.

Tolerances need to be even more precise because a machine won’t be able to make a judgement call on the fly about smoothness, efficiency, or inaccuracies. Even if the system can tell something is wrong, that won’t stop the intervening process from either going into downtime or pushing through faulty results until the problem is fixed. Those are both unacceptable outcomes.

That leads me to safety. As we all know, a sensory system is only as good as what it can recognize, and a machine brain can only adapt if it’s told what new information means. Until the new information is defined, that machine can’t adjust to new circumstances. So, the unexpected is a problem. No matter how big or “complete” any data set is, there is no way to anticipate the unknown unknowns, and that is where safety problems live. Physical AI systems need to have safeguards upon safeguards, redundancies, and humans in the loop, all of which can introduce inefficiencies and slow down the process.

I won’t be surprised (or upset) if I start hearing that Physical AI systems are slower than human-operated systems, at least for a while.

I’m not trying to be a downer here, either. Ever since Jensen Huang of NVIDIA popularized (and likely coined) the term, it’s been present and almost synonymous with industrial automation and AI in nearly every conversation I’ve had over the past year. It’s fine if we all agree to use the new term, but to me, that means we need to be very careful about the new implications of what we’re calling Physical AI.