The term Physical AI appears in every vendor keynote and research call-for-papers in 2025 — NVIDIA’s Jensen Huang uses it, robotics labs use it, and now Korean ministry grant applications are starting to use it too. But when I sit down to build a curriculum around it, the term gets slippery fast.
Here is the definition I’ve settled on for teaching purposes:
Physical AI is machine intelligence that perceives, reasons about, and acts within the physical world in real time, typically on or near the hardware doing the sensing and actuation.
Three words do the heavy lifting: perceives, acts, and real time. A model that classifies images in the cloud is not Physical AI. A model that runs on a microcontroller attached to a robot arm, reads a sensor, and moves a servo in response — that is.
Why “Physical” and not “Embodied”?
Embodied AI is the older academic term and carries strong connotations from cognitive science: it implies a body, proprioception, environmental coupling. Physical AI is the industrial/engineering framing — it emphasises the hardware constraints (power budget, latency, memory), the deployment context (edge, embedded, real-time OS), and the feedback loop with the physical world.
For curriculum design, “Physical AI” maps more directly to what students will build: sensors, actuators, constrained inference, and control loops.
Three axes for curriculum design
When I plan a track or module, I ask where the content sits on three axes:
- Perception → Action: Is this primarily about sensing the world, reasoning about it, or changing it?
- Cloud → Edge → Bare-metal: Where does inference happen?
- Prototype → Production: Are we exploring possibilities or engineering reliability?
Most university courses live in the upper-left corner (perception, cloud, prototype). Physical AI curriculum deliberately pulls toward the lower-right: actuation, bare-metal, production-grade thinking.