01Definition
NVIDIA popularized the term industry-wide — CEO Jensen Huang's CES 2025 keynote framed physical AI as AI that understands physics, friction, inertia, and cause-and-effect. At its core is "physical reasoning": predicting how a ball will roll, estimating how much force to use when grasping an object without damaging it, or inferring that a pedestrian may be standing behind a car. (NVIDIA's official blog)
02How it differs from generative AI
Generative AI
Output is a document, image, or line of code. Judged mainly by benchmarks and human preference on a screen.
Physical AI
Output is physical motion. Judged by whether it works reliably in the messy, unstructured real world, not a leaderboard.
03The perceive → reason → act loop
Physical AI combines AI models — machine learning, computer vision, natural language processing — with robotics hardware: sensors and actuators. This three-step loop is what lets a system sense its environment, decide what to do, and take physical action.
04Why it's hard
Three things make this loop much harder than a chatbot's: it has to run in real time (a robot arm can't pause for a second to "think" before catching a falling object), it has to handle a long tail of situations no training set fully covers (every kitchen, warehouse, and sidewalk is arranged differently), and mistakes have physical, sometimes safety-critical consequences instead of just an awkward reply. That combination is why physical AI is judged on reliability in messy real-world conditions rather than a benchmark score.
05Where it shows up
Humanoid robots, industrial arms, autonomous vehicles, and drones are the main hardware categories physical AI gets built into — see the Company & Robot DB for who's building them, or the history of the robot industry for how the field got here.
A growing slice of the industry, though, doesn't build robot hardware at all — it builds the "reasoning" layer only, as a foundation model meant to run across many different robot bodies. Two companies tracked in the Company & Robot DB illustrate that split from the humanoid makers above.
π0 / π0.5
General-purpose robot policies meant to run across many different robot bodies — no hardware of its own.
Skild Brain
A general-purpose "brain" model meant to work across humanoids, quadrupeds, and industrial arms.
Looking at this from a public-markets angle — which chipmakers, robotics manufacturers, and industrial suppliers have exposure to physical AI — is covered separately by the listed-stock tracker, rather than in this glossary entry.
06Related terms
07FAQ
Q.Is physical AI just another word for robotics?
A.Not quite — robotics is the broader engineering discipline covering mechanical design, control theory, and manufacturing that's existed since the 1960s. Physical AI specifically refers to the AI/foundation-model layer (perception, reasoning, learned control policies) now being built on top of that hardware.
Q.Why did the term suddenly become common around 2025?
A.NVIDIA CEO Jensen Huang popularized the term at CES 2025, framing it as AI's next frontier after text and image generation — AI that "understands the laws of physics." The framing stuck industry-wide because it gave a name to work (robot foundation models, sim-to-real transfer) that was already underway.
Q.How is this different from embodied AI?
A.They overlap heavily. Physical AI tends to emphasize deployed, real-world hardware; embodied AI is the broader academic term that also covers agents trained entirely in simulation.
Q.Do all physical AI companies build their own robots?
A.No. Some, like Figure AI and Apptronik, build both the hardware and the AI. Others, like Physical Intelligence and Skild AI, build only the foundation-model "brain" and aim to run it on other companies' robot bodies.