What is physical AI? Physical AI vs embodied AI vs robotics, explained

Physical AI is artificial intelligence that perceives, reasons and acts in the real world through a machine: a robot, a self-driving car, a drone or a smart camera. NVIDIA, which popularised the term, defines it as AI that lets “autonomous systems like cameras, robots, and self-driving cars perceive, understand, reason, and perform or orchestrate complex actions in the physical world.”

Why it matters: the phrase now names funding rounds, national budgets and product lines. Knowing what it covers, and how it differs from embodied AI and plain robotics, is the first step to reading the numbers attached to it.

What does physical AI mean?

Physical AI means an AI model whose output is an action in the physical world rather than text, images or code. A chatbot predicts the next word. A physical AI system has to predict what happens when a gripper closes or a wheel turns, and deal with friction, gravity, occlusion and people walking past. NVIDIA’s glossary lists warehouse robots, robot arms that adjust their grip to an object’s position, surgical robots and autonomous vehicles as examples.

The term spread quickly as a marketing and policy label from 2025. When NVIDIA launched its Cosmos world foundation models at CES on January 6, 2025, it framed them as a platform to “accelerate physical AI development,” and chief executive Jensen Huang said “the ChatGPT moment for robotics is coming.” The International Federation of Robotics, the industry’s statistics body, adopted the term in its February 2026 position paper on AI in robotics, describing physical AI as technology that lets robots “train themselves in virtual environments and operate by experience, rather than programming.”

What is the difference between physical AI and embodied AI?

Mostly emphasis. Embodied AI stresses that intelligence comes from having a body and learning through it; physical AI stresses that the system acts on the physical world. NVIDIA defines embodied AI as “the integration of artificial intelligence into physical systems, enabling them to interact with the physical world,” which is close to its physical AI definition. The academic usage is older and broader. A widely cited survey by Jiafei Duan and colleagues in IEEE Transactions on Emerging Topics in Computational Intelligence contrasts “internet AI,” which learns from curated images, video and text, with embodied AI, which learns “through interactions with their environments from an egocentric perception similar to humans.” In that research tradition, an agent navigating a simulated house counts as embodied AI even though no hardware exists.

Usage also splits by geography. English-language industry and Korean policy favour “physical AI.” China’s policy language favours embodied intelligence: in June 2026 its industry ministry and state assets regulator launched a national programme to test “humanoid robots and embodied artificial intelligence systems” in real settings, according to the Global Times.

Key terms compared, as of September 2026.

Term	Definition	Example	Who uses the term
Physical AI	AI that lets autonomous systems perceive, understand, reason and act in the physical world (NVIDIA)	Warehouse robots, self-driving cars, humanoids	NVIDIA, the IFR, Korea’s government and investors
Embodied AI	AI integrated into physical systems (NVIDIA); agents that learn through interaction from an egocentric view (Duan et al.)	Navigation agents in simulators, humanoid robots	Academic researchers; China’s policy and industry (“embodied intelligence”)
Robotics / robot	“A programmed actuated mechanism with a degree of autonomy to perform locomotion, manipulation or positioning” (ISO 8373:2021)	A welding arm on a car line, a delivery robot	ISO, the IFR, manufacturers and integrators
Vision-language-action (VLA) model	A vision-language model adapted to output robot actions	Robot foundation models such as NVIDIA’s GR00T family	AI labs and robot makers
World model	A model that predicts how the physical world will change, often in response to actions	NVIDIA Cosmos	AI labs, simulation vendors, investors

Robotics is the oldest and most concrete of these words. By the ISO definition, a robot needs only programming, actuators and some autonomy. An industrial arm that repeats a fixed welding path is a robot but not physical AI. It becomes physical AI when a learned model, rather than a hand-written program, decides what it does.

What are the layers of the physical AI stack?

Physical AI splits into five layers, and value is being fought over at each one.

• Hardware. Bodies, actuators, hands, sensors and on-board compute. This is where volumes are counted, as our tracker of humanoid shipments by company shows, and where prices set adoption (humanoid robot prices in 2026).
• Foundation models. Vision-language-action models translate camera input and an instruction into motor commands. Our VLA explainer covers the architecture and our model comparison the contenders. World models that predict the physical consequences of actions are the next research bet.
• Data. Robots lack an internet-scale corpus of actions. Teleoperation, human video and purpose-built collection sites fill the gap, most visibly in China’s robot data factories.
• Simulation. Training in simulation is cheap and safe, but policies have to survive the transfer to hardware. Sim-to-real explains why that gap is narrowing.
• Deployment and software platforms. The operating layer that runs models across different robots, manages fleets and connects to business systems. Our piece on the race for the layer above the robot maps it.

Money has moved faster than revenue. Our analysis of where physical AI revenue is booked today found it in surgical robots, warehouse systems and robotaxis rather than humanoids. IDC counted about 18,000 humanoid robots shipped worldwide in 2025, according to the Economic Observer, a small base next to the funding the sector attracts.

Why does Korea use the term physical AI?

Because it is a budget category. Korea’s government proposal for 2027, released on September 1, 2026, sets ₩3.1 trillion for physical AI as one of three “mega-projects” alongside AI semiconductors and AI data centres, within a ₩21.3 trillion package for those projects and AI, Seoul Shinmun reported. The money covers about 2,000 domestically made AI robots for eight sectors, including manufacturing, construction, agriculture, defence, elder care, logistics and ports, plus ₩1.5 trillion of core technology R&D and money for training centres and data factories. Within the Ministry of Science and ICT’s ICT R&D budget, ₩384.9 billion is earmarked for physical AI work on world models, general-purpose foundation models and AI accelerators aimed at cars, construction and manufacturing. The figures are a proposal and still need National Assembly approval. Our breakdown of Korea’s 2027 physical AI budget has the line items.

The term fits Korea’s economy. “Physical AI” puts factories, shipyards and cars inside the AI story, where Korea’s manufacturers already have scale, rather than centring humanoids alone. For which Korean companies actually ship product in each layer, see our ranking of Korea’s physical AI firms.

Our view: treat “physical AI” as an umbrella rather than a technology. When a company or government uses it, ask which layer it means. A humanoid maker, a chip vendor and a simulation company can all claim the label and be exposed to very different risks.