Physical AI defined in plain terms, set against robotics, embodied AI and autonomy, with IFR and Eurostat data on how far deployment has got.
Physical AI is software that turns sensor data into movement. A model takes in camera images, depth readings and force or position signals, builds a working picture of the space around a machine, decides what to do next and sends commands to its motors. The result is an action: a gripper closing on a part, a mobile robot changing route, a car braking.
No standard definition exists yet, though the main ones are close. NVIDIA's glossary names "cameras, robots, and self-driving cars" among the systems that physical AI lets "perceive, understand, reason, and perform or orchestrate complex actions in the physical world". The same page calls it an extension of generative AI "with an understanding of spatial relationships and the physical behavior of the 3D world".
For the machine itself, a public definition comes from the US National Science Foundation. Its robotics programme defines a robot as "intelligence embodied in an engineered construct, with the ability to process information, sense, plan, and move within or substantially alter its working environment".
The phrase is older than the current interest in it. In November 2020, Aslan Miriyev and Mirko Kovac, of the Swiss materials research institute Empa and Imperial College London, wrote in Nature Machine Intelligence that "synthesizing robots via physical artificial intelligence is a multidisciplinary challenge for future robotics research". Their paper describes the field as "the theory and practice of synthesizing nature-like intelligent robotic systems".
NVIDIA has since made the term part of its product launches. At CES on 6 January 2025 it launched Cosmos, a platform of generative "world foundation models" and tools for developing robots and autonomous vehicles, partly by generating synthetic training data. Jensen Huang, its founder and chief executive, said: "The ChatGPT moment for robotics is coming." The press release itself lists that sentence among its forward-looking statements, "subject to risks and uncertainties".
Researchers also use the term more widely. A survey posted to arXiv in October 2025 by Kun Xiang and co-authors at Sun Yat-sen University, ETH Zurich and other institutions treats physical AI as AI grounded in physical laws, tracing four stages from perception and reasoning to modelling and embodied interaction, and covering symbolic physics reasoning and generative models alongside embodied systems. On that reading, a model that reasons about physics from text or video counts as physical AI, with no robot involved. This article uses the narrower sense, AI that acts through hardware, because that is what the installation data and the public programmes measure.
The IFR's World Robotics 2026 report, published in September 2026, gives worldwide figures for 2025.
Industrial robots. The operational stock rose 9% to a record 5 million units, and installations grew 11% to more than 600,000. "Industrial automation is progressing at high speed," said IFR President Jane Heffner, adding that the five million mark "is more than double the number seven years ago". China installed 354,000 robots, 59% of the world total, and Chinese suppliers held 55% of their home market. The United States moved into second place with almost 38,500 installations, up 12%, ahead of Japan (36,219, down 19%), the Republic of Korea (30,000) and Germany (fewer than 25,000, down 8%).
Service robots. Sales of professional service robots rose 24% to almost 250,000 units. Transport and logistics was the largest application, with 117,500 units and a 47% share. Medical robots reached 15,300 units, including 8,400 surgical robots, up 57%. The IFR adds that its service robot figures come from a sample of 238 of 1,001 manufacturers, with no projection to the whole industry. Inspection is one of the professional uses, and our piece on quadruped robots follows legged machines from the lab to that work.
Humanoids. About 7,000 full-size humanoid robots, taller than 140 cm, were sold worldwide in 2025. In the IFR's words, most current applications "remain specialized and often require human teleoperation", and it names "safety standards, high training and maintenance costs, and the lack of a strong business case" among the obstacles. Its September presentation placed humanoids at the step from laboratory to pilot projects, ahead of real deployments. For every full-size humanoid sold in 2025, factories installed more than 80 industrial robots.
Use inside companies. Eurostat's survey of EU enterprises with ten or more employees and self-employed persons asks which AI technologies they use. In 2025, 19.95% used at least one. The least common type was AI "enabling machines to physically move by observing their surroundings and taking autonomous decisions", used by 1.39% of enterprises. Among large enterprises the share was 8.55%, again the lowest of the AI types surveyed.
The IFR counts machines, whatever software runs inside them. Five million industrial robots is a count of hardware, and the report does not say how many of them use learned models. It does say that "advances in artificial intelligence, machine vision and sensing are increasing robot capabilities".
Vehicles sit outside the robot statistics. Robotaxis and driverless trucks fall within NVIDIA's definition and Eurostat's category, and none of the official sources used here gives a global count for them comparable to the IFR series.
Eurostat's 1.39% comes closest to measuring AI-driven machines in business use, and it covers the EU only. It also counts firms. An enterprise with one autonomous forklift and another running a fleet of them each count once, so the figure shows how many companies have started and says little about how far any of them has gone.
The European Commission presented its Apply AI Strategy on 8 October 2025. Its sectoral flagships include "targeted measures to boost AI adoption across 10 key industry sectors", robotics among them. In its opinion on the strategy, published in the Official Journal in April 2026, the European Economic and Social Committee wrote that "the focus on 'physical AI' and collaborative robotics is credible and commendable". It added that the strategy "should more explicitly address labour market impacts, including upskilling, job transitions and social dialogue".
In the United States, the National Science Foundation's Foundational Research in Robotics programme, run jointly by its computing and engineering directorates, supports research on "robotic systems that exhibit significant levels of both computational capability and physical complexity".
It means AI that perceives the physical world through sensors and acts in it through a machine. Robots, autonomous vehicles and camera systems that trigger actions are the usual examples. The term refers to the learned software; the machine itself belongs to robotics.
Generative AI models are trained mostly on text and images gathered from the internet, and they produce content. Physical AI has to produce motion that obeys physical rules, so NVIDIA's glossary says it needs physically accurate data, much of it generated in simulation. The two are converging: NVIDIA describes physical AI as generative AI extended with an understanding of space and physics, and its Cosmos models are generative models used to create training data for robots and vehicles.
NVIDIA uses the term for AI that lets robots, vehicles and camera systems act in the physical world. Its glossary describes a chain of three computers for building it: one to train models, one to simulate the world and generate synthetic data, and one on board the robot or vehicle. Cosmos, launched in January 2025, sits in the simulation step.
The IFR counts 174 humanoid robot manufacturers alone, more than half of them in China. When NVIDIA launched Cosmos, it named 1X, Agile Robots, Agility, Figure AI, Foretellix, Uber, Waabi and XPENG among the first companies to adopt it. On the buying side, transport and logistics was the largest application for professional service robots in 2025, and 8,400 surgical robots were sold for clinical use.
Largely. Both describe AI that perceives and acts through a body. Embodied AI is the established research term, with surveys of its own; physical AI is the label used by NVIDIA and in EU policy texts. Some researchers use physical AI more broadly, for any AI that reasons about physical laws, with or without a robot.
Physical AI and robotics is one of the four keynote-led panels at the Davos Super Intelligence Summit 2027, on 19 January 2027 at the Grandhotel Belvedere during Davos week. Prof. Sangbae Kim, Robotics Architect, Meta & MIT, is among the speakers announced, and the full programme sets out the four panels. Access is by invitation only.