Last reviewed:
What is physical AI? Definition, examples and what it means for an industrial SME
Physical AI refers to artificial intelligence systems that perceive the real world through sensors, reason about it, then act on it through machines: robots, autonomous vehicles, industrial equipment. The term is mainly promoted by NVIDIA and has no normative definition; it differs from generative AI in that a mistake has material consequences.
The term has no legal or standards-based definition. NVIDIA, which brought it into common use, describes it as AI that lets autonomous machines perceive, understand and perform complex actions in the physical world. IBM words it similarly: AI models combined with sensors, actuators and control systems to perceive an environment, reason, act and learn from the results. The applications cited range from warehouse robots and robot arms to autonomous vehicles, surgical robots, humanoid robots and factories run with cameras and digital twins. The difficulty lies in data and simulation. Training data requires machines to interact with a real environment, which is slow and costly. A model that performs well in simulation can fail in the field (friction, wear, noise). Hence the interest in world models: on 6 January 2025 NVIDIA launched Cosmos, a family of world foundation models that generates realistic synthetic data to train and evaluate physical AI models, under an open model license. Another player: on 12 March 2025 Google DeepMind introduced Gemini Robotics, a vision-language-action model built on Gemini 2.0 that controls robots directly. On regulation, AI embedded in an already regulated product falls under Annex I of the AI Act, applicable on 2 August 2028 since the Omnibus (Regulation (EU) 2026/1744). Machinery follows a specific route, see the FAQ.
Concrete example
Illustrative case: a precision-engineering SME with 120 employees considers a cell where a camera-guided robot arm sorts and loads parts of varying shapes, until now handled by hand. The supplier says the arm learns "by example", without programming. Before signing, the company asks three questions: under what conditions does the arm stop on contact with an operator, who is responsible for the compliance of the cell (the supplier or the company), and what happens to the system's behaviour after a model update. It obtains a three-week trial on its own parts, a safety file for the complete cell and a commitment that no update will be deployed without written validation. The project starts on a single workstation, before any extension.
Comparison
| Generative AI | Agentic AI | Physical AI | |
|---|---|---|---|
| What it produces | Content (text, image, code) | A series of software actions towards a goal | A movement or a machine command |
| Environment | Digital | Software, data, connected tools | The real world: objects, places, people |
| Consequence of an error | Wrong content, which a human can discard | A wrong action in a system | Material damage or injury |
| Key technical constraint | Quality and accuracy of answers | Reliability of tools and chains of steps | Real time, sensors, safety, simulation to reality |
| Regulatory frame (EU) | AI Act: transparency, general-purpose models | AI Act depending on use, GDPR | AI Act Annex I (products), Machinery Regulation |
| Example | Drafting a meeting summary | Handling a complaint end to end | Robot arm sorting parts |
FAQ
What is physical AI?
Physical AI refers to AI systems that perceive the real world through sensors, reason about it and act on it through machines: robots, autonomous vehicles, industrial equipment. The term is mainly used by NVIDIA and has no normative definition.
What is the definition of physical AI?
According to NVIDIA, it is AI that lets autonomous machines perceive, understand and perform complex actions in the physical world. IBM describes it as AI models connected to sensors, actuators and control systems that act, then learn from the results.
What is the difference between physical AI and generative AI?
Generative AI produces digital content (text, image, code) that a human can choose not to use. Physical AI drives a machine that acts on the world: a mistake can damage a product or injure someone. Generative models are nevertheless a key ingredient of physical AI.
What is the difference between physical AI and agentic AI?
Agentic AI pursues a goal using software tools. Physical AI gives an agent a body: it acts on real objects, with real-time and safety constraints that software alone does not face.
What is a world model?
It is a model that learns how an environment evolves, in order to simulate situations. NVIDIA offers Cosmos (January 2025), world foundation models that generate realistic synthetic data to train and evaluate physical AI models when real data is slow, costly or dangerous to collect.
Who is working on physical AI?
NVIDIA (Cosmos, simulation platform), Google DeepMind (Gemini Robotics, March 2025) and robot makers such as Figure AI, Agility or 1X, listed by NVIDIA among the first adopters of Cosmos. Vendors' claims should be read as commercial positioning.
Does the AI Act apply to physical AI?
Yes, when the AI is embedded in an already regulated product (Annex I): application on 2 August 2028 since the Omnibus. For machinery, the Omnibus moved the Machinery Regulation to Section B of Annex I: according to an analysis of the text, sector-specific delegated acts (applicable by 2 August 2028) carry the requirements instead of the full high-risk regime. The Machinery Regulation (EU 2023/1230) itself applies from 20 January 2027.
Can an SME take an interest in physical AI?
Yes, through suppliers: cobots, mobile robots, camera-based quality control. The right starting point is a precise problem (handling, sorting, inspection) and a bounded pilot, not the purchase of a "general-purpose" robot.
See also
Further reading
Sources
- What is Physical AI?, NVIDIA (glossary). https://www.nvidia.com/en-us/glossary/physical-ai/
- NVIDIA Launches Cosmos World Foundation Model Platform to Accelerate Physical AI Development, NVIDIA, 6 January 2025. https://nvidianews.nvidia.com/news/nvidia-launches-cosmos-world-foundation-model-platform-to-accelerate-physical-ai-development
- What is physical AI?, IBM. https://www.ibm.com/think/topics/physical-ai
- Gemini Robotics brings AI into the physical world, Google DeepMind, 12 March 2025. https://deepmind.google/blog/gemini-robotics-brings-ai-into-the-physical-world/
- Regulation (EU) 2026/1744, Articles 2 and 3 (sectoral amendments, machinery), euaiact.com. https://www.euaiact.com/digital-omnibus-ai/sectoral-amendments
- EU Machinery Regulation (EU 2023/1230), application on 20 January 2027, Baker McKenzie. https://www.bakermckenzie.com/en/insight/publications/resources/product-risk-radar-articles/machinery-regulation