When Intelligence Becomes Physical

So-called physical AI combines models, sensors, robots and control systems to make intelligence operate in material environments. Unlike a text assistant, a robot must perceive changing conditions and act within safety constraints. Its learning depends on simulation and data collected during real operations, but the relationship between installations, data and AI capability is not automatic.

According to the International Federation of Robotics, factories worldwide had around five million industrial robots in operation at the end of 2025, 9% more than a year earlier. More than 600,000 new units were installed during 2025. China installed 195,000 robots in its domestic market, up 15%; Chinese manufacturers reached 55% of domestic sales, down from 57% the year before.

The figures describe a significant shift: China is both a major market and an increasingly prominent producer in its own industrial ecosystem. They do not, however, show that every robot collects data usable to train general models, or that those data are accessible to manufacturers. Ownership, interoperability, quality and usage rights remain separate variables.

An automated machine often follows defined tasks. A physical AI system instead aims to generalise across materials, objects or conditions that are not identical to those encountered before. This requires reliable sensors, quality data, simulation and feedback on the outcome of actions.

Robot makers and users can collect valuable experience: material variation, failed attempts, corrections and process times. But quantity is not enough. Data without labels, context or permission may be hard to use or impossible to reuse. The strategic question is who can turn factory experience into shared knowledge and better products.

For a European company, dependence may concern the robot, its software, support and access to data. Evaluating a supplier therefore means asking where data are processed, who can reuse them, how they can be exported and what happens if the service changes. In sectors with demanding processes, from eyewear to precision engineering, manufacturing expertise could become a resource for training systems suited to specialist tasks—if it is documented and governed.

Competition cannot be measured only by counting robots or companies that make them. The ability to connect operational data, outcomes and usage rights also matters. Are European manufacturers treating experience from their production lines as a strategic asset?

IFR figures concern industrial robots and installations; they are not a direct measure of AI datasets, their quality or technological sovereignty. The connection between automation and model learning is presented here as a question to investigate, not an established outcome.