Physical AI, artificial intelligence takes shape in the factory

On the same frontier moves Sinapsi.ai, with an application in precision mechanics, with a solution that combines artificial vision and AI models trained on the customer's real defects, integrating with CAD/CAM software to make automatic corrections in real time. A solution that, the company explains, could lead to up to a 40% reduction in scrap and a 50% reduction in quality control time. Estimates that show well the industrial value of the subject: bringing quality control into the process, instead of leaving it only downstream.

The second strand is robotics outside the traditional cell. Eagleprojects presents a vision of Physical AI applied to robots and drones capable of perceiving, deciding and acting in the field. The main areas are autonomous patrolling for the security of industrial sites and critical infrastructure, and inspections in confined spaces or hazardous environments. Here the value lies not in the individual robot, but in the integration with sensors, drones, GIS systems, digital twin and control rooms. Robotics becomes an operational extension of human control: it enters where humans are most at risk, collects data and links it to decision-making.

Then there is the frontier of cobots that learn from humans. SUPSI, the University of Applied Sciences and Arts of Southern Switzerland, is working in the RoboGym area of its laboratory on the European Fluently project on an environment in which people and robots can train to interact more fluidly and improve industrial processes. The project aims to develop intelligent interfaces that enable industrial robots to better understand human behaviour and cooperate more effectively. In other words, the collaborative robot must not only be safe next to the operator, but must better understand the operator's intentions, gestures and operational needs.

In terms of research, the University of Pisa and the Italian Institute of Technology are taking soft robotics to the next level with the Pisa/IIT SoftHand, a simple, robust and effective robotic hand for gripping and soft handling. An example of physical intelligence embedded in the body of the robot, because not all the ability to adapt to the real world lies in the algorithm, part of it is in the mechanics, compliance and morphology of the device.

Alongside visible robotics, however, there is also less conspicuous but decisive AI: that which organises industrial knowledge. Cefriel, for example, is working on the issue of corporate knowledge augmented by artificial intelligence, starting from a very real problem: skills dispersed among documents, emails, archives and people, difficulties in onboarding, loss of know-how and slowing down decision-making processes. And it is in this direction that the Arca platform will be presented at the fair, which was created with the aim of making the company's information assets accessible and updatable, transforming them into an operational tool for technicians, designers and new entrants.

However, the issue of trust remains central. AI4I, the Italian Institute of Artificial Intelligence, places the issue within the IT4LIA AI Factory project, a national initiative designed to make computing infrastructures, services and skills available to companies, start-ups, public administrations and the world of research. In manufacturing, the focus is also on the validation of complex autonomous systems: robots, drones and intelligent machines must be tested not only in ordinary cases, but also in rare and critical scenarios in which an error can block a process, generate a waste or create an operational risk. If AI is to move in the physical world, it is not enough that it works in most cases. It must be verifiable, robust and integrable into industrial processes without introducing new fragilities.

The picture that arrives from Parma is therefore less futuristic and more industrial: not generalist humanoid robots ready to occupy the factory, but specific systems that observe better, decide more quickly, assist the operator, reduce errors and bring automation into more variable contexts. The promise of Physical AI is this: to give algorithms a body. The challenge for companies will be to turn demos into reliable, scalable and cost-effective solutions. Because in manufacturing, innovation is not measured by technological fascination, but by the ability to withstand the pace, constraints and responsibilities of real production.