Inside a factory, a worker wearing a headset commands a robot equipped with agentic artificial intelligence: "Darvis, Process No. 6 looks fine visually, but the system is flagging a problem. Figure it out." The worker wears only a helmet, with nothing in either hand. Darvis analyzes the situation briefly and responds: "Shuttle No. 6 appears empty to the naked eye, but the loading sensor seems to be detecting incorrectly." It then flags the need for inspection and a reboot, moves the faulty equipment to a maintenance zone, and switches the mode to allow manual inspection upon arrival. This is a scene from KAIROS, a physical AI-based dark factory demonstration platform built by the Korea Advanced Institute of Science and Technology (KAIST).
The Ministry of Science and ICT (MSIT) and KAIST unveiled the system on Monday, announcing they would move forward with a physical AI strategy that integrates and controls an entire factory through a single intelligent operating system (OS). According to KAIST, the core of KAIROS lies in software and an OS that runs an entire factory like an orchestra. People manage processes from outside the factory alongside AI, while robots move according to those instructions. KAIST presented this as the central concept of a "dark factory" — a fully unmanned production facility.
During the live demonstration, the focus went beyond simple question-and-answer exchanges to show AI overseeing multiple robots and equipment in an integrated manner. On the digital twin screen, autonomous mobile robots (AMRs), warehouses, production equipment and humanoids were connected as a single system. The AI agent Darvis not only summarized equipment status in its responses but immediately executed actions as well.
Another strength of KAIROS is its optimization of multi-robot operations. Jang Young-jae, the KAIST professor who led the presentation, said: "In existing systems, deadlocks between robots occur frequently, but by applying our proprietary reinforcement learning-based algorithm, robots can yield to each other, adjust priorities and move organically." He added that "in simulations involving 200 robots, performance improvements of 30 to 50 percent were achieved compared to conventional systems."
The government plans to use the demonstration confirmed at KAIST to flesh out its strategy for securing core competitiveness in physical AI. KAIST will divide roles with Jeonbuk National University, taking charge of factory operation software and the OS, with plans to expand the integrated solution to the Gyeongnam and Ulsan regions. In particular, the initiative will also pursue free cloud distribution of an "AI Factory Manager for All" service so that small and mid-sized enterprises can verify the effects of robot adoption without expensive foreign simulation solutions. The long-term vision extends beyond exporting products to exporting the AI systems and frameworks that design, build and operate entire factories.