Doosan Robotics CEO: AI agents, not humans, will run robots within a decade

"In five or ten years, I think there will be no disagreement that it is agents, not humans, who operate robots."

Kim Min-pyo, CEO of Doosan Robotics, said Monday that the robotics industry in the era of physical AI will shift from direct human operation to AI agent-driven control. The company's strategy is to gather data from real industrial settings through its collaborative robots and expand that foundation into intelligent solutions and agents.

Kim made the remarks at the 2026 Global Mechanical Technology Forum held at the Korea Science and Technology Center in Gangnam, Seoul, where he delivered a keynote titled "The Present and Future of Physical AI at Doosan Robotics."

"Traditional industrial robots require you to spell out the how — exactly what needs to be done and in what way," Kim said. "But with agents and robots, you only need to state the goal, and they replan, learn and become skilled workers until that goal is achieved." He added that once the world moves from expressing how to do something to simply instructing what needs to be done, robots will become as accessible as rice cookers or dishwashers — devices anyone can operate without specialized knowledge.

To that end, Doosan Robotics is restructuring its robot software into a form that agents can handle directly. "We are rebuilding our robot software so that agents can operate it," Kim said. "Because we own the lowest-level controller, we can make use of all the logs, data and sensor outputs that controller generates."

However, Kim said significant hurdles remain before physical AI can be realized in the real world. "Digital intelligence is exploding right now, and the intelligence within it is becoming remarkably sharp — but in fact, even the smartest model cannot pick up a cup the way a human does," he said. "Physical AI is not an extension of large language models. It is better viewed as a separate problem entirely."

Kim also pointed to the difficulty of acquiring and validating data in the physical world, and identified the "sim-to-real gap" — the divide between simulated and actual environments — as a key challenge. "Contact, friction, retention, deformation and recovery from failure are still areas that are difficult to replicate in simulation," he said. "The contact force, friction, delicacy and success judgment that industrial sites actually require remain open challenges."

Doosan Robotics has chosen to narrow its focus rather than try to solve every problem at once. "Our strategy is to narrow the problem," Kim said. "Instead of trying to solve everything, we chose the hardest skilled-labor tasks and are pursuing a strategy centered on achieving completeness in specific tasks — deriving the form factor and intelligence those tasks require."

The company plans to gather data from real industrial environments through its collaborative robots and scale that into intelligent solutions. "Our perspective is to develop a humanoid robot by deriving a compact form factor that optimizes these tasks in a highly practical way," Kim said. "Doosan Robotics, which has been hardware-centric, is now incorporating AI and solutions as it prepares for a true UX innovation — one where future agents configure and operate robots."