Doosan Robotics leads ₩98.9bn push to run cobots on Korean AI chips

Doosan Robotics said on October 6, 2026 that it will lead two government R&D projects worth about ₩98.9 billion in total, about ₩68.1 billion of it government funding, to build collaborative robots and an industrial humanoid that run AI on Korean-made chips and to automate welding of nuclear components. Why it matters: it is a sizeable state-funded test of whether Korean on-device AI chips can carry robot perception and control without an external PC or cloud server.

Some English-language reports misread the Korean figure as ₩989 billion. The Korea Herald reports ₩98.9 billion and attributes the remaining ₩30.8 billion to Doosan.

What will the robots run on?

The first project puts a neural processing unit (NPU) and real-time control onto a domestic AI system-on-chip mounted in the robot. According to eToday’s account of the company announcement, the robots will process camera, tactile, force/torque and voice data in real time and perceive, decide and control on site without connecting to an external PC or the cloud. The Korea Herald reports Doosan’s pitch for this: lower latency and lower security risk in defence, nuclear and biopharmaceutical settings.

The hardware line-up is three cobots with 5 kg, 10 kg and 20 kg payloads and an industrial humanoid capable of dual-arm work and omnidirectional movement. Partners in the 54-month project include AiDin Robotics, which makes force/torque sensors and robot hands, the NPU designer Mobilint, and the Korea Electronics Technology Institute (KETI). The target is commercial standard products in 2031, after a 72-hour continuous demonstration and international safety certification.

Mobilint is also working on a separate robot chip under the same K-On-Device AI programme. On October 1, SEMIFIVE signed a turnkey contract to develop it, aimed at agricultural robots and outdoor use where links to external servers are limited, with LPDDR6 memory, PCIe Gen6 and UCIe-S chiplet interconnect. Neither announcement says that chip will go into Doosan’s robots.

What is the nuclear welding project?

The second project, running 45 months, applies AI and a digital twin to high-difficulty welding to nuclear-equipment standards. The robot recognises the workpiece with a camera, corrects for its position, learns skilled welders’ movements and generates its own weld paths. The targets are a path match of at least 95% against skilled welders, error within ±3 mm, and multi-layer welding automation that cuts work time by 50% or more. Trials will run at Doosan Enerbility’s nuclear centre covering tack welding, preheating and main welding, before extension to shipbuilding, plant and defence work.

“We will combine on-device AI chips with our own cobot platform to build physical AI cobots that learn and reproduce even the tacit knowledge of skilled workers,” said Doosan Robotics president Park In-won.

Why does the chip matter more than the robots?

Doosan already sells cobots, and its chief executive set out a path to a 2028 industrial humanoid proof-of-concept last month. The new element is the compute. In China, AgiBot has started putting domestic chips into joint controllers, motion control and communications while its high-level model still runs on NVIDIA Thor, as we reported in China’s robot makers start swapping in domestic chips. The Doosan project aims to run perception and decision-making on a Korean chip inside the robot, which is the harder half.

The test is narrower than the framing. A 5–20 kg cobot doing vision-guided pick-and-place or weld-path generation is a bounded workload for an edge NPU; a dual-arm mobile humanoid is not. The 72-hour non-stop demonstration and safety certification are the milestones to watch, and the 2031 date puts commercial product five years out.