Korea to start robot foundation model R&D in 2027 with two consortia

South Korea’s Ministry of Science and ICT will begin developing robot foundation models and world models in 2027, along with a library of physical-world training data, Song Chang-jong, head of the ministry’s Device AX Innovation Team, said on September 22, 2026 at Humanoids Summit Seoul. Money Today Network (MTN) reported on September 30 that the programme will fund two consortia in parallel, chosen by open call, and that the first models will go to postal logistics run by Korea Post.

Why it matters: Korea has spent 2026 funding robot hardware and state procurement. This programme is aimed at the model layer that decides what those robots do, and Song said Korea has about three years before that layer is settled elsewhere.

Two consortia, no mid-term cut

MTN contrasts the structure with Korea’s LLM foundation model project, which used a concentrated-support model. Here two consortia keep their funding without an interim evaluation, so that different training approaches can be tried side by side. MTN names two in current use: imitation learning from human motion data collected with body cameras, and reinforcement learning, in which a robot works out its own rules of behaviour by trial and error. The report does not say whether each consortium will be tied to one method.

The programme builds on prototypes from the ministry’s existing Physical AI Pioneer Technology Development project, whose participants include LG Electronics, Maum AI, KT and ROBOTIS. MTN lists LG Electronics, RLWRLD, Maum AI and NC AI as likely contenders on the software side; no consortium has been selected. Song described a large R&D effort that brings AI models, data, robots and demand-side companies together, coordinated through a Physical AI Alliance of industry, universities and research institutes, with other ministries including trade and industry and SMEs and startups involved.

A three-year window and no budget figure

Song said forecasts suggest the “golden time” in physical AI could close within three years, and that budgets have not been fully confirmed. That puts the start date, structure and first customer on the record while leaving the scale open. Korea’s 2027 budget proposal, approved by cabinet on September 1, raises the physical AI line from ₩900 billion to ₩3.1 trillion; how much of that goes to foundation models was not stated.

The field Korea is entering is not empty. As of late September, at least eight developers had a named generalist robot model, and NVIDIA’s Isaac GR00T N1.7 ships open weights that a Korean team could fine-tune rather than build from scratch. A domestic model has to justify itself against that. The postal pilot is the answer the ministry has so far: a defined task with a public buyer from the first day.

Patent strategy before the research

On September 30, the Intellectual Property Office and the Ministry of Science and ICT presented a plan at the 13th Science and Technology Ministers’ Meeting to set patent strategy before national R&D projects start rather than after. Physical AI goes first. A national patent strategy map, which analyses global supply chains and patent distribution to steer research direction, is to be built for physical AI from the second half of 2026 and then extended to other strategic technologies. The plan also calls for a “K-Patent Pool” built in stages, in which Korean patents in areas such as humanoids are used jointly; Patent Lawyer Magazine describes it as aggregating standard-essential patents.

Patent Lawyer Magazine also reports about ₩4 trillion planned for 2027 across seven SEED technology projects. Those cover small modular reactors, fusion, renewable energy, quantum, aerospace, advanced biotechnology and supply chains; physical AI is not one of the seven. The patent work matters for robots because Korea is starting from behind: its share of humanoid patent filings at the five largest patent offices was 4.6% in the first half of 2026, against China’s 60.8%.