Korea to launch independent physical AI brain development project next year
The government will begin developing a physical AI foundation model, the 'brain' of humanoid robots, starting next year. The Ministry of Science and ICT announced it will build a physical AI version of the 'independent foundation model' project, which developed GPT-style large language models with homegrown technology. This follows up on a plan announced in July by Deputy Prime Minister and Science and ICT Minister Bae Kyung-hoon to 'create a physical AI foundation model within three years.'
The project will proceed on a two-track basis: two consortia will be selected through an open call and funded without a mid-term evaluation. The parallel-track approach was adopted to allow exploration of different learning methods, including imitation learning, which trains robots on human motion data collected via body cams, and reinforcement learning, where robots learn behavioral rules on their own through trial and error without human intervention.
The project will build on prior research. Once early models (prototypes) are unveiled under the 'Physical AI Leading Technology Development' project involving LG Electronics, Maum AI, KT, and Robotis, two newly formed consortia will carry research and development forward. The completed physical AI model will first be applied to the postal logistics sector overseen by Korea Post.
Realworld has open-sourced its proprietary robotics foundation model 'RLDX-1' and ranked first in the world on performance evaluations across eight global benchmarks.
While domestic companies remain in the early stages, global big tech is expanding its network of partnerships and growing its influence. NVIDIA is pursuing a strategy of offering a full-stack service for robot development, spanning not only its robot foundation model 'Isaac GR00T' but also the 'Isaac Sim' simulation environment, the 'Omniverse' digital twin platform, and the compact high-performance computer 'Jetson Thor.' Developers such as Field AI and Skild AI are building general-purpose robot brains using NVIDIA's Cosmos world model and Isaac simulation framework, and domestic player Robotis has also demonstrated task-performing capabilities for its humanoid 'AI sapiens' using NVIDIA technology.
Google has focused on the robot brain. 'Gemini Robotics 2,' a vision-language-action (VLA) model unveiled in July, converts visual information captured by cameras and verbal commands into actual movement, capable of controlling everything from robot arms to humanoids. It is conducting joint research with Apptronik, Boston Dynamics, and Agile Robots. In Korea, Hyundai Motor Group subsidiary Boston Dynamics is a strategic partner of Google DeepMind, in a structure where Boston Dynamics provides the body and Google DeepMind the brain.
A domestic robotics industry official said, "Google and NVIDIA are so interested that whenever someone posts a humanoid demo video, they ask which platform it was trained on," adding, "For now, it is difficult to adopt a domestic foundation model."
Kim Eun-chan, a professor of information systems at Hanyang University, said, "Rather than chasing general-purpose models where global big tech already excels, it would be better to build specialized physical AI models focused on industries where we have strengths, such as manufacturing, shipbuilding, logistics, and automotive."
