Upon entering the ISC building at LG Science Park in Magok, Gangseo-gu, Seoul, on Tuesday, LG's humanoid robot "Cloid" greeted reporters. Cloid displayed a smiling expression on the OLED screen mounted on its head and interacted naturally with people, exchanging fist bumps. Next to Cloid, Robotis's robot "AI Worker" mimicked a researcher's movements, picking up objects and moving them into a basket. From paper boxes to plastic packaging, it delicately adjusted the force of its five fingers according to the material and shape of the objects, accumulating data on finger movements, force control methods, and responses to different objects in the process. A researcher on site explained, "Future data collection needed to build the robot foundation model (RFM) will also be carried out in this way."
The government has launched the development of "physical AI leading technology" worth 34 billion won together with LG Electronics, Robotis, and the Korea Advanced Institute of Science and Technology (KAIST). The plan is to develop core source technologies such as world models and robot foundation models (RFM) so that AI-based robots can stably perform tasks in the real world, where many variables exist, and to deploy these at domestic manufacturing and logistics sites to secure competitiveness. The Ministry of Science and ICT (MSIT) and the Institute of Information & Communications Technology Planning & Evaluation (IITP) unveiled the plan at the "Physical AI Leading Technology Development Project Kickoff Meeting" held at LG Science Park on the same day, stating, "We will achieve world-class performance within a short time."
The government will invest a total of 34 billion won over the next two years to secure proprietary world model source technology and verify domestic simulator technology linked to it. A world model is a platform that generates large volumes of synthetic data to supplement real-world behavioral data and allows actual verification processes to be replaced with virtual simulations. When simulation data learned from the world model is linked to the RFM, which serves as the robot's "brain," the robot can respond to changing environments and perform various tasks. The government and companies plan to collect data on a scale of 110,000 hours through the verification process.
The project involves a consortium of 10 industry-academia-research institutions, including LG Electronics, Maum AI, Holiday Robotics, Robotis, Alchera, Crowdworks, KT, KAIST, Seoul National University, and the Telecommunications Technology Association. The consortium will build a verification pipeline that runs from "world model learning to RFM linkage to verification and performance evaluation to case analysis and retraining," and will conduct a total of four rounds of repeated verification over two years.
The goal is to raise the robots' task performance to world-class levels. The consortium said it would raise the final action success rate of robots applying its independently developed world model by more than 20 percentage points compared with robots without the world model. This exceeds the 14.5 percentage point improvement of OpenGVLab, currently considered the global best. In particular, the consortium plans to secure differentiated competitiveness by enhancing understanding of manufacturing environments.
To this end, the consortium will use multimodal models such as KT's "Mi:dm K AI" and the LG AI Research's "Exaone 4.5" as the foundation for developing the world model and RFM. It also plans to advance the technology by analyzing existing global models such as Nvidia's "Cosmos," combining their strengths and supplementing their limitations. Kim Young-joon, head of LG Electronics' Artificial Intelligence Lab, said, "We will share the development results in an open-source system so that they can also be used at the industrial sites of small and medium-sized enterprises."
