[Financial News] NC AI has officially entered the global physical AI competition by unveiling its research achievements in the ‘World Foundation Model’ (WFM), a critical technology in robotic intelligence. The company demonstrated success in the world model field, which typically requires large-scale computational resources, by attaining about 80% of state-of-the-art (SOTA) performance while reducing the use of computing resources to just 25%, proving its efficiency.
NC AI announced on the 16th that it has successfully demonstrated its WFM and confirmed its significant performance in high-difficulty robot manipulation tasks. World models are key technologies in physical AI that enable robots to understand, predict, and act within the physical world, attracting substantial investments from global big tech companies.
One of the major technical challenges in the robot AI industry is the ‘Sim2Real’ gap, where robots trained in virtual environments behave unpredictably due to minor frictions or physical variables in real-world settings. To address this, global companies are investing heavily in developing robot foundation models.
NC AI approached this issue using a world model capable of precisely predicting real-world physical laws, beyond mere visual imitation. Unlike existing world models that infer actions through a Vision-Language Model (VLM) after generating images, NC AI’s model is designed to generate actions directly from latent space information before image creation. This approach reduces the image generation and inference stages, enhancing computational efficiency while maintaining accuracy.
Performance metrics indicated practical applicability. Testing based on 24 complex robot manipulation tasks showed average performance at approximately 70% of top global models. In particular, for 18 key tasks directly linked to field application, it achieved about 80% success rate compared to top-performance models like NVIDIA’s Cosmos.
Another notable feature of this research is resource efficiency. NC AI performed world model training using only about 25% of the GPU resources needed for fine-tuning top global performance models. This demonstrates the technological feasibility at a global top level without massive infrastructure investments through an optimized learning structure.
NC AI also addressed the data problem necessary for robot learning by establishing an environment that quickly generates large-scale industrial data through a synthetic data generation pipeline based on world models. They state that an A100 GPU can generate about 10 seconds of video in approximately 80 seconds, and using 100 H100 GPUs, it can generate about 10,000 hours of synthetic video data in 11 days.
This technology is significant in that it can generate various scenarios needed in actual industrial fields, like semiconductor clean rooms, steel processes, and shipyard operations, in a virtual environment for robot learning.
NC AI is currently participating in the ‘K-Physical AI Alliance’ along with companies like Realworld, Samsung SDS, Seamless, Rainbow Robotics, research institutions like ETRI, KETI, the Korea Automotive Technology Institute, and academic entities like KAIST, Seoul National University, and Korea University. Their strategy is to expand industrial field applications by building an ecosystem that spans from precise physical simulations to robot validation.
Lee Yeon-soo, the CEO of NC AI, stated, “The significance of our WFM research achievements lies in proving substantive validity at a global top-tier level through precise physical understanding and optimized learning architecture, moving beyond traditional robot AI development methods reliant on vast computational resources. We will cultivate this as a core competitive edge leading the global physical AI supremacy.”