NC AI announced on the 16th that it has successfully demonstrated its World Foundation Model (WFM), a core technology for robot intelligence.
The biggest challenge facing the global physical AI industry is the so-called "Sim2Real gap," where robots that perform flawlessly in virtual simulations malfunction when confronted with subtle friction and physical variables in the real world. While U.S. and Chinese big tech companies are racing to pour astronomical investments into robot foundation models to solve this problem, NC AI has presented a solution through its WFM, which goes beyond visual imitation to accurately predict the sophisticated physical laws of reality, the company explained.
While conventional WFMs generate video, process it through a Vision Language Model (VLM) for inference, and then select actions, NC AI's WFM generates actions directly from latent space information—the data that exists before video generation—achieving both efficiency and accuracy. By eliminating the video generation and inference stages, the model increases speed while improving action accuracy through extensive training on data generated by a high-precision physics engine.
Combined with the extensive virtual world-building expertise accumulated over more than 20 years of operating large-scale MMORPGs since the NCSoft era, and NC AI's proprietary "VARCO 3D"—Korea's only 3D generation model—the company can implement 3D environments and simulators closely resembling the real world.
The most notable aspect of this research is "resource efficiency." NC AI successfully trained its WFM using only 25% of the GPU resources required for fine-tuning global top-performing models.
Performance metrics have also reached levels suitable for practical application. In tests measuring simulator-level precise predictions across 24 high-difficulty robot manipulation tasks involving complex robotic arm movements, the model achieved 70% of state-of-the-art (SOTA) performance across all 24 tasks. Notably, when measured on the top 18 core tasks directly related to field deployment and commercialization, it recorded task success rates reaching 80% of top-performing models such as NVIDIA Cosmos.
This demonstrates that NC AI has achieved global top-tier technical validity through optimized post-training without massive infrastructure costs, positioning it as a strong competitive advantage for democratizing robot AI technology and enabling cost-efficient industrial adoption.
NC AI also plans to implement a large-scale synthetic data generation pipeline through its world model to fundamentally solve the "data scarcity" problem essential for robot learning. Previously, collecting real video footage of various real-world variables—such as snow falling in factories, nighttime logistics centers with lights off, or unexpected human intervention—required enormous time and costs.
However, in NC AI's WFM environment, video data of such extreme conditions can be generated in large volumes simply through prompt manipulation. With high efficiency of 80 seconds to generate a 10-second video on a single A100 GPU, utilizing 100 H100 GPUs—which offer three times the performance of A100—enables generation of 10,000 hours of new synthetic video data in just 11 days.
Based on this capability, NC AI plans to supply "domain-specific customized" synthetic data tailored to Korean manufacturing characteristics, including semiconductor cleanrooms, steel processing, and shipyard blocks, perfectly complementing gaps in actual field data.
"This WFM research achievement holds great significance in proving global top-tier practical validity through precise physics understanding and optimized learning architecture, breaking away from conventional robot AI development methods that relied solely on massive computational resources," said Yeonsu Lee, CEO of NC AI. "Going forward, we will firmly establish a Korean industry-specific robot ecosystem together with the K-Physical AI Alliance based on NC AI's unrivaled world model technology, and develop it into a core competitive advantage leading global physical AI dominance."