"The Core of Robot Intelligence"... NC AI Demonstrates World Foundation Model

First Domestic Demonstration of WFM

NC AI, the artificial intelligence (AI) specialist division of NCSoft, announced on March 16 that it has successfully demonstrated the "World Foundation Model (WFM)", which is considered the core of robot intelligence. This marks the first demonstration by a domestic game company in the field of world models, an area that has been led by global big tech companies such as Google.

NC AI, an AI specialist company under NCSoft, announced on the 16th that it has successfully demonstrated the 'World Foundation Model (WFM)', a core of robot intelligence. Video predicted by NC AI's WFM (left) and video of a robot actually moving in the simulator (right). Provided by NC AI

NC AI introduced WFM as a solution to robot malfunctions by going beyond visual imitation to accurately predict the sophisticated physical laws of the real world. Even when robots train on vast amounts of data in virtual simulations, they can malfunction in the real world due to factors like subtle friction or physical variables. NC AI’s WFM secures both efficiency and accuracy by employing a model that generates actions directly from latent space information, which precedes video generation, instead of generating actions from videos themselves. In addition, by removing the video generation and inference steps, the model’s speed is increased, and it is trained on a large volume of data generated using a high-precision physics engine.

NC AI stated that it has achieved resource efficiency and performance metrics with WFM. The company trained WFM using only 25% of the graphic processing unit (GPU) resources required to fine-tune globally top-performing models. In experiments predicting outcomes for 24 challenging robot manipulation tasks involving complex movements of a robotic arm, WFM achieved 70% of the performance of state-of-the-art (SOTA) models across all 24 tasks. For the top 18 tasks, which are more closely related to real-world deployment and commercialization, WFM recorded a success rate equivalent to 80% of the best-performing models, such as NVIDIA’s Cosmos.

NC AI plans to address the 'data scarcity' problem in robot learning by implementing a large-scale synthetic data generation pipeline using its world model. Collecting diverse video data to reflect real-world variables—such as snow-covered factories or unlit night-time logistics centers—typically requires immense time and cost. However, in the WFM environment, a large quantity of video data representing extreme conditions can be generated simply by manipulating prompts. NC AI stated, "We will supply 'domain-specialized and customized' synthetic data tailored to the characteristics of Korea’s manufacturing sectors, such as semiconductor clean rooms, steel processes, and shipyard blocks," adding, "WFM will serve as a foundational technology within the value chain, enabling various advancements."

Lee Yeonsoo, CEO of NC AI, commented, "This WFM demonstrates a global top-tier level of practical efficacy, achieved not by relying solely on massive computational resources as in conventional robot AI development, but through precise physical understanding and an optimized learning structure. Going forward, based on NC AI’s unparalleled world model technology, we will solidify a Korea-specific industrial robot ecosystem and lead the global physical AI landscape."

This content was produced with the assistance of AI translation services.