NC AI said on Monday it demonstrated a World Foundation Model (WFM), a core technology for robot intelligence, and issued a challenge to the global physical AI market.

NC AI said it carried out training and validation using its own research infrastructure in the world model field, which requires large-scale computing infrastructure, and recorded success rates in key tasks that confirm the model's potential for practical use.

The company said existing WFMs generate video, reason over it with a vision-language model (VLM) and then select actions, while NC AI's WFM uses a model that generates actions directly from latent space information, which precedes video generation, securing efficiency and accuracy at the same time.

The standout aspect of the research results is resource efficiency. NC AI said it successfully trained the WFM using just 25 percent of the graphics processing unit (GPU) resources required to fine-tune global state-of-the-art models.

It also raised performance metrics to a level suitable for practical application. In tests of simulator-level precise predictions across 24 high-difficulty robot manipulation tasks controlling complex movements of a robot arm, it secured 70 percent of state-of-the-art performance across all 24 tasks. In results measured on 18 key tasks most directly tied to field deployment and commercialisation, it recorded a high task success rate of 80 percent of top-performing models such as Nvidia Cosmos.

NC AI said it also plans to implement, using the world model, a large-scale synthetic data generation pipeline that will fundamentally address data shortages, an essential element of robot training. It said collecting real video to capture variables that can occur in reality, such as a factory during snowfall, a night-time logistics centre with lights out, or unexpected human intervention, required huge time and cost. In NC AI's WFM environment, it said, prompt manipulation alone can generate large volumes of video data for such extreme conditions.

Based on this, NC AI said it will supply domain-specific, customised synthetic data tailored to the characteristics of South Korea's manufacturing industry, such as semiconductor cleanrooms, steel processes and shipyard blocks, to fully complement missing pieces of real on-site data.

NC AI CEO Yeon-su Lee (이연수) stressed, "Going forward, based on NC AI's unparalleled world model technology, we will work with the K-Physical AI Alliance to firmly build a robot ecosystem specialised for Korean industry, and further develop it into core competitiveness that leads global physical AI dominance."