NC AI said it will move to crack the global physical artificial intelligence (AI) market by unveiling research results on the World Foundation Model (WFM), a core technology for robot intelligence.
NC AI said on the 16th that it proved its competitiveness in the global physical AI market by successfully demonstrating WFM, regarded as a key technology for implementing robot intelligence. It added that in the world model research field, which requires large-scale computing resources, it used its in-house research infrastructure to train and validate the model and secured performance at a level viable for real-world application.
In today's physical AI industry, the "Sim2Real (simulation-to-reality) gap," in which robots trained in virtual environments malfunction due to subtle physical variables in the real world, is cited as a major challenge. As global big tech corporations continue to invest heavily in robot foundation model research, NC AI proposed a solution with WFM technology capable of precise physical prediction.
Conventional WFMs generate video and then have a vision-language model (VLM) infer actions to decide behavior. In contrast, NC AI's WFM applies a model that generates actions directly from latent space information, a stage prior to video generation. This reduced the video generation and inference steps to boost processing speed, and it improved action accuracy by using training data generated with a high-precision physics engine.
NC AI said it also implemented a 3D simulation environment that closely resembles reality by combining large-scale virtual world-building technology accumulated since the NCSOFT era with its in-house 3D generative model VARCO 3D.
The standout aspect of this research result is high resource efficiency. NC AI trained the WFM using about 25% of the graphics processing unit (GPU) resources required to fine-tune a world-leading model.
It also achieved meaningful results in terms of performance. In tests on 24 high-difficulty manipulation tasks controlling complex movements of a robot arm, it secured about 70% of the performance of the global state-of-the-art (SOTA) on an overall task basis. In particular, on the top 18 key tasks, it recorded a success rate of about 80% compared with leading models such as Nvidia Cosmos.
NC AI plans to build a large-scale synthetic data generation pipeline to solve the data shortage problem needed for robot training. Previously, securing diverse environmental data—such as factories with snowfall or nighttime logistics centers—required significant time and expense, but in a WFM environment, various situational data can be generated with prompt input alone, it said.
According to the company, on a single Nvidia A100 GPU, it takes about 80 seconds to generate a 10-second video. Using 100 H100 GPUs, it can generate about 10,000 hours of synthetic video data in about 11 days.
Based on this technology, NC AI plans to provide synthetic data tailored to Korea's manufacturing environments, such as semiconductor clean rooms, steel processes, and shipyards, to solve the on-site data shortage.
It is also joining the "K-Physical AI Alliance," which includes corporations such as RealWorld, Samsung SDS, CMES, Config Intelligence, Rainbow Robotics, NdotLight, and FunctionBay; research institutes such as the Electronics and Telecommunications Research Institute (ETRI), the Korea Electronics Technology Institute (KETI), and the Korea Automotive Technology Institute; and academia including KAIST and Seoul National University, to help build the ecosystem.
Lee Yeon-su, head of NC AI, said, "This WFM research result is meaningful in that it proved the validity of world-leading technology through precise physical understanding and an optimized training structure, moving away from the conventional robot AI development approach that relied on massive computing resources," and added, "Together with the K-Physical AI Alliance, we will build a robot ecosystem tailored to Korean industry and strengthen global physical AI competitiveness."