[DigitalToday reporter Chi-kyu Hwang] Physical AI company Real World (RLWRLD) is showcasing its in-house robotics foundation model (RFM) RLDX-1 and speeding up its push into the domestic market.

Real World recently held an event, Dexterity Night in Seoul, to share its physical AI vision and laid out RLDX-1 performance and a partnership-based business strategy.

The event was part of a global tour that wrapped up in Seoul after stops in San Francisco, Tokyo and Taipei.

Real World CEO Ryu Jung-hee (류중희) stressed a vision of fully automating manufacturing-site labor with AI robots, work he said is impossible without human-level dexterity. "Even the most automated factories in Korea only reach 75 percent automation," he said. "The remaining 25 percent is still handled directly by people. Japan, China and the United States are not much different. Nearly half of labor in the overall labor market still depends on human hands." He added that tasks such as assembling small parts or body components cannot be done without human-level dexterity and said Real World aims to automate this "last mile" of labor with AI.

According to him, RLDX-1 focuses on implementing human-level AI dexterity. The core elements are data, model and deployment.

On data, Real World built its own data pipeline. Ryu said it collects human data by having a robot follow human movements. He said Lotte Hotel is cooperating by providing hotel space free of charge to capture movements of professional hoteliers. He said the company analyzes collected data with AI and converts human-level dexterity into robot dexterity.

Real World also uses a synthetic data pipeline to advance its data.

Ryu said securing only 20 percent of the overall data allows AI to augment the remaining 80 percent to create 100 percent data. He stressed that the company built a model with little change in quality even when trained this way and verified that Real World's synthetic data pipeline outperforms Nvidia's.

On deployment, Real World places heavy emphasis on proof-of-concept tests in cooperation with multiple hardware partners. Ryu said domestic and overseas companies such as Nvidia, Hanwha and Lotte are investing in Real World or conducting joint experiments.

Real World also stresses that RLDX-1 has overcome fundamental limitations of existing robot AI models.

According to Real World chief scientist Jinwoo Shin (신진우), interest is rising in foundation models in robotics that handle various tasks with a single model.

Examples include Nvidia's GR00T series. These models connect scene understanding and language reasoning capabilities of vision-language models (VLMs) to robot motion control, but have limits because they make judgments based on only a single screen.

The solutions can be summarized in 3 main points.

To catch a moving object, a robot must read changes in motion over time, which is impossible from a single still image. Some tasks require past context, such as remembering what is inside a box. Other tasks, such as inserting a plug into a socket without looking or picking up a card, cannot be done without feedback on touch and force.

Shin said RLDX-1 is the first among existing robotics foundation models to incorporate all three in a single model. He stressed that in both simulation and real-robot tests, RLDX-1 outperformed all competing models, including Nvidia GR00T. He said it was the first robotics foundation model to exceed 78 percent on a simulation benchmark and recorded more than 10 times GR00T's success rate on precise manipulation tasks requiring human dexterity.

Real World said there is broad agreement that dexterity matters, but there is still no clear answer on what level is needed for commercialization. It is working with Nvidia to create a global dexterity standard. It recently introduced DexBench, a dexterity standard metric, and aims to spread it as a global standard in cooperation with Nvidia.