South Korean Robotics Firm XYZ Unveils 'Field Data + Simulation' Training Loop, Claims 96% Success Rate for Folding Tasks

Hwang Sung-jae, CEO of South Korean robotics startup XYZ (엑스와이지), presented the company's development methodology for Physical AI at the "NVIDIA AI Day Seoul 2026" event held in Seoul on November 10, 2026. Physical AI refers to AI systems that move and perform tasks directly in the physical world, such as robots.

According to Hwang, optimizing AI models alone is insufficient for robots to autonomously judge and work in real environments; instead, field data collection, virtual environment learning, and actual robot execution must be integrated into a coherent development system. The company stated that XYZ collects data from 50 service sites operated by its subsidiaries and uses NVIDIA's robot simulation technology to train and validate robots in virtual environments. These learning results are then applied back to actual robot operations, creating a closed-loop structure where "field data and virtual learning mutually reinforce each other" to handle diverse working environments.

XYZ focuses on developing technology that enables robots not only to understand human requests but also to autonomously plan and execute long-sequence tasks. The company uses its proprietary robot intelligence framework "BrainX" (브레인엑스) to connect customer interaction, task planning, and physical manipulation, applying it to products like the semi-humanoid robot "DEUX" (듀스).

In the presentation, XYZ reported specific performance improvement data: the time taken for a robot to fold clothes was reduced from 50 seconds to 25 seconds, while maintaining a 96% success rate. The company claimed this case demonstrates improvements in both accuracy and the operational speed required for actual services.