XYZ Shows Physical AI Built on Field Data at NVIDIA AI Day Seoul

A field-to-simulation flywheel

The company combines NVIDIA's AI computing platform with virtual robot simulation infrastructure. Behavior data gathered at 50 service sites run by a subsidiary is fed into a digital-twin environment for training and repeated validation. The results are then applied back to physical manipulation tasks on real robots. XYZ calls this loop of field collection, virtual learning and physical execution an intelligence flywheel, and says it raises how ready the robot intelligence is for commercial use in unstructured work environments.

The core software is BrainX, XYZ's in-house robot intelligence framework. According to the company, it lets robots go beyond passively recognizing natural-language commands: they assess the situation on site and plan and carry out long-horizon tasks. It combines user interaction, task path planning and precise physical manipulation in one system, and XYZ is rolling it out step by step on hardware including its semi-humanoid robot DEUX.

Laundry-folding results

XYZ also released results from a study that assumed a real commercial service setting. Using robot manipulation, it cut the time needed to fold clothes, a well-known hard case for handling deformable objects, from 50 seconds to 25 seconds, a 50% reduction, while reaching a 96% success rate. The company says this shows working speed and economics suited to immediate deployment in industry.

Hwang said that getting robots out of the lab requires an integrated development system in which field data collection, high-precision simulation and real-device control fit together seamlessly, beyond simple model tuning. He said XYZ will keep working with NVIDIA's global ecosystem to improve robot learning efficiency and speed up global commercialization of its physical AI solutions.