RealWorld Advances Robot Hand Manipulation Learning Using NVIDIA Platform

Physical AI company RealWorld (RLWRLD) announced on the 6th that it is enhancing the object manipulation learning of its high-dexterity 5-finger robotic hand using the NVIDIA platform. The high-dexterity 5-finger robotic hand has a complex control system due to the numerous finger joints. The process of recognizing the surroundings with a camera, understanding the situation, and then precisely adjusting the fingers to move without mistakes must occur seamlessly. RealWorld is benchmarking alongside developing its own foundation model based on its proprietary data pipeline and an improved version of NVIDIA's GR00T N1.5 tailored for the 5-finger robotic hand. CTO Bae Jae-kyung stated, "We have conducted additional training optimized for the high-dexterity 5-finger robotic hand based on GR00T, which significantly increased the speed of repetitive experiments through quick fine-tuning," and added, "With iterative training, the previously insufficient data is reinforced, and as learning is optimized, the robotic hand is advanced to a level where it can be utilized in real-world applications." To alleviate the issue of data scarcity, RealWorld is conducting real robot experiments alongside virtual environment-based development tools. They utilize NVIDIA Isaac Lab during the construction of the robotic learning pipeline and are also in the process of testing their self-developed model on NVIDIA's Jetson Thor. This information was officially disclosed via NVIDIA at CES 2026. RealWorld plans to unveil its independent VLA (Vision-Language-Action) model in the first half of this year. Designed to have large-scale parameters, it aims to enhance language and vision understanding performance while increasing the precision and flexibility of the robotic hand's manipulation.