Physical AI startup RealWorld (RLWRLD) is joining forces with NVIDIA to develop performance benchmarks and data standards for humanoid hands.
This collaboration aims to establish a universal language for robot manufacturers and research institutions worldwide.
Humanoid Robot Competition Enters a Standardization Phase
Industry sources revealed on Tuesday that RealWorld has officially launched the DexBench Initiative in partnership with NVIDIA.
The DexBench Initiative centers on three key objectives: creating a dexterity performance evaluation benchmark, standardizing data formats for 5-finger humanoid manipulation, and integrating with NVIDIA’s Isaac Lab and Isaac Lab-Arena.
The humanoid robotics field is currently transitioning from a focus on algorithms and hardware to a phase of standardization.
China recently unveiled its Humanoid Robot and Embedded Artificial Intelligence (AI) Standards System (2026 edition), outlining a national strategy that encompasses industry standards, safety protocols, and ethical considerations. Meanwhile, tech hubs in the United States and Europe are ramping up efforts to develop comprehensive testbeds and guidelines.
RealWorld’s DexBench is built around five domains and 18 tasks that simulate real-world industrial environments.
The benchmark draws inspiration from actual industrial tasks, including assembly, sorting, packaging, and picking operations.
It defines five core domains: Grasp Diversity, Spatial Precision, Temporal Precision, Contact Precision, and Context Awareness, which are further broken down into 18 specific tasks.
A standout feature is the integration with NVIDIA’s Isaac Lab-Arena environment, enabling consistent validation across both simulated and real-world scenarios. This approach addresses the challenge of comparing robotic models across varied settings, offering a standardized method for assessing dexterity under uniform conditions.
On the data front, both companies are collaborating to design a universal format for 5-finger dexterity humanoids, ensuring seamless compatibility with the Isaac Lab pipeline.
Ryu Joong-hee, Chief Executive Officer (CEO) of RealWorld, emphasized that to advance the commercialization of dexterity AI, it needs a common language for measuring and replicating precise robotic hand movements. The partnership with NVIDIA to establish DexBench and data standards will set a new benchmark in the global humanoid AI ecosystem.
NVIDIA sees the DexBench Initiative as a catalyst for expanding its Isaac ecosystem.
Amit Goel, head of NVIDIA’s Robotics Ecosystem, stated that quantifiable and replicable precision in robotic manipulation is crucial for wider adoption of robotics in industrial settings. The integration of DexBench with the Isaac platform will provide the robotics community with standardized metrics and data infrastructure, accelerating the development of reliable, high-precision manipulation technologies.
