RLWRLD and Nvidia Create Universal Standard for Evaluating Humanoid Robot Dexterity

South Korean startup RLWRLD has partnered with Nvidia to develop DexBench — the first universal benchmark for evaluating how precisely robots control their hands when performing physical tasks.

The Problem DexBench Solves

The market for humanoid robots is growing rapidly today, but without unified standards. Each manufacturer tests manipulators differently: some measure grip speed, others measure positioning accuracy, still others measure applied force or resistance to failures. Comparing results between different companies is practically impossible.

This is a problem not only for researchers, but first and foremost for industrial buyers. A factory or warehouse choosing a humanoid robot cannot objectively assess which system will better handle a specific task. The solution is a single reference test that all market players pass under the same rules.

DexBench is being developed precisely as such a standard. Test scenarios will include tasks involving grasping heterogeneous objects, fine assembly, moving fragile items, and situational responses to unexpected obstacles — everything needed in manufacturing and logistics.

The Roles of RLWRLD and Nvidia

RLWRLD is a South Korean startup founded by Chongi Ryu. It positions itself not as a manufacturer of another robot, but as a developer of the infrastructure layer for the entire industry: standards, benchmarks, and evaluation methodology without which the industry cannot develop normally. Nvidia joined the partnership as a technological anchor with enormous influence in the robotics hardware stack. The company is already building an ecosystem through its Isaac platform — a set of tools for simulation, reinforcement learning, and deployment of robotic models. DexBench fits organically into this strategy: Nvidia is interested in industry evaluation standards being formed on its foundation. Specific partnership tasks:

• Develop unified metrics for evaluating the dexterity and precision of robot hand movements
• Create standardized test scenarios for industrial and warehouse applications
• Open access to the benchmark for a wide range of manufacturers
• Integrate DexBench into Nvidia Isaac's training infrastructure
• Establish results as an industry standard for the next generation

A Race Without Rules

The market for humanoid robots is experiencing a boom: Figure AI, Apptronik, Agility Robotics, Boston Dynamics, and dozens of Asian startups are racing to launch new models. Investors have poured billions of dollars into the sector, and major manufacturing companies are beginning their first pilots.

"We intend to develop next-generation standards for humanoid robotics," —

Chongi Ryu, founder and CEO of RLWRLD

But it is precisely now, while standards have not yet been established, that key players have a chance to set the rules of the game for years to come. If DexBench becomes the de facto standard with Nvidia's support, manufacturers will be forced to adopt it, and buyers will finally get an objective frame of reference for comparison. The market will be able to consolidate around measurable metrics, not marketing claims. The analogy with ImageNet is telling: when this computer vision benchmark appeared in 2010, it accelerated industry progress by an entire decade. The creators of DexBench are counting on a similar effect for physical AI.

What This Means

For the industry, DexBench is an attempt to do with robotics what ImageNet did with computer vision: provide a common reference point and launch competition for measurable results. If RLWRLD and Nvidia's initiative takes hold, standards for humanoid robots could be formed much faster than expected — and this will reshape the balance of power in the market.