Seoul-based physical AI startup RLWRLD and NVIDIA on June 9, 2026 announced DexBench, a universal benchmark and data standard for evaluating humanoid robot dexterity. The initiative defines 18 atomic tasks across five manipulation regimes and plugs directly into NVIDIA Isaac Lab and Isaac Lab-Arena, giving researchers a shared yardstick for skills that have so far been measured ad hoc.
Five Pillars Of Dexterity
DexBench, published at dexbench.org, scores robots on grasp diversity, spatial precision, temporal precision, contact precision and context awareness. Each of the 18 atomic tasks targets a real-world manipulation primitive — from opening cabinet doors to pouring liquids — that humanoid foundation models are expected to handle as a baseline. RLWRLD CEO Junghee Ryu framed the benchmark as the missing layer between simulation accuracy and field deployment.
Integration With NVIDIA Isaac
NVIDIA will integrate DexBench into Isaac Lab-Arena so that the same evaluation runs across simulation and real-world hardware. The companies also commit to a shared data standard for dexterous manipulation training, which RLWRLD says will reduce the friction of swapping datasets between humanoid platforms. The work builds on the company's RLDX-1 foundation model, which already outperformed NVIDIA's GR00T N1.6 and Physical Intelligence π0.5 on eight established simulation benchmarks earlier in 2026.
Why It Matters
Humanoid investment hit a record $23 billion in 2026, but vendors still describe dexterity progress in incompatible terms. By tying DexBench to Isaac Lab, RLWRLD and NVIDIA give buyers — from automotive OEMs to logistics operators — a way to compare humanoid hands across vendors. The launch follows the recent Isaac GR00T Reference Humanoid with Unitree H2 Plus and Sharpa hands, and the Cadence-NVIDIA simulation pact that aims to close the sim-to-real gap. RLWRLD also confirmed growing demand for robot foundation models.
Reporting based on coverage from PR Newswire, Robotics & Automation News and NVIDIA.