RLWRLD Partners with NVIDIA to Build 'DexBench' as Next-Generation Industry Standard for Humanoid AI

RLWRLD, a developer of advanced physical AI foundation models (CEO: Ryu Jun-hee / Japan Representative: Lee Hoon), announced on June 9, 2026, a collaboration with NVIDIA to set next-generation industry standards for humanoid robot AI. The partnership focuses on three key areas: 1) developing 'DexBench,' a shared benchmark for evaluating dexterous manipulation; 2) standardizing data used for learning dexterous movements; and 3) strengthening integration with NVIDIA's robotics development platforms, 'Isaac Lab' and 'Isaac Lab-Arena.'

'Dexterous Manipulation,' the ability of humanoid robots to perform delicate tasks like precision assembly, sorting, and packaging, is a critical technical area for the commercialization and industrialization of humanoid AI. Currently, however, the industry lacks common benchmarks to objectively evaluate performance or standardized data foundations for training related AI models. These gaps hinder technical progress and commercial expansion.

RLWRLD CEO Ryu Jun-hee commented, 'Without a common language to evaluate and reproduce precision robot hand movements, the commercial potential of humanoid AI remains limited. Through this collaboration with NVIDIA, RLWRLD is committed to developing industry-wide infrastructure by establishing the 'DexBench' benchmark and data standards, going beyond mere model development.'

Amit Goel, Head of Robotics Ecosystem at NVIDIA, stated, 'Measurable, reproducible dexterous manipulation capability is essential to scale industrial use of humanoid robots. By integrating DexBench with the NVIDIA Isaac platform, the robotics community can develop more reliable, high-precision manipulation models, accelerating the practical application of next-generation robotics.'

'DexBench' is integrated with NVIDIA’s 'Isaac Lab-Arena' to evaluate performance in both simulated and real-world environments. It assesses 18 major industrial tasks—such as assembly, sorting, and packaging—across five dimensions: 'Grasp Diversity,' 'Spatial Accuracy,' 'Temporal Accuracy,' 'Contact Accuracy,' and 'Situational Awareness.' This provides a common evaluation metric for manufacturers and research institutions.

Additionally, RLWRLD’s foundation model, 'RLDX-1,' has demonstrated state-of-the-art performance across eight established simulation benchmarks, outperforming cutting-edge models such as NVIDIA GR00T N1.6 and Physical Intelligence π₀.₅.