RealWorld unveils RLDX-1 humanoid model, claims benchmark wins over NVIDIA GR00T N1.6

RealWorld publicly unveiled RLDX-1, its humanoid precision manipulation foundation model, at a 500-person invitation event called Dexterity Night in Seoul on June 10 at Raum Art Center in Gangnam.

The company says RLDX-1 outperformed NVIDIA's GR00T N1.6 and π₀.₅ across eight open benchmarks: RoboCasa Kitchen, RoboCasa GR-1 Tabletop, and LIBERO-Plus among them. Built on a MAST (Multi-Stream Action Transformer) architecture, the model contains three core modules—Motion, Physics, and Memory—and processes vision, language, applied force, and tactile sensation in a single system. It runs across multiple robot platforms via fine-tuning rather than locking into specific hardware.

Two days earlier, on June 8, RealWorld announced a collaborative project with NVIDIA. The partnership covers DexBench (a dexterity performance evaluation benchmark), 5-finger humanoid data standards, and integration with the NVIDIA Isaac platform. DexBench spans five evaluation areas: grip variety, spatial precision, temporal precision, contact precision, and situational awareness. It includes 18 Key Atomic Tasks designed around actual industrial operations. The plan is to integrate DexBench with the NVIDIA Isaac Lab-Arena environment for verification in both simulation and real-world settings.

"Measurable and reproducible precision manipulation capabilities are essential for creating robots usable in industrial settings," said Amit Goel, Head of NVIDIA's Robotics Ecosystem.

RealWorld CEO Ryu Jung-hee framed the market opportunity around automation gaps. Even South Korea, which he described as the world's most automated country, sits at a 75% factory automation rate. That leaves 25% still dependent on human labor. RealWorld's goal from the start has been to automate all remaining labor markets globally.

Attendees included working-level representatives from Lotte Hotel, CJ Logistics, LG Innotek, AWS, and NVIDIA. Kim Jae-hwan of Hotel Lotte called hotels "the most suitable yet challenging places where physical AI should be applied," citing the need for humanoids to raise customer satisfaction without changing existing infrastructure. Gu Seong-yong of CJ Logistics pointed to overseas logistics center exports as a use case, noting that local labor in foreign markets cannot match domestic levels.

RealWorld operates three partnership tiers. Platform partnerships include NVIDIA on DexBench standards and AWS on the full data pipeline from collection through deployment. Domain partnerships connect with industrial customers who hold training data. Tech partnerships bring in hardware and sensor companies for industrial proof-of-concept validation.

The company secured an additional 39 billion won in a Seed 2 round earlier this year, bringing cumulative investment to 60 billion won. A Series A round is next. Development of RLDX-2 has already begun.

RealWorld outlined a regional strategy: build an East Asian physical AI ecosystem centered on South Korea, Taiwan, and Japan, then expand it to global standards through comprehensive collaboration with US big tech including NVIDIA, AWS, and Microsoft. Dexterity Night events have previously been held in San Francisco and Japan; Seoul was the latest stop.