Nvidia to Back Korean Physical-AI Startup RLWRLD as DexBench Lands

The planned investment, announced at Nvidia AI Day Seoul, funds RLWRLD's robot foundation model and a new benchmark for robotic hands.

Nvidia plans to take an equity stake in RLWRLD, a Korean startup building robot foundation models, the company confirmed on November 9 at Nvidia AI Day Seoul 2026. The size of the investment was not disclosed, and the deal has not yet closed — what Nvidia announced is a plan, not a completed transaction.

RLWRLD, also known as RealWorld, develops the robotics foundation model RLDX-1. The funding will go toward upgrading the model and expanding deployments at industrial sites.

The same event produced the concrete deliverable: DexBench, a benchmark for measuring how precisely humanoid robots handle objects. Built with Nvidia on the Isaac Lab-Arena simulation platform and the Newton physics engine, it runs 15 tasks — product placement, unloading goods, connector insertion, screw fastening — across four settings: retail and logistics, assembly, and hotel food and beverage.

DexBench scores more than task success. It also tests generalization: whether a robot can handle objects and settings it has not trained on. RLWRLD plans to publish the code, evaluation environments, training data and a performance leaderboard.

Chief executive Ryu Jung-hee put it plainly: progress in robot AI only connects to real adoption if precision manipulation can be measured objectively in industrial settings. The investment is part of a wider Korea push: Nvidia also announced planned stakes in language-model developers Upstage and Motif Technologies at the same forum, after its venture arm backed Point2 Technologies in April.