NVIDIA invests in RLWRLD as the Korean startup releases DexBench, a 15-task test for humanoid hands built on NVIDIA's simulator

NVIDIA said on Nov. 9 that it will invest in RLWRLD, a Seoul startup building robot foundation models, at a startup forum it co-hosted with South Korea's science ministry during AI Day Seoul. The money comes from NVIDIA itself rather than its NVentures arm, Seoul Economic Daily reports, and the same announcement named two Korean language-model developers, Upstage and Motif Technologies. Neither side gave an amount. The Elec reports that RLWRLD is raising about ₩200 billion at a pre-money valuation of roughly ₩1 trillion.

What RLWRLD showed on stage explains NVIDIA's interest better than the round does. DexBench, a benchmark for how well humanoid robots handle objects with five-fingered hands, is built on NVIDIA's Isaac Lab-Arena and the Newton physics engine. A stake in the company that writes the test helps make NVIDIA's simulator the place where robot hands get graded.

Fifteen tasks in four workplaces, scored on what the model has not seen

The version shown in Seoul has 15 tasks set in retail, logistics, assembly and hotel food service, with hands and products modeled on real specifications, according to the company's announcement. Besides success rates, it measures how well a model generalizes along three axes: unfamiliar objects, different physics and changed layouts. When RLWRLD and NVIDIA first described DexBench in June, they listed 18 "atomic" tasks across five skill domains.

RLWRLD says it will publish the code, environments, simulation assets, training data, baseline models and a leaderboard, without giving a date. Genesis AI and Mimic Robotics are testing it through early access. "Robot AI progress can only be tied to real deployments if we can objectively measure precise manipulation in real industrial settings," chief executive Ryu Jung-hee said.

The weak point is who goes first. The initial baseline comes from RLDX 1.1, RLWRLD's own next model, and in June the company said its RLDX-1 beat NVIDIA's GR00T N1.6 and Physical Intelligence's π0.5 across eight simulation benchmarks. A leaderboard whose first entry is its author's model reads as marketing until outside labs post their own scores.