ProductSouth Korea

RLWRLD releases DexBench: 15 humanoid hand tasks in 4 industries, confirms NVIDIA investment

Published 2 min read
Source
Data graphic: Tasks: 15 dexterous-manipulation…

RLWRLD, a physical AI start-up with offices in San Francisco, Seoul and Tokyo, said on Nov 9 it released DexBench, an open benchmark of 15 dexterous-manipulation tasks for five-finger humanoid hands across retail, logistics, assembly and hotel food and beverage, built on NVIDIA Isaac Lab-Arena and the Newton physics engine.

The company also confirmed an NVIDIA investment announced at AI Day Seoul 2026, without giving the amount. Code, environments, training data, baselines and a leaderboard are open.

RLWRLD, the robotics foundation-model company led by chief executive Ryu Jung-hee, said in a release on Nov 9 that it had published DexBench, a benchmark for humanoid dexterous manipulation in industrial settings. The company presented it at NVIDIA's AI Day Seoul 2026 at COEX the same day and confirmed an investment that NVIDIA announced at the event. RLWRLD did not disclose the amount, the round or other investors; it said the money goes to upgrading its foundation model and expanding deployments with industry partners.

What DexBench measures

DexBench contains 15 tasks in four industries: retail, logistics, assembly and hotel food and beverage. Examples in the release include picking and placing products, unloading goods from roll cages, inserting connectors and fastening screws, and stacking plates and pouring liquid. Each task is tested for generalisation along three axes: unseen objects, changed physics and new spatial layouts. The benchmark targets five-finger robot hands and runs on NVIDIA Isaac Lab-Arena and the open-source Newton physics engine. NdotLight, SPACE AI and PhysicsSimLab supplied 3D assets and environments. RLWRLD says it releases the code, evaluation environments, simulation assets, training data, baseline models and a leaderboard at dexbench.ai, with baselines measured on its own RLDX 1.1 model. The release lists Genesis AI and Mimic Robotics as early-access users.

The benchmark is the first output of a June initiative with NVIDIA to set shared evaluation and data standards for dexterous manipulation, the company said. RLWRLD has offices in San Francisco, Seoul and Tokyo and runs commercial pilots in Korea, Japan, the EU and the US.

MEKAWIRE analysis

Whoever writes the test for robot hands gets to define what good looks like for buyers.

Industrial customers today compare humanoid hands on vendor demos. A shared benchmark backed by NVIDIA's simulation stack could become the reference, and RLWRLD's model is the first baseline on it. The counter-argument is that simulation scores often fail to transfer: DexBench is simulation only, and real-world results on the same 15 tasks are what buyers will ask for. RLWRLD has no disclosed funding round in our database yet; the size of the NVIDIA investment is the missing number. Signals to watch: third-party entries on the leaderboard, a real-robot version of the tasks, and a disclosed round.

Good or bad news for the company?

 

Key facts
Tasks
15 dexterous-manipulation tasks
Release
Nov 9, 2026, AI Day Seoul 2026
Platform
NVIDIA Isaac Lab-Arena and Newton physics engine
Industries
4: retail, logistics, assembly, hotel F&B
Open assets
Code, environments, data, baselines, leaderboard
NVIDIA investment
Confirmed, amount not disclosed
Generalisation axes
3: object, physics, spatial
Sources
MEKAWIRE Intelligence
  1. 01

    Why it matters

    DexBench gives five-finger humanoid-hand developers 15 tasks across retail, logistics, assembly and hotel food and beverage, with open code, environments, training data, baselines and a leaderboard. Its object, physics and spatial generalisation axes could help buyers compare adaptability across conditions. The supplied material does not establish whether benchmark results predict performance in deployed operations.

  2. 02

    Rival analysis

    RLWRLD develops RLDX robot foundation models using in-house motion-capture data, while NVIDIA supplies the underlying simulation platform and is a confirmed investor. That relationship places NVIDIA in both infrastructure and financing roles; it does not establish RLWRLD's technical lead. No competing developers' scores or comparable deployment results are supplied, so a competitive ranking is unsupported.

  3. 03

    Valuation context

    NVIDIA's investment amount and terms are undisclosed, and no RLWRLD valuation, total funding or financing stage is supplied. NVIDIA is publicly listed, but its market capitalisation is not provided; market cap is distinct from a disclosed private financing valuation. The supplied facts support neither a valuation estimate nor an assessment of the investment's financial materiality.

  4. 04

    Related papers

    RLWRLD lists PointWAM: 3D World Action Modeling for Dexterous Robotic Manipulation and PaceVLA: Adaptive Temporal Action Quantization in Vision-Language-Action Models. NVIDIA lists ChunkTrust, Magic-W0 and AlphaNIST; no paper results or demonstrated links to DexBench performance are supplied.

  5. 05

    Supply-chain implications

    DexBench is built on NVIDIA Isaac Lab-Arena and the Newton physics engine, making that software stack a relevant dependency for users adopting its environments. RLWRLD also identifies in-house motion-capture data as an input to its RLDX models. Hand hardware suppliers, compute procurement, component costs and manufacturing arrangements are not disclosed.

  6. 06

    Signals to watch

    Watch for independent leaderboard submissions and results across all three generalisation axes. Evidence connecting benchmark scores to physical-hand reliability, throughput and intervention rates would clarify commercial relevance, but none is supplied. Disclosure of NVIDIA's investment amount and terms would allow a more grounded financing assessment.

  7. 07

    Analyst view

    DexBench could give RLWRLD influence over how dexterous manipulation is evaluated if outside developers adopt it. Open assets lower access barriers, but adoption, commercial traction and a durable model advantage are not established by this release. Investors should treat the benchmark launch and NVIDIA investment as developments to investigate, with technical and financial significance still unquantified.

  8. 08

    Questions you should be asking

    How closely do DexBench scores track physical-hand performance, and which independent developers have submitted results? What investment amount, ownership terms and commercial commitments accompany NVIDIA's backing? Which hand platforms are supported, and what compute and integration costs do users face?