How we actually test a robot hand

TL;DR

Short-form video from RLWRLD — the Korean physical-AI company behind the RealDex1 hand-agnostic foundation model — walking through how they actually stress-test the 10+ commercial dexterous robot hands they’ve procured and disassembled. Anchors the launch of All Hands Up! (allhandsup.org), an open interactive platform with URDF viewers, joint sliders, side-by-side spec sheets, and clinical thumb-opposition (“Kapandji test”) results across every hand in the archive, plus RLWRLD’s private DexBench simulation benchmark that surfaces failure modes on real-workplace skills that datasheets miss. The pitch: the field can’t answer “which robotic hands actually perform in the real world?” from spec sheets alone, and RLWRLD is publishing the empirical hardware-bottleneck findings from having every hand in one lab.

Key claims

All Hands Up! covers 10+ dexterous robotic hands with interactive URDF visualizations that let users actuate individual joints and verify grasp shapes in-browser, plus side-by-side specifications [project page].
The archive includes RLWRLD’s empirical findings on the hardware bottlenecks and design trade-offs that “gate true dexterity” — framed as what actually matters beyond datasheet numbers [project page].
The Kapandji test — 11 clinical thumb-opposition landmarks — is used as the standard cross-hand dexterity metric, measured in simulation by sweeping each hand’s thumb against the landmarks on its own URDF model [All Hands Up! Kapandji panel].
No perfect robotic hand exists yet — the framing of the entire platform is “now you have the data to make smarter decisions”, not “here is the winner” [tweet].
RLWRLD’s DexBench is a private complementary benchmark: many skills required by real workplaces run inside a simulator against the URDFs of many commercial hand architectures, and “we see many failures” — hardware options don’t account for actual actions [podcast interview].
RealDex1, RLWRLD’s foundation model, is deliberately hand-agnostic — a hardware-agnostic intelligence layer that works across a wide variety of robotic hands, fusing vision, tactile, force, torque, and motion signals [podcast interview].

Method

Short technical explainer video (transcript not available at filing time). The associated platform is a browser-native robot-hand archive: each entry loads a URDF into a Three.js viewer, exposes per-joint sliders, and offers pose presets including the 11-level Kapandji thumb-opposition test measured in-simulation against each hand’s own URDF. Specifications sit alongside for side-by-side comparison. The launch context is that RLWRLD physically procured most of the commercial robot hands on the market, disassembled them, and ran them through their internal DexBench simulator on workplace-relevant manipulation skills — the video presumably walks through parts of that testing methodology, and the site publishes the aggregate spec / URDF / Kapandji results.

Results

Coverage numbers only, from the launch materials: 10+ dexterous hands in the archive at launch; Kapandji results published per hand; DexBench findings summarized as “no perfect hand exists” with per-hand hardware-bottleneck notes. No quantitative per-task success rates or head-to-head rankings are reported publicly at filing time — the DexBench numbers themselves remain internal, with the public artifact being the URDF viewer + spec comparison + Kapandji sweep.

Why it’s interesting

Anchors a design-space position that Tactile sensing for manipulation has been circling: placement dominates sensor type (Tactile Genesis: Exploring Tactile Sensors at Scale for Learning Dexterous Tasks) and whole-hand coverage beats fingertip-only concentration — but those findings came from a simulator, whereas RLWRLD is running the actual commercial hands in one lab and publishing the cross-hand comparison the field has been missing. Complements the policy-side hand-agnostic bet in Solving Dexterity: A Full-Stack Approach (mimic hand M1 + wearable U1) (mimic hand M1’s “every actuator doubles as a force sensor” recipe) and Tactile-Reactive Dexterous Hand: High-Frequency Physical Interaction (variable-rate MoT fusing kHz touch with slower vision) by taking the hardware-audit position instead: don’t propose a new hand, publish empirical data on the ones that exist. Also relevant to Human-to-Robot Retargeting — the Kapandji sweep is a compact, embodiment-agnostic dexterity metric that any human-hand retargeting pipeline needs to match at the target embodiment, and having Kapandji scores for 10+ hands makes the “which hand can even reach this pose” question checkable before designing a retargeting recipe.