All Hands Up!
RLWRLD
A browser-based robot hand archive for exploring joint motion, hand geometry and simulated thumb reach across multiple hand designs.
What you can explore
Inspect a model before shortlisting hardware: compare presets, individual joints and the assumptions behind its thumb-reach visualization.
Before you use it
A research viewer, not a hardware driver or an independent product certification. Gallery illustrations are AI-generated; model revisions and physical behavior need separate verification.
Original contributors
Max Chang Hwan Kim and Mark Seungjae Lee
License / access: Website reuse terms not verified. Dexhands uses separately sourced model packages; individual MIT, Apache-2.0 or BSD licenses are shown in the native viewer.
Explore licensed 3D models on Dexhands
Explore the hand. Right here.
Rotate, zoom and adjust individual model joints. Downloads begin when you choose Load 3D model. This is geometry, not a physical performance test.
Open larger view ↗ · 27 licensed models available
All Hands Up!: inspect a hand before choosing hardware
- Environment
- Browser with 3D/WebGL support; no robot required
- Embodiment
- Publisher-selected hand models; confirm exact generation, size and side
What you need
Start with the native Dexhands viewer above: 27 licensed models from original publisher packages, with rotation, zoom and joint controls. Choose Load 3D model to download only the selected model. Each has a license and source revision. The separate All Hands Up gallery has 21 entries in the reviewed snapshot and additional tools; its code, generated images and motion presets are not copied here.
First useful check
Choose a task contact, such as thumb-to-index opposition. Use the same question for two hand models and record the exact model name before comparing their poses.
A small first experiment
Match the model before moving a joint
Select the manufacturer and full model name. Record generation, side and sensor configuration if available; mark missing fields as unknown. Flex 2 and Prima 1 have separate entries, as do LEAP v1 and the ORCA generations. A family resemblance does not establish the same joint ordering or actuator layout.
Choose a model in the original gallery →Compare presets, then isolate one motion
Try Open, Grasp and Spread, then adjust one visible joint slider at a time. Watch where the fingertip moves and whether other joints follow. Use Prima 1 as one example, then repeat with a different design. Treat the controls as model coordinates; do not copy their values into real hardware.
Inspect the Prima 1 model →Read thumb reach as a geometric question
Inspect the Kapandji visualization and its K0–K10 landmarks. The site counts reached targets out of 11 in its modeled sweep. Ask which contact location matters to your task, not merely which total is larger. A reached landmark does not establish usable force, collision-free manipulation or access while holding an object.
Inspect the Flex 2 model →Turn the visual comparison into a test brief
Write down two candidate models, the required contact, the relevant joint motions and anything the viewer cannot tell you. Confirm geometry and actuation with the manufacturer. Next request a matched physical task with the actual object, sensing configuration, software revision and all attempts recorded. Keep the viewer observation separate from any later test result.
Build a fair hardware comparison →
What to record
- Save the exact viewer model name, page link and observation date; record unresolved side, size or revision differences.
- Describe the task contact and the motion you observed. Separate model coordinates, mechanical joints and independently controlled axes.
- Keep a short evidence sheet: viewer observation / manufacturer confirmation / physical test still required.
- Inspect model-specific license notices before downloading or reusing assets. A robot-description license does not automatically cover the website, article text or images.
Common problems and what to check
- A gallery image looks different from the product
- The publisher identifies its gallery illustrations as AI-generated from public information. Use manufacturer photographs for product identification and inspect the actual model separately.
- A larger thumb-reach count seems to mean a better hand
- The result depends on model geometry, joint limits and the sweep method. It does not compare sensing, force, reliability or task completion.
- The viewer specification differs from a catalog record
- Keep the original sources and definitions visible. Total DOF, actuated axes, palm-only mass and complete installed mass are not interchangeable.
- The 3D model or controls do not finish loading
- Try a browser that supports WebGL and reload the selected model. A failed load is not evidence of missing joints or poor hand performance.
- A linked model is assumed to fit a physical unit
- Family links and beta/generation labels need revision checks. The external viewer is not an SDK or a validated controller.
Where it stops
The viewer illustrates the loaded model. It does not measure physical force, durability, tactile quality, control latency or success on an object. Its Kapandji display counts reached landmarks in a modeled thumb sweep; it is not a universal dexterity ranking or a clinical assessment.
Keep exploring
MuJoCo Menagerie
A collection of robot models for MuJoCo, including LEAP, Allegro, Shadow Hand and DEX-EE models.
Dex Retargeting
Software for mapping human hand motion to the joint configurations of robotic hands.
Editorial reference · checked 2026-10-25