Shadow Hand

What the Shadow Hand can learn.

Each skill pairs neural pathways with target poses and a test the robot must pass with stimulation and fail without it.

1. 01Close fist1 pathway · 6 s testThe Shadow Hand closes into a fist only while the “fist” pathway is ON.
2. 02Pinch1 pathway · 6 s testThumb and index fingertips meet in a pinch only while the “pinch” pathway is ON.
3. 03Count fingers5 pathways · 11 s testFive pathways, one per finger. Each finger folds only while its own pathway is ON.

What the neurons can move.

20 controls in all. A skill sets their targets; the simulator does the rest.

Wrist

tilt · flex

Thumb

rotate · flex · hub · bend · tip

Index finger

spread · flex · curl

Middle finger

spread · flex · curl

Ring finger

spread · flex · curl

Little finger

cup · spread · flex · curl

Real physics, on our servers.

The Shadow Hand is simulated with MuJoCo 3.12, the official WebAssembly build, on the server; your browser loads the same model only to draw it. The model comes unmodified from Google DeepMind’s MuJoCo Menagerie.

Engine

MuJoCo 3.12, WebAssembly, server-side

Model

MuJoCo Menagerie · shadow_hand

In the simulator

Right Shadow Dexterous Hand on a fixed forearm, palm up. 20 position actuators; each finger’s “curl” drives its two outer joints together through a tendon (0 to 3.14 rad).