Robot hand plays rock-paper-scissors

How it works

• A RealSense feeds MediaPipe hand landmarks. A finger-extension classifier turns them into one of three gestures, in real time.
• The gesture maps to a target pose, sent through Tesollo's SDK over Modbus-TCP.
• Every pose is twenty joint angles: five fingers, four motors each.

Why classify by extension

Counting extended fingers off the MediaPipe skeleton needs no training data, and when it fails you can see which fingertip landmark drifted. Rock is none extended, scissors two, paper five.