A personal creative project

EMG-Gesture-Interface

Eight channels of muscle signal in, four hand gestures out — a wearable interface that reads the electrical activity of the forearm and turns it into software input.

How it was built Watch the demo

The recorded talk

From the idea to the classifier

At the whiteboard: why the system exists, how the gesture data was recorded, and what the classifier does with those sixty-four numbers.

From an idea to a working classifier

Seven steps, in the order they happened.

Turning muscle into input

Keyboards and mice cannot read intent. The goal of this project was to make the forearm itself an input device.

Choosing the tooling

An off-the-shelf MYO armband for the sensors, and nukleos — an open-source app — for recording. The software built on top of them is my own work.

Recording the gestures

Four gestures: rock, scissors, paper, and the OK sign. Each was held and recorded six times for twenty seconds — 120 seconds per class.

Cutting the signal into windows

At 200 Hz, eight consecutive readings of all eight sensors take 40 ms. That window becomes one row of 64 numbers.

Training the classifier

Sixty-four numbers in, one class out. The classifier never sees the raw stream — only the window it is handed.

Holding the variables still

Electrode placement changes the signal. Every recording therefore came from the same right forearm, within one short timespan, so the classes stayed comparable.

The live demo

The recorded talk walks through the whole system end to end: the idea, the recording setup, and what the classifier does with those sixty-four numbers.

01 / 07

How the system works

Eight electrodes sit on the forearm. Each reading captures all eight at once, eight times over, and those sixty-four numbers are the only thing the classifier ever sees — from them it decides which of four gestures you are holding.

Key parameters

Hardware

The MYO armband carries eight sensors against the skin of the forearm. Each one measures the electrical activity produced by the muscles beneath it.

Data

Every dataset line is eight consecutive readings of all eight sensors — 64 columns of EMG — plus one final column holding the gesture that was being held.

Classifier

Given those 64 numbers the classifier predicts one of four classes: rock, scissors, paper, or the OK sign.

Dataset

Each gesture was recorded six times for twenty seconds — 120 seconds per class — all from the same right forearm in a short timespan, then concatenated into one CSV per class.

Watch the full walkthrough

The recorded talk goes through the idea, the recording setup, and what the classifier does with those sixty-four numbers.

Play the video

Hardware and tooling

MYO armband · nukleos

The armband and the recording app are third-party tools. The pipeline, the feature windows, and the classifier on top of them are my own work.