Guides

Every task guide, grouped by what you are trying to do. If you are new, start with EMG classification: it runs in one command and establishes the App + Stream + Pipeline pattern everything else assumes.

Start here

Complete runnable experiments, walked through line by line against a real file in examples/.

• EMG classification - two-class Rest/Fist with CatBoost on RMS+MAV features. The best first read.
• EMG regression with VHI - continuous 5-DOF control, recorded through the Virtual Hand and pushed back to it.
• Examples directory - one paragraph on each runnable example, with its command and what is worth changing.

Connecting hardware

Getting samples out of an amplifier and into a Stream.

• Pick a device from the UI - DevicePicker: choose the amplifier from a dropdown at runtime instead of hardcoding it. Start here if you are new to the hardware.
• Connect OTB devices - the OT Bioelettronica sources in detail: wiring, modes, channel geometry, dropout counters.

Driving a device

Anything that moves is a target: three methods, and a TOML control map naming which model output drives which control. Concepts › Controls explains the system itself.

• Drive your own device - your prosthesis, motors or cursor, from Python. Start here: the copyable example runs without hardware.
• Drive a remote target - the contract to serve if your device is already its own program.
• Build a remote target, stage by stage - that same contract, built in seven stages with a checkpoint at each.
• Integrate the Virtual Hand - the one remote target this project ships.
• Install the Virtual Hand - the installer CLI, and where a build is looked for.

Recording

• Enable on-disk recording - the four lines that start writing sessions.
• Record and replay - capture sessions, read them back programmatically.
• Record good training data - cycle-style recording, how many cycles you actually need, verifying templates before training.
• Record for proportional control - the graded counterpart: why three cued classes make a three-class model whatever you fit, and the followed-cursor block that fixes it.
• Track a force target - isometric force tracking against a trapezoid, on an amplifier AUX channel, with the target recorded as its own stream.

Models and features

• Add a custom model - wire extract / train / predict for any ML library.
• Record for proportional control - if your regressor is dead below 30 % effort or goes the wrong way when the subject contracts harder, the fix is in the recording protocol.
• Feature extraction cookbook - copy-paste @pipeline.extract snippets (RMS+MAV, bandpass+envelope, spectral, sliding RMS, onset detection, multi-stream fusion).
• Use the recipe feature set - the shipped feature recipes.
• Keep state between pipeline stages - rolling windows, stateful models, gating side effects on change.
• Post-process predictions - PostProcessor and myogestic.outputs.filters for output smoothing.

Extending the framework

• Add a custom source - implement the Source protocol for a new device, file format or transport.
• Add a custom widget - a class with a .ui(ctx) method that draws ImGui commands.
• Publish a data stream - an Outlet is a paced sender for telemetry: predictions to a recorder, a stream another application reads. If something moves, you want a target, not this.

Operations

• Run headless (no GUI) - unattended recording and prediction; signal handling; protocol-driven scripts.