Exohand Updated

ExoHand

EMG-controlled assistive robotic hand. Built for ECTE351 by Team Orvyn (Group 19).

Overview

ExoHand reads surface EMG (sEMG) from the forearm, classifies the muscle activity into a hand gesture, and drives four servo-actuated fingers to match. The current system recognizes rest, fist, and individual-finger gestures (pinky&ring, middle, index, thumb), running the trained classifier directly on the microcontroller. VR integration is planned; not yet in this codebase. — see docs/architecture.md. Runs the trained classifier on-device (no PC needed for inference).

EMG-control concept

1. Sample raw EMG at 1 kHz from a single sensor channel.
2. Extract statistical features (MAV, RMS, variance, std dev, waveform length, zero-crossing rate, slope sign changes, skewness, kurtosis, IQR) over a rolling 30-sample window.
3. Classify the window with a Random Forest trained on EMG collected by the team (plus validation against the public NinaPro benchmark dataset).
4. Debounce the result (4 agreeing windows + 150ms hold) before committing to a new gesture.
5. Drive 4 servos through a PCA9685 to match the confirmed gesture, moving smoothly rather than snapping.

Full signal-flow diagram: docs/architecture.md.

Hardware stack

• Teensy 4.0 — main MCU for EMG sampling, feature extraction, and on-device RF inference
• PCA9685 servo driver (I2C, 4 channels)
• MG996R servos (or equivalent)
• SEN0240 EMG sensor
• ESP32-C3, for live telemetry/graph relay (firmware/current/esp32_bridge.ino — self-hosted WiFi AP + WebSocket server)

Repo layout

ORVYN/
├── firmware/        Teensy/ESP32 .ino code — current/ is what's in active use, archive/ is superseded versions
├── web/             static site + the FastAPI live dashboard
├── hardware/        KiCad schematics/PCB, BOM, perfboard build docs
├── signal-processing/  EMG filtering, feature extraction, trained model, datasets
├── docs/            build-log, architecture, decisions, posters
├── media/           screenshots, recordings, concept art
└── _quarantine/     duplicate/superseded files pulled out of the tree (gitignored, see MANIFEST.txt)

Nothing failed or superseded was deleted — old firmware versions, every dated PCB backup, duplicate dataset exports, and earlier notebook runs are all kept under the relevant archive/ folders. See docs/build-log.md for the full story of what was tried and what changed.

How to build / run

Firmware: open firmware/current/teensy_final.ino in Arduino IDE with Teensyduino (board: Teensy 4.0), plus the Adafruit PWM Servo Driver and Watchdog_t4 libraries. It includes ai_classifier_model.h, which is checked into firmware/current/ and compiles as-is. To regenerate it from a retrained model, run python3 signal-processing/export_model_to_header.py (loads ExoHand_RF_Model.joblib + ExoHand_Scaler.joblib, writes the header back to firmware/current/ai_classifier_model.h).

Signal processing / training: pip install -r signal-processing/requirements.txt, then signal-processing/ExoHand_EMG_Classifier_RF_V2.ipynb to retrain, or signal-processing/live_predict.py to run live PC-side inference over serial.

Live dashboard: web/dashboard/app.py (FastAPI). Install web/dashboard/requirements.txt, run with uvicorn app:app, open the browser page it serves. Includes a /api/demo mode that fakes EMG data if no hardware is connected.

Website: web/website.html is the current main site (Live Rehab dashboard view, live EMG chart, camera-based sensor-placement assistant). web/archive/ORVYNUpdated.html is an earlier, less developed version.

Hardware: schematics and PCB in hardware/schematics/, BOM in hardware/bom/, perfboard wiring/assembly in hardware/perfboard/.

Credit

Team Orvyn (Group 19). My role: led firmware (Teensy), EMG signal processing, hardware build. Data collection was team effort. Benchmark data from NinaPro.