EMG Hand Control — Prosthetic Hand Gesture Recognition
Real-time EMG-based hand gesture recognition for prosthetic arm control using CNN+LSTM with Quick Calibration. Works for any new user after ~2 minutes of calibration.
Performance
| Gesture | Accuracy (Self) | Accuracy (New User after Calibration) |
|---|---|---|
| rest | 100% | ~85% |
| fist | 100% | ~85% |
| grasp | 99% | ~80% |
| index | 97% | ~75% |
| middle | 100% | ~85% |
| ring | 100% | ~85% |
| pinky | 95% | ~75% |
| thumb | 100% | ~85% |
| wrist_rotate_out | 92% | ~80% |
| wrist_rotate_in | 90% | ~80% |
| Overall | 97.2% | ~85% |
Architecture
Input: 150 samples × 8 EMG channels (750ms @ 200Hz) ↓ CNN Block: Conv1D(64) → BatchNorm → ReLU Conv1D(128) → BatchNorm → ReLU → MaxPool → Dropout(0.3) Conv1D(256) → BatchNorm → ReLU → MaxPool → Dropout(0.3) ↓ Bidirectional LSTM (hidden=128, layers=2, dropout=0.3) ↓ FC: Dense(128) → ReLU → Dropout(0.4) → Dense(10) ↓ Output: 10 gesture classes
Quick Calibration System
New user calibration in ~2 minutes:
User wears Myo Armband Performs each gesture 3 times × 5 seconds System fine-tunes last layer on user data (30s) Ready — no full retraining needed
Gestures
| Label | Gesture |
|---|---|
| 0 | rest |
| 1 | fist |
| 2 | grasp (cylindrical grip) |
| 3 | index finger extension |
| 4 | middle finger extension |
| 5 | ring finger extension |
| 6 | pinky finger extension |
| 7 | thumb extension |
| 8 | wrist rotate out (pronation) |
| 9 | wrist rotate in (supination) |
Dataset
- 4 personal recording sessions
- 10 gestures × 10 rounds × 5 seconds each
- ~400,000 EMG samples
- Myo Armband — 8 channels — 200Hz
Training Details
| Parameter | Value |
|---|---|
| Window | 150 samples (750ms) |
| Step | 75 samples (50% overlap) |
| Optimizer | AdamW (lr=5e-4) |
| Loss | Weighted CrossEntropy + Label Smoothing |
| Epochs | 100 (best at epoch 78) |
| Batch size | 128 |
| Normalization | Global fixed (saved as .npy files) |
Key Technical Contributions
- Block-level train/test split — prevents data leakage from overlapping windows
- Fixed global normalization — ensures train/inference consistency (critical fix that improved accuracy from 2% to 94% on rest gesture)
- Quick calibration via last-layer fine-tuning — enables cross-user generalization without full retraining
- Hysteresis voting — stable real-time predictions for actuator control
Files
| File | Description |
|---|---|
models/best_model_hand.pt |
Trained PyTorch weights |
models/hand_norm_mean.npy |
Global normalization mean |
models/hand_norm_std.npy |
Global normalization std |
code/train_hand.py |
Training pipeline |
code/quick_calibration.py |
Calibration + real-time inference |
code/collect_hand_data.py |
Guided data collection |
Usage
# Quick Calibration for new user
python3 quick_calibration.py
# Records 2 min → fine-tunes → real-time inference
Hardware
- Sensor: Myo Armband (8 EMG channels, 200Hz, Bluetooth)
- Controller: Raspberry Pi (deployment)
- Prosthetic: 3D-printed hand with servo motors + tendon system
- Communication: Serial (115200 baud)
Related
- Drone Control Model: malansi/EMG-Gesture-Recognition
- Dataset: malansi/EMG-Gesture-Dataset
Demo
https://youtube.com/shorts/JOfaNj-Af6Q
Author
Mohammed Alansi AI & Biomechatronics Research — EMG-Based Prosthetic Arm Control