Bypass EMG Control

Real-time EMG-driven control of a bypass prosthetic hand using Dynamixel servo actuators, a Myo armband, and LibEMG. A deep learning classifier decodes surface EMG signals into hand gestures and translates them into position and velocity commands sent to two Dynamixel motors over a 60 Hz control loop.

---

Overview

A bypass prosthesis is worn over an intact limb and mechanically couples to it, allowing the user's residual EMG activity to drive an external robotic hand. This system closes the loop between EMG acquisition, real-time gesture classification, and actuator control without any per-session calibration step — the classifier is loaded from a pretrained checkpoint and operates immediately on new users.

The pipeline is:

Myo armband (8-ch, 200 Hz)
    -> LibEMG online streamer
    -> OnlineEMGClassifier
    -> UDP socket (probabilities + velocity)
    -> input_thread (gesture -> motor command mapping)
    -> Bypass control loop (60 Hz, Dynamixel SDK sync write)

---

Hardware Requirements

• Myo Armband — 8-channel sEMG, 200 Hz, signed 8-bit ADC
• Dynamixel servos — two motors, Protocol 1.0
  • WRIST_ID = 0 — wrist pronation/supination
  • GRIP_ID = 1 — grip open/close
• Serial adapter — connected at COM6, 3 Mbaud (update DEVICENAME in bypass.py for your port)
• CUDA-capable GPU — DEVICE defaults to 'cuda' in utils.py

---

Software Dependencies

torch
libemg
dynamixel_sdk
numpy
scikit-learn
pandas
matplotlib
joblib
tqdm

Install via pip:

pip install torch libemg dynamixel-sdk numpy scikit-learn pandas matplotlib joblib tqdm

---

Repository Structure

Bypass_EMG_Control/
    main.py       -- Entry point: model loading, LibEMG setup, UDP listener, logging
    bypass.py     -- Bypass class: Dynamixel initialization and 60 Hz control loop
    models.py     -- Model definitions and loss functions
    utils.py      -- Hyperparameters, data loaders, training and evaluation routines

---

Models

Three architectures are defined in models.py. All scale raw EMG by / 128.0 internally to account for the Myo's 8-bit ADC range.

CNN (primary model) — multi-horizon dilated convolutional encoder. Three parallel Conv1d branches with dilations 1, 2, and 4 capture features at different temporal scales, concatenated and passed through a final conv layer, adaptive average pooling, and a 128-dimensional embedding head. The classifier head produces 5-class logits.

CNN_GRL — CNN with an attached gradient reversal layer (GRL) and a secondary user-identity classifier head. Intended for domain-adversarial training to reduce user-specific features in the embedding space.

MLP — simple fully connected baseline operating on handcrafted LibEMG features (default: WENG).

Pretrained weights are loaded from the pickles/ directory at startup. The active model can be switched at runtime by updating SharedContext.active_model_name.

---

Gesture-to-Command Mapping

The classifier outputs one of five classes. input_thread in main.py maps each predicted gesture to a (wrist, grip) position index:

Gesture ID	Gesture	Wrist command	Grip command
0	Rest (NM)	hold	hold
1	Hand close	pronation	hold
2	Flexion	hold	open (or close if flipped)
3	Extension	hold	close (or open if flipped)
4	Hand open	supination	hold

Each DOF has three discrete positions (min, mid, max). Velocity is proportional to the EMG signal strength computed by LibEMG's velocity estimator and is clipped to [0, 1] before scaling to the motor's maximum velocity.

The flip_lr flag in SharedContext swaps the flexion/extension grip direction to accommodate left/right arm differences.

---

Configuration

Key parameters are defined at the top of their respective files:

bypass.py

Parameter	Default	Description
DEVICENAME	'COM6'	Serial port for Dynamixel USB adapter
BAUDRATE	3000000	Communication baudrate
RATE	60	Control loop frequency (Hz)
WRIST_MIN_POS / WRIST_MAX_POS	800 / 2500	Wrist position limits (encoder counts)
GRIP_MIN_POS / GRIP_MAX_POS	1780 / 2340	Grip position limits (encoder counts)
WRIST_MAX_TORQUE / GRIP_MAX_TORQUE	500 / 400	Torque limits
WRIST_MAX_VEL / GRIP_MAX_VEL	300 / 200	Maximum motor velocities

utils.py

Parameter	Default	Description
SEQ	40	EMG window length (samples)
INC	2	Window increment (samples)
CH	8	Number of EMG channels
CLASSES	5	Number of gesture classes
SAMPLING_RATE	200	Myo sampling rate (Hz)
DEVICE	'cuda'	PyTorch device

---

Running

1. Connect the Myo armband and ensure the Myo driver/daemon is running.
2. Connect the Dynamixel USB adapter and confirm the correct COM port in bypass.py.
3. Place pretrained model weights (.pt files) in the pickles/ directory. The filename must match the model name listed in model_names in main.py.
4. Run:

python main.py

On startup the system will:

• Load all models listed in model_names
• Initialize the LibEMG Myo streamer and online classifier
• Open a UDP socket on port 12346 to receive prediction packets
• Start the input thread for gesture-to-command translation
• Initialize Dynamixel motors (enable torque, set limits)
• Start the 60 Hz control loop
• Log raw EMG to emg_logs/<NAME>/ and motor state to bypass_log/<NAME>/

To stop, trigger an out-of-range position index from the command mapping (or Ctrl+C). The control loop will disable motor torque and close the serial port cleanly before exiting.

---

Data Logging

Two parallel logs are written on every run:

EMG log (emg_logs/<NAME>/bypass_<timestamp>) — raw streamed EMG via LibEMG's log_to_file.

Bypass log (bypass_log/<NAME>/bypass_log_<timestamp>.csv) — one row per control cycle at 60 Hz containing:

time, w_pos_d, w_vel_d, w_trq_d, g_pos_d, g_vel_d, g_trq_d,
w_pos, w_vel, w_trq, g_pos, g_vel, g_trq,
velocity, probs_0, probs_1, probs_2, probs_3, probs_4

_d columns are commanded (desired) values; unprefixed columns are present (feedback) values read via bulk read from the Dynamixel registers.

---

Notes

• The / 128.0 scaling applied inside every model forward pass is bit-depth normalization for the Myo's signed 8-bit ADC. It is not a learned or per-user normalization step.
• The MultiModelWrapper in main.py supports hot-swapping between loaded models at runtime by updating SharedContext.active_model_name from an external process or UI. The switch is detected on the next forward call.
• A monkey patch is applied to Protocol1PacketHandler.bulkReadTx to work around a parameter signature mismatch in some versions of the Dynamixel SDK. If you upgrade the SDK, verify this is still necessary.
• SharedContext is a multiprocessing.Manager().Namespace() object, allowing state to be shared safely between the main process (Bypass control loop) and the LibEMG subprocess.

---

License

MIT — see LICENSE.