This repository is the continuation of the QT-Touch project.
The goal of QT-jacket is to extend the scripts to capture more sensors from a piezo-resistive and piezo-electric jacket.
This architecture is used to classify different types of touches on the robot's jacket. We defined 5 classes for the data acquisition:
- Class 0 : Nothing / Background noise (Robot movements, parasite noises).
- Class 1 : Tap_Attention (Left or Right Arm) -> Short peaks.
- Class 2 : Comforting_Caress (Back or Arm) -> Light, continuous and sliding pressure.
- Class 3 : Tickles (Left + Right Torso) -> Fast and irregular pressure variations.
- Class 4 : Hug / Cuddle (Global) -> Enveloping, increasing then maintained pressure.
- Acquisition: 1000 points per second (Hz) per sensor. Each acquisition event lasts 5 seconds, resulting in a total of 5000 points per event.
- Processing: Decimation to 500 points * 5 sensors.
- Dataset: 120 examples per class (equally distributed).
- CNN Network: Input dimension
[500, 5].
Check out the tutorial to get started with the Raspberry Pi setup and data acquisition.
uv venv
uv pip install -r requirements.txt
uv run python train/preprocess_data.py
# Training (Best Configuration: Derivative + Dropout)
uv run python train/train.py --derivate --dropout
# Other available options:
# uv run python train/train.py # Raw baseline (5 channels)
# uv run python train/train.py --extend # 10 channels
# Export the trained model to ONNX for the Raspberry Pi
uv run python export_onnx.py --derivate --dropoutHere is the evolution of the training performances and the resulting confusion matrices:
For inference on a Raspberry Pi or any local device, you only need the inference.py script and the exported tactile_deriv_drop_model.onnx file. The tutorial contains all the necessary details to set this up.
This project is licensed under the Apache License 2.0.