An OpenBCI brain-computer interface that classifies EEG with EEGNet and LDA and turns it into S/F/B/R/L commands for a model car via Arduino and HC-05.
A project building a non-invasive brain-computer interface (BCI) that captures electroencephalography (EEG) signals from an 8-channel OpenBCI Cyton board, cleans them, segments them into time windows, and classifies them into five motor states (rest, right hand, left hand, right foot, left foot movement), then translates the classification into driving commands sent over a serial link (USB or HC-05 Bluetooth) to an Arduino Uno that switches four relays controlling the motors of a model car. The project was carried out by a team of students with technical assistance from Techno Enjaz.
The system is written in Python around a unified processing core that serves three tracks: offline training on the public PhysioNet dataset, personal calibration through live recording, and real-time inference with immediate command transmission. The best documented results are 86.22% accuracy for distinguishing eyes open from eyes closed, about 69.42% for binary motor tasks on PhysioNet, and a preliminary end-to-end latency under 500 ms from signal acquisition to command; personal motor-imagery models, however, stayed close to chance level, which the book states openly.
| Item | Details |
|---|---|
| Project type | Software and hardware BCI system controlling a model car |
| Field | Brain-computer interfaces, biosignal processing, machine learning, embedded systems |
| Project status | Working educational research prototype; not a medical device |
| Techno Enjaz role | Technical support and assistance to the students during project development |
| Core technologies | OpenBCI Cyton, BrainFlow, Python 3.11, SciPy, MNE, scikit-learn, PyTorch (EEGNet, ShallowConvNet), Shrinkage LDA, FastAPI, Tkinter, Arduino Uno, HC-05 |
| Outputs | Predicted class with confidence, signal-quality reports, trained models and a model registry, S/F/B/R/L driving commands for the car |
EEG offers a non-invasive, low-cost way to capture neural activity, open-source boards such as OpenBCI make experimental systems feasible, and the PhysioNet dataset provides recordings of 109 people in more than 1,500 recordings. Moving from a model trained on public data to live control of a real device, however, runs into a gap the project describes in four dimensions:
The project was carried out academically by the students, with technical assistance and support from Techno Enjaz during development. It is presented among the office's work as an example of its support for projects that connect biosignal processing and machine learning to a microcontroller-based physical control layer.
| Model | Description in the project |
|---|---|
| EEGNet | Compact convolutional network designed for EEG, about 5.5 thousand trainable parameters in the three-class version |
| ShallowConvNet | Shallow network that learns temporal and spatial filters and then computes log-scaled power, about 14.4 thousand parameters in the binary version |
| Shrinkage LDA | Linear discriminant analysis with covariance shrinkage on band-power features from five frequency bands and log-covariance matrices; the preferred track for small personal datasets |
| One-Class SVM | Anomaly detector with an RBF kernel |
| DeepConvNet | Architecture present in the code, with no trained model as of the review date |
Deep networks were trained in PyTorch with weighted cross-entropy loss to handle class imbalance, the Adam optimizer with an initial learning rate of 0.001, and early stopping after 15 epochs without improvement. To prevent leakage, the updated pipeline splits data stratified at the trial level before windows are extracted.
| Component | Role |
|---|---|
| OpenBCI Cyton | 8-channel EEG acquisition board at 250 samples/second |
| Arduino Uno (ATmega328P, 16 MHz) | Receives serial commands, decodes them, and drives the relays |
| Relay module (4 channels used) | Isolates the controller circuit (5 V) from the motor circuit and switches the motors per command |
| HC-05 Bluetooth | Wireless serial link appearing as a virtual COM port with a range of about 10 m; USB can be used instead |
| Step-down converter | Converts the car battery voltage (7.2–12 V) to the stable voltages the controller and modules need |
| Model car | Physical test platform for the closed control loop |
The project uses a single repository with a shared core used by three interfaces: a web interface (FastAPI and the Uvicorn server with plain HTML, CSS, and JavaScript), a desktop interface (Tkinter), and a command-line interface (CLI). Its layers cover signal sources (the board, PhysioNet files, local files, and a Mock signal generator for testing without hardware), raw storage, shared processing, learning, services, and outputs.
A 17-table data model was designed in four groups (data collection and sessions, signal quality, dataset building and training, models and predictions). It mirrors the current file-based storage (EDF, CSV, JSON, NPZ, PT, joblib) and uses anonymous participant IDs such as S001. The system also checks each model's metadata for compatibility (channels, sampling rate, window length, normalization), blocks loading of outdated incompatible models, and promotes a model only when it passes a performance threshold.
The interface includes screens for collecting five-movement data, viewing the live signal, configuring the session and protocol timings, checking electrode readiness, running the session stages (baseline, preparation, movement, rest), training and testing models, browsing the model registry, and starting live prediction.
| Experiment | Result |
|---|---|
| Eyes open vs. eyes closed | 86.22% (highest recorded accuracy) |
| Best binary motor task on PhysioNet | About 69.42% |
| Three-class personal calibration (rest, right hand, left hand) with deep learning | Test accuracy between 0% and 21.43% |
| Personal calibration with LDA (cross-validated balanced accuracy) | 33.63%, close to chance level (33.33%) |
| End-to-end closed-loop latency (preliminary results) | Under 500 ms from signal acquisition to command |
The project concludes that the engineering goal was achieved: a unified architecture running across all three tracks and a closed control loop tested on a model car. The scientific validity of personal models was not yet achieved; a typical personal calibration session had no more than 33 trials, and validation was confined to the same session. The project also observed that a classical classifier can outperform deep networks when personal data are scarce.
The system is an educational research tool, not a medical device. Any use with people or in open environments requires physical and software safety layers and a thorough risk assessment, and controlling a real vehicle or hazardous machinery would require standards and certification beyond the scope of this project. Brain signals are also sensitive biological data, and the data model was designed to use anonymous participant IDs.
If you are working on a brain-computer interface or bio-signal control system and need technical help developing the idea or building a prototype, you can contact the Techno Enjaz office to discuss your project's requirements and the appropriate scope of support.
An OpenBCI Cyton board captures 8-channel EEG at 250 samples/second; the signal is filtered with a 50 Hz notch and an 8–30 Hz band-pass, re-referenced, and normalized. EEGNet, ShallowConvNet, or LDA classifies each time window, and a command mapper turns the class into S, F, B, R, or L, sent over USB or HC-05 Bluetooth to an Arduino Uno that switches the model car's motor relays.
The eyes open/closed model reached 86.22% accuracy, the best binary motor task on PhysioNet reached about 69.42%, and preliminary tests showed under 500 ms from signal acquisition to command. Three-class personal calibration stayed near chance: 0% to 21.43% with deep learning and 33.63% balanced accuracy with LDA.
The software is written in Python with BrainFlow, SciPy, MNE, scikit-learn, and PyTorch, with a FastAPI web interface and a Tkinter desktop interface. The hardware includes an OpenBCI Cyton board, an Arduino Uno, a relay module with four channels in use, an HC-05 Bluetooth module, a step-down converter, and a model car.
The project attributes it to scarce calibration data (no more than 33 trials in a typical session), poor quality on some channels, and the domain shift between PhysioNet (64 channels, 160 Hz) and OpenBCI (8 channels, 250 Hz). It proposes multiple sessions with at least 20 valid trials per class, a strict quality gate, and splitting at the session or subject level.
No; it is an educational research tool, not a medical device. Any use with people or in open environments requires safety layers such as an emergency stop, speed limits, and rejection of low-confidence commands, plus a thorough risk assessment.
Reach out to Techno Enjaz engineering team to discuss technical feasibility, prototype hardware, and development scope.
Discover more engineering prototypes and academic implementations
A remotely controlled ground robot prototype combining Arduino UNO, three wireless links (Bluetooth, Wi-Fi, 433 MHz RF) and a camera whose images are analyzed with YOLO.
A robotic hand that mimics a human hand opening and closing in real time via camera hand tracking, sending finger states over Bluetooth to an Arduino UNO driving servos.
A prototype that measures a tank's water-level percentage with an HC-SR04 sensor and Arduino Nano, showing it on an LCD and sending it via HC-05 Bluetooth to a phone app.
