TinyML-Based Self-Balancing Robot Control with Arduino Nano 33 BLE Sense

Overview

This project demonstrates the implementation of a Tiny Machine Learning (TinyML) model for controlling a two-wheeled self-balancing robot using the Arduino Nano 33 BLE Sense. The objective is to replace a traditional control algorithm with a lightweight neural network capable of performing real-time inference on resource-constrained embedded hardware.

Project Workflow

The development process consists of the following steps:

1. System Modeling and Controller Design

   • A mathematical model of the self-balancing robot was developed in MATLAB/Simulink.
   • A Full-State Feedback (FSB) controller was designed using the Linear Quadratic Regulator (LQR) method to achieve stable balancing performance.

2. Data Collection

   • While the robot was controlled by the FSB controller, data were collected for machine learning.

   • The recorded features included:

     • Pitch angle (estimated using a complementary filter)
     • Angular velocity
     • FSB controller output (control signal)

3. Model Training

   • The collected dataset was imported into Edge Impulse.
   • A TinyML model was trained to approximate the behavior of the LQR-based controller.

4. Deployment

   • The trained model was converted to TensorFlow Lite Micro.
   • It was quantized to 8-bit integer precision to reduce memory usage and computational complexity.
   • The final model was deployed and executed on the Arduino Nano 33 BLE Sense for real-time balance control.

Technologies Used

• Arduino Nano 33 BLE Sense
• MATLAB / Simulink
• Linear Quadratic Regulator (LQR)
• Full-State Feedback (FSB) Control
• Edge Impulse
• TensorFlow Lite Micro
• Tiny Machine Learning (TinyML)

Repository Structure

├── Arduino/          # Arduino source code
├── MATLAB/           # MATLAB/Simulink models and scripts
├── Model/            # Trained TinyML model and TensorFlow Lite files
├── Data/             # Collected datasets 
└── README.md

Requirements & Setup

Before running the Arduino code, the following setup steps are required:

1. Install Required Library

You must manually add the following library to the Arduino IDE:

ei-complimentary_98-arduino-1.0.6.zip (located in the text model/ folder)

To install it:

Open Arduino IDE

Go to Sketch → Include Library → Add .ZIP Library

Select the file from the model folder

2. IMU Calibration

IMU calibration was required to ensure accurate sensor readings for stable balancing performance.

The calibration procedure and code are included in the Arduino/ folder. Make sure to run the calibration script before uploading the main control code.

Results

The trained TinyML model successfully reproduces the behavior of the LQR-based controller while being compact enough to run on the Arduino Nano 33 BLE Sense. Through 8-bit quantization and TensorFlow Lite Micro deployment, the model achieves efficient real-time inference suitable for embedded applications.

Future Work

• Improve model accuracy through additional training data.
• Compare TinyML performance with classical control methods under disturbances.
• Optimize inference latency and power consumption.
• Explore more advanced neural network architectures for embedded control.

License

This project is released under the MIT License. Feel free to use, modify, and distribute it for research and educational purposes.