Low-Cost Embedded Computer Vision and Neural Network-Enabled Autonomous Navigation for a Four-Wheeled Mobile Robot Using an ESP32 and Arduino Framework

Volume 17, Issue 2 (November 2026), pp. 65–72

Download PDF

@article{jocse-17-2-10,
  author={Abhiram Kaakarla},
  title={Low-Cost Embedded Computer Vision and Neural Network-Enabled Autonomous Navigation for a Four-Wheeled Mobile Robot Using an ESP32 and Arduino Framework},
  journal={The Journal of Computational Science Education},
  year=2026,
  month=nov,
  volume=17,
  issue=2,
  pages={65--72},
  doi={https://doi.org/10.22369/issn.2153-4136/17/2/10}
}

This paper presents the design, construction, and testing of a low-cost autonomous navigation system built on the Freenove 4WD robot car. An ESP32-CAM microcontroller handles image capture and onboard machine learning while an Arduino Uno controls the four DC motors, with all hardware costing under 103 US dollars. A convolutional neural network trained through Edge Impulse was compressed to an 18.2 KB quantized model that runs directly on the ESP32, classifying 32 by 32 pixel grayscale camera frames into five sign categories: stop sign, left turn, right turn, T intersection, and forward arrow. An OpenCV image processing pipeline running alongside the neural network performs Canny edge detection and Hough line transform for lane tracking and pixel-density-based obstacle avoidance. Navigation decisions are encoded into short serial packets with CRC error checking and sent from the ESP32 to the Arduino, which converts them into PWM motor signals. Across 500 controlled test runs, the robot correctly identified signs 91.78 percent of the time under even lighting and 69.46 percent under uneven lighting. The full decision loop from camera capture to motor response averaged 66.6 milliseconds. A systematic failure analysis across all 500 trials categorized errors into hardware, environmental, and software causes; five targeted fixes reduced the overall failure rate from 36.8 percent to 11.4 percent. The project also served as a comprehensive self-directed computational science education exercise, and the paper includes a dedicated student reflection documenting technical lessons, methodological insights, and educational impact. All firmware, training data, and experimental records are publicly available for reproduction and improvement by other students and educators.