Jeremy Wang¹ and Jonathan Sahagun², ¹USA, ²California State Polytechnic University, USA

Within a work environment, the cleanliness and organization of tools contribute greatly to the workers' mental state and psychological well-being, allowing for more creativity and productivity in the workplace [1]. This is largely an issue within the industrial industry, where machines need to be built, maintained, and repaired. While being automated in larger factories, the tedious task of sorting hardware, such as screws, nuts, and bolts, is often performed manually in the context of smaller-scale factories and repair centers. However, the tediousness of this task often undermines the importance of the task and is often neglected or performed poorly. Existing industrial solutions to this issue are expensive and inflexible. This paper presents a low-cost, autonomous hardware sorting system that uses a custom-built Convolutional Neural Network (CNN) object detection model trained using TensorFlow and Keras. The system runs entirely on a Raspberry Pi 5 and uses a Microsoft Lifecam Studio webcam together with an H-bot gantry for mechanical sorting. The primary focus of research is on the optimization of the CNN for real-time deployment on resource-constrained hardware. Multiple lightweight architectures such as YOLOv8-Nano and MobileNetV2-SSD are proposed for examination and evaluated. A custom dataset was created and labeled using Roboflow, with images consisting of three hardware classes: screws, nuts, and standoffs. The trained model reached a mean average precision (mAP) of 91.5%, with ∼125 ms for each inference while on the Raspberry Pi. When integrated with the mechanical system, the full pipeline sorted hardware at an average rate of 18.6 parts per minute with an accuracy of 90.0%. As the project is built with a budget of $300, this project demonstrates the feasibility of deploying lightweight deep learning models for automation tasks on embedded systems.

Convolutional Neural Network, Keras, Tensorflow, Edge AI, Model Optimization