Vision Based Robotic Autonomous Refueling System for Highway Service Areas

Abstract

Highway service-area fueling stations face recurring challenges in labor cost, centralized management, and operational safety. This paper presents a prototype autonomous visual unmanned gas-station system that combines YOLOv5s object detection, a Raspberry Pi-controlled two-degree-of-freedom manipulator, and an ESP32-based Internet-of-Things (IoT) interface. A camera detects the fuel-inlet proxy target, and its image coordinates are transmitted through a serial link to an incremental PI controller that aligns the nozzle. Users select a fueling amount through a mobile interface connected to the Bafa Cloud platform. A 900-image dataset was collected under varied viewpoints and backgrounds; the model was trained for 400 epochs with an 80/20 train–validation split. The prototype achieved detection accuracies above 80% on the tested target shapes, while nozzle positioning errors remained below 3 cm. The results demonstrate the feasibility of integrating lightweight deep learning, embedded control, and cloud interaction for unattended fueling, while also revealing the need for larger real-vehicle datasets and additional safety sensing.