A Deployment-Oriented Benchmarking of You Look Only Once (YOLO) Models for Orange Detection and Segmentation in Agricultural Robotics

The deployment of autonomous robots is critical for advancing sustainable agriculture, but their effectiveness hinges on visual perception systems that can reliably operate in natural, real-world environments. Selecting an appropriate vision model for these robots requires a practical evaluation that extends beyond standard accuracy metrics to include critical deployment factors such as computational efficiency, energy consumption, and robustness to environmental disturbances. To address this need, this study presents a deployment-oriented benchmark of state-of-the-art You Look Only Once (YOLO)-based models for orange detection and segmentation. Following a systematic process, the selected models were evaluated on a unified public dataset, annotated to rigorously assess real-world challenges. Performance was compared across five key dimensions: (i) identification accurac, (ii) robustness, (iii) model complexity, (iv) execution time, and (v) energy consump-tion. The results show that the YOLOv5 variants achieved the most accurate detection and segmentation. Notably, YOLO11-based models demonstrated strong and consistent results under all disturbance levels, highlighting their robustness. Lightweight architectures proved well-suited for resource-constrained operations. Interestingly, custom models did not consistently outperform their baselines, while nanoscale models showed demonstra-ble potential for meeting real-time and energy-efficient requirements. These findings offer valuable, evidence-based guidelines for the vision systems of precision agriculture robots.