Sensing-based Robustness Challenges in Agricultural Robotic Harvesting
This paper presents the challenges agricultural robotic harvesters face in detecting and localising fruits under various environmental disturbances. In controlled laboratory settings, both the traditional HSV (Hue Saturation Value) transformation and the YOLOv8 (You Only Look Once) deep learning model were employed. However, only YOLOv8 was utilised in outdoor experiments, as the HSV transformation was not capable of accurately drawing fruit contours. Experiments include ten distinct fruit patterns with six apples and six oranges. A grid structure for homography (perspective) transformation was employed to convert detected midpoints into 3D world coordinates. The experiments evaluated detection and localisation under varying lighting and background disturbances, revealing accurate performance indoors, but significant challenges outdoors. Our results show that indoor experiments using YOLOv8 achieved 100% detection accuracy, while outdoor conditions decreased performance, with an average accuracy of 69.15% for YOLOv8 under direct sunlight. The study demonstrates that real-world applications reveal significant limitations due to changing lighting, background disturbances, and colour and shape variability. These findings underscore the need for further refinement of algorithms and sensors to enhance the robustness of robotic harvesters for agricultural use.