Realizing an intelligent agricultural robot: An analysis of the ease of tomato harvesting

Abstract

• Developed a tomato harvesting robot capable of autonomous motion. • Designed a harvesting strategy with variable approach directions. • Demonstrated 81% success rate through real harvesting experiments. • Statistically identified spatial features influencing harvesting success. • Implemented image-based analysis to estimate success rate and approach direction. Labor shortages and the need for labor-saving solutions in agriculture have driven the development of autonomous harvesting robots. However, challenges such as unreliable harvest performance and limited reproducibility hinder their practical deployment. This study investigated the structural and spatial factors influencing harvesting success and approach direction with the aim of improving the autonomy and reliability of robotic tomato harvesting in plant factories. A tomato-harvesting robot was developed and tested in an actual plant factory environment. The robot was equipped with a vehicle, multi-axis manipulators, a gripper-type end-effector, and an RGB-D camera. A total of 100 tomatoes were targeted, and harvesting was attempted from three directions: front, left, and right. Chi-square tests and logistic regression analyses were performed to identify the dominant variables influencing harvest success. The results indicated that obstacles located in front of the fruit, such as peduncles, significantly reduced the success rate, whereas certain spatial configurations, such as peduncles above the fruit, improved success. To generalize these findings, this study developed an image-based method using YOLOv8 for object detection and semantic segmentation. The detected features were linked to logistic models to estimate the harvesting success probability and determine the approach direction. The results demonstrate that statistical analysis with computer vision can enable autonomous, environment-aware decision-making in robotic harvesting. Although challenges, such as variations in fruit detachment and finger control, remain, these findings offer a promising path toward reliable and intelligent harvesting systems and set the foundation for virtual simulation-based refinement in future work.