Key technologies of tomato-picking robots based on machine vision

To address the challenges of harsh harvesting environments, high labor intensity, and low picking efficiency in tomato harvesting, this study investigates the key technologies related to the end-effector design, detection and recognition, and spatial localization of tomato-picking robots. A non-contact cavity-type end-effector is designed, which effectively prevents tomato damage caused by compression during picking while preserving the peduncle. Additionally, the motion of the robotic arm is simulated for performance analysis. Subsequently, tomato images are captured and annotated for training deep neural network models. Both the original YOLO v8n and the improved YOLO v8n models are used for tomato image detection, with a focus on the impact of varying light intensities and different tomato maturities on recognition and localization accuracy. Experimental results demonstrate that the robot’s vision system achieves optimal recognition and localization performance under light intensities ranging from 20 000 to 30 000 lx, with an accuracy of 91.5%, an average image detection speed of 15.1 ms per image, and an absolute localization error of 1.55 cm. Furthermore, the prototype tomato-picking robot’s end-effector successfully performed stable grasping of individual tomatoes without damaging the skin, achieving a picking success rate of 83.3%, with an average picking time of approximately 9.5 s per fruit. This study provides a technical support for the automated harvesting of tomato-picking robots.