A robotic harvesting system for occluded cucumbers using F2SA-YOLOv8 and HVSC

• Type-2 Fuzzy Self-Attention improves detection robustness under occlusion. • F2SA-YOLOv8 outperforms baseline YOLOv8 in occluded cucumber detection. • Multifunctional end-effector integrates cutting, grasping, and foliage-pushing. • A staged Hybrid Visual Servo Control enhances spatial alignment under occlusion. Cucumber harvesting in greenhouse environments faces challenges such as occluded cut-points and overlapping plant structures. This study presents a fully integrated robotic harvesting system that combines perception, control, and end-effector innovations to address these issues. In the perception module, a Type-2 Fuzzy Self-Attention (F2SA)-enhanced YOLOv8 module improves detection under occlusion, achieving an F1-score of 0.92635 with an average inference time of 256 ms per image, compared to 0.92056 F1-score and 903.9 ms for the YOLOv8n. The multifunctional end-effector enables simultaneous cutting and grasping, simplifying the mechanism and improving efficiency, while an eye-in-hand depth camera allows physical interaction with foliage to reveal occluded cut-points. A staged Hybrid Visual Servo Control (HVSC) strategy ensures reliable cut-point alignment even under occlusion. Experiments across five occlusion scenarios achieved success rates over 90% in laboratory trials and 81% in greenhouse tests, with average harvesting times of 55 s and 73.16 s per plant. These results demonstrate the effectiveness of a compact and occlusion-resilient design for real-time cucumber harvesting in cluttered greenhouse environments.