Dual-range vision-based cutting-point localization and end-effector positioning for robotic table grape inflorescence trimming

Inflorescence trimming is essential for producing high-quality table grapes, but it is labor-intensive and demands skilled labor. Automating it is difficult: grape inflorescences have a complex branching structure, are frequently occluded, and are viewed under highly variable illumination, and the cutting tool must be positioned precisely. This study presents perception and motion control methods for a grape inflorescence trimming robot. The methods localize the cutting point on the inflorescence and position the cutting tool at that point, using a dual-range framework. At far range, a detection model identifies the target inflorescence and guides the manipulator to a suitable observation distance; at close range, a segmentation-based method localizes the cutting point on the main stem. This study covers cutting-point localization and tool positioning only. The robot stopped when the scissors reached the cutting point; the stem was not cut, so the trimming result itself was not evaluated. To improve perception robustness, multi-frame detection and real-time illumination augmentation were introduced. Two motion control strategies were developed and compared: sequential trajectory control (STC) and hybrid feedback control (HFC), the latter using real-time visual feedback for continuous motion correction. The system was evaluated in a greenhouse vineyard on 30 preselected inflorescences of the table grape cultivar Shine Muscat grown under a pergola training system. The mobile platform was positioned manually in front of each target. Each inflorescence was processed once by each control strategy, yielding 60 trials in total. Positioning was counted as successful when the main stem was between the scissor blades and within the effective cutting range. STC achieved a positioning success rate of 90.0 %, a mean positional error of 2.1 cm, and a mean operation time of 31.9 s. HFC achieved a positioning success rate of 83.3 %, a smaller mean positional error of 1.4 cm, and a longer mean operation time of 35.4 s. In successful trials, HFC placed the scissors where a cut would keep exactly the two uppermost lateral branches in 52.0 % of cases, compared with 33.3 % for STC. Visual feedback therefore reduced positioning error but did not increase positioning success or shorten operation time. Real-time illumination augmentation raised the target detection rate from 41.6 % to 100.0 % across 125 frames covering cloudy and sunny conditions at multiple locations in the vineyard, and multi-frame aggregation eliminated the 3.0 % single-frame target selection error. Perception took 58.5 % of the total operation time under STC and was the main time bottleneck. These results show that the proposed methods can localize the cutting point and position the cutting tool on a single cultivar in a greenhouse vineyard under natural illumination and occlusion.