Abstract: Here, an intelligent impurity removal system was developed for fresh Lycium barbarum (L. berries) using a Delta robot, machine vision, and an air-suction end-effector. Four modules consisted of material conveying, visual detection, a central control, and sorting actuation. A “conveying–stopping–recognition–grasping–resuming” mode was adopted during operation. The conveyor transported the fruit-leaf mixture into the detection area and then stopped temporarily, thus reducing the influence of material motion on image acquisition and robotic grasping. An Intel RealSense D435i depth camera was mounted above the working space of the Delta robot to collect aligned RGB and depth images. A YOLOv8n model was used to detect leaf impurities. The center point of each bounding box was taken as the target pixel coordinate. Combined with the depth value, the target point was converted into three-dimensional coordinates in the camera coordinates using the pinhole camera model. According to the three-dimensional points, a hand-eye calibration was then used to transform the target coordinates into the robot base coordinates. A hydrophobic breathable fabric was added to the suction port of the air-suction end-effector. Leaf release reliability was then improved after suction, according to geometric profiles and mass of fresh goji leaves. The average longitudinal and transverse diameters of sampled leaves were 57.65 and 16.84 mm, respectively, while the maximum mass was 0.48 g after measurement. Three suction port configurations were compared: a conventional flexible silicone suction cup, a rigid suction port without a suction cup, and a hydrophobic breathable fabric composite suction port. Among them, the silicone suction cup provided good sealing, but the leaf release time was 2–5 s after the air source was shutdown, because of residual negative pressure. The rigid suction port reduced dead volume, but liquid bridge adhesion occurred in 78% of repeated grasping tests. By contrast, the composite suction port also reduced the leaf release time to less than 0.5 s, and the liquid bridge adhesion probability was lower than 5%. The hydrophobic breathable fabric was used to suppress liquid film spreading and then accelerate pressure recovery after suction shutdown. A ROS2 control system was constructed to coordinate visual perception, target localization, trajectory planning, execution control, and state scheduling. The image dataset contained 1,000 original images of mixed goji berries and leaves under stable indoor illumination, of 10% of which were the degrees of leaf-leaf or fruit-leaf occlusion. After data augmentation, the dataset was expanded to 4,000 images and then divided into training and validation sets at a ratio of 9:1. After that, YOLOv8n was compared with YOLOv5n, YOLOv8s, YOLOv8m, and YOLOv9c under the same dataset and hardware. The results showed that the YOLOv8n model achieved an mAP@0.5 of 99.2%, an mAP@0.5:0.95 of 69.4%, a Precision of 97.3%, and a Recall of 98.6%, with an inference speed of 185.2 frames per second (FPS) and a model size of 5.95 MB. YOLOv8n model was selected for real-time recognition of leaf impurity, according to detection accuracy, inference speed, and model size. The hand-eye calibration test showed that the average spatial positioning error was 1.53 mm after coordinate transformation, and the root mean square error was 1.91 mm, fully meeting the positioning requirement for leaf suction grasping. Four joint-space trajectory planning approaches were compared, including point-to-point motion , linear segments with parabolic blends, S-Curve interpolation, and quintic polynomial interpolation. Specifically, Point-to-point motion achieved the shortest single operation cycle time of 2.906 s, with a theoretical throughput of 20.6 operations per minute. Compared with linear segments with parabolic blends, S-curve interpolation, and quintic polynomial interpolation, the point-to-point motion operation cycle time was reduced by 48.4%, 49.9%, and 50.7%, respectively. There was a difference in the pure motion time of the robot. According to the low mass of goji leaves and the short-cycle operation, PTP was adopted as the motion strategy for impurity removal tests. Orthogonal experiments were conducted to clarify the effects of conveyor speed and impurity density on system performance. Conveyor speed was set at 7.9, 23.6, and 39.3 mm/s, while impurity density was set at 1, 5, and 10 leaves per 10 fruits. Range analysis showed that impurity density had a greater effect on both impurity removal rate and single operation cycle than conveyor speed. As impurity density increased from 1 leaf per 10 fruits to 10 leaves per 10 fruits, the average impurity removal rate decreased from 100.0% to 87.3%, and the single operation cycle increased from 2.98 to 3.10 s. The removal performance was attributed to leaf overlap, occlusion, stacking, and severe curling, leading to detection errors, localization deviation, or insufficient sealing between the suction port and the leaf surface. An impurity removal test was performed under the conveyor speed of 23.6 mm/s and an impurity density of 5 leaves per 10 fruits. In 20 repeated tests, the system achieved an impurity removal rate of 97.0%, a loss rate of 0.5%, and a single operation cycle of 3.127 ± 0.278 s. The single-grasp success rate was 89%. The failure modes were visual localization deviation, visual recognition error, and severe leaf curling, accounting for about 45%, 35%, and 20% of grasping failures, respectively. Overall, the online recognition, three-dimensional localization, selective grasping, and impurity placement were realized to remove postharvest impurity from fresh goji berries. The findings can provide a technical reference for the low-damage impurity removal for small berry crops.