Asparagus harvesting can be confined to the efficacy of robotic vision in recent years. Asparagus spears are characterized by a slender morphology in their natural growth state. These tender stems are highly prone to mutual occlusion and overlapping when growing densely in field conditions. Furthermore, the stout mother stems can simultaneously present as the complex background interference. Collectively, it is often required for the high accuracy of the multi-target segmentation and recognition using machine vision. In this study, the lightweight instance segmentation model (YOLO11n-seg) was adopted as a baseline, in order to improve the precise positioning and harvesting performance of the robotic end-effector. Consequently, an optimized model named YOLO11n-SAL was also proposed to specifically tailor the slender, occluded targets with high fidelity. Two modules were introduced to enhance the feature extraction and attention mechanisms in the architectural framework. Firstly, the multi-scale edge enhancement Module (MEEM) was conceptually designed and integrated in order to mitigate the challenge wherein the edge features of the slender asparagus targets were inherently weak and easily lost during convolutional operations. Multi-scale decomposition was performed on the convolutional feature maps. The MEEM effectively extracted and intensified the edge and contour information before feature fusion. The sensitivity to the target boundaries was significantly elevated for the high segmentation precision, thereby enhancing the perceptual capability of the targets with the slender morphological structures. Secondly, the separated and enhancement attention module (SEAM) was introduced to rectify the feature confusion and data incompleteness caused by inter-target occlusion. Attention separation over both channel and spatial dimensions was also utilized to adaptively perceive the local and global features of the occluded asparagus at the varying scales. These features were selectively enhanced and effectively fused to better position the visible subjects of the partially masked targets, while suppressing the background noise and distractor information. The robust performance of the detection and recognition was maintained even within the complex and cluttered environments. A series of experiments was conducted to verify the effectiveness of the improved model. Quantitative evaluation results indicate that the improved YOLO11n-SAL model achieved significant gains over all key performance indicators, compared with the baseline model. In the detection task of the target bounding box, the superior performance was achieved with a detection precision of 94.2%, a recall rate of 83.1%, a mean average precision at IoU threshold 0.5 (mAP_0.5) of 91.2%, and a mean average precision at IoU threshold 0.5-0.95(mAP_0.5-0.95) of 76.2%. In the more granular instance mask segmentation, the model also performed impressively. The segmentation precision, recall, mAP_0.5 and mAP_0.5-0.95 reached 93.4%, 77.9%, 90.7%, and 62.7%, respectively. Furthermore, the heatmap analysis demonstrated that the YOLO11n-SAL model was markedly improved to perceive the asparagus edge features over different scenarios, with the superior multi-target segmentation and recognition under occluded conditions. The high accuracy of the segmentation and recognition was achieved to reduce the interference in the complex multi-scenario environments, compared with the baseline. Finally, a series of asparagus recognition, positioning, harvesting, and grasping trials were carried out using depth cameras and mechanical arms, in order to validate the cognition and position performance in the actual deployment scenarios. The empirical results showed that a positioning success rate of not less than 90% was accompanied by effective harvesting and grasping performance. These findings can provide reliable technical support for the advancement of robotic harvesting in precision agriculture.

Fruit and vegetable harvesting robots have been the cutting-edge research hotspots in the important branches of agricultural machinery. The automation level can also be enhanced for labor-saving in the agricultural industry. In this study, a systematic review was presented on the current status and trends of the fruit and vegetable harvesting robots in the world. The technological breakthroughs and engineering practices were analyzed in the evolution of the typical machine from laboratory prototypes to commercial products. The current progress has focused on the top international journals and the industry over the past five years. Key mechanisms and core technologies were summarized for the fruit and vegetable harvesting robots. The technical path of the visual perception was then examined from the two-dimensional imaging to three-dimensional multimodal perception. A comparison was also made on the differences in the rate range, accuracy, and environmental adaptability between passive stereo vision and active structured light, as well as the time-of-flight methods. The matching failure was determined after the texture loss in the unstructured environment of the farmland. In the walking mechanism section, there was the terrain adaptation of the wheeled, tracked, and hybrid drive modes in different scenarios, such as flat land, hills, and greenhouses. The performance was also compared on their obstacle crossing, turning radius, energy consumption efficiency, and the impact of the vibration on the end precision. Among them, the end effector was the core working component. Three mainstream schemes were then classified: negative pressure adsorption, tool shearing, and flexible grasping type. Grasping-picking coordination strategies were evaluated as suitable for the different fruit and vegetable scenarios, such as apples, strawberries, and tomatoes. Each scheme was explored on the damage rate, single fruit operation time, success rate, and crop characteristic constraints. In the core technology, the navigation and positioning evolved from the single GNSS to the multi-sensor fusion SLAM. Some strategies were proposed for the satellite signal occlusion under dense planting environments in orchards, such as the tightly coupled vision-inertial-wheel speed. Furthermore, the target recognition was also summarized from conventional image processing to deep learning. A trade-off optimization was conducted on the Faster R-CNN, Mask R-CNN, and YOLO series models in terms of detection accuracy and real-time performance. The performance was degraded under occlusion, drastic changes in lighting, and interference from similar color tones. In mechanical arm path planning, the swarm intelligence algorithms were evaluated to rapidly expand the random trees, graph search, and artificial potential field in the static obstacle avoidance and dynamic response. Motion redundancy and trajectory smoothing were observed in the multi-arm collaboration and mobile grasping. The current technical bottlenecks were attributed to the five aspects: low flexibility of the end effector leading to mechanical damage rate; low dynamic adaptability of planning, with the lagging responses to dynamic obstacles, such as the swaying branches and leaves; severe degradation of the perception accuracy in the complex environments; lack of the multi-machine collaboration resulting in the repeated or missed operations; all-weather operation with the dust- and water-proof grades and battery life unable to meet the continuous operation requirements of the large fields. Finally, a perception-decision-control system was integrated to break through the existing technical barriers in smart agriculture. Some recommendations were then proposed, including the collaborative design of the software and hardware, multi-modal data fusion, agricultural large models training, and swarm intelligence scheduling, picking robots towards high efficiency, low damage, and generalization. The finding can also provide the core equipment support for the future smart agriculture, particularly from single-machine autonomy to swarm intelligence.