Harvesting Robot Research Articles

MFENet: Multi-scale feature extraction network for images deblurring and segmentation of swinging wolfberry branch

Wolfberry harvesting faces challenges related to factors such as wind-induced branch movement and mechanical arm collisions, which lead to undesirable swinging of wolfberry branches. This motion results in blurred images captured by imaging devices, which hampers the effectiveness of the harvesting robot arm in identifying and grasping the wolfberry branches. Clear, unblurred images are imperative for precise branch recognition and successful harvesting. Unfortunately, prior research on wolfberry branch identification has largely overlooked the issue of motion blur. In this study, we propose a novel solution by introducing the Multi-Scale Feature Extraction Network (MFENet) specifically tailored for image deblurring tasks. Our approach aims to enhance the accuracy of identifying and tracking dynamically moving wolfberry branches. MFENet achieves this by harnessing a combination of multi-scale image content features, incorporating modules for multi-scale depth-separable convolution, multi-scale receptive field structure, and attention refinement residual processing. In this study, we propose a novel solution by introducing the Multi-Scale Feature Extraction Network (MFENet) specifically tailored for image deblurring tasks. Our approach aims to enhance the accuracy of identifying and tracking dynamically moving wolfberry branches. MFENet achieves this by harnessing a combination of multi-scale image content features, incorporating modules for multi-scale depth-separable convolution, multi-scale receptive field structure, and attention refinement residual processing. Our findings underscore the efficacy of the proposed MFENet in addressing the challenge of blurred images resulting from the swinging motion of wolfberry branches. Furthermore, through experimental validation, we demonstrate that the improved SparseInst segmentation method achieves exceptional precision and faster segmentation speeds when applied to the branch dataset. This research not only provides a practical solution to mitigate branch swinging during wolfberry harvesting but also offers valuable theoretical and technical support to advance this field.

Fruit growing direction recognition and nesting grasping strategies for tomato harvesting robots

AbstractIn recent years, the potential of robotic harvesting in greenhouse tomato production has garnered significant attention within the tomato industry. However, there is a lack of sufficient research on the complete replacement of manual harvesting with this technology. In this paper, we propose a tomato harvesting method that utilizes a nesting approach to simplify the process and minimize damage. The paper describes the tomato harvesting robot prototype, the visual system equipped with three vision‐based tomato detectors: YOLOv5_CBAM, which incorporates a convolutional block attention module; YOLOv5_SE, enhanced with a squeeze‐and‐excitation block; and a standard YOLOv5s model. Additionally, a novel shear gripping method for fruit bunches is presented, utilizing a bottom‐up snapping technique during harvesting. Point cloud data are utilized to determine the position of the tomato's main stem and bunch. The paper includes field tests and experimental findings, which indicate that the YOLOv5_CBAM model achieves the highest precision (82.62%) and recall (82.57%), outperforming YOLOv5_SE and standard YOLOv5s. Field experiments demonstrate that the improved end‐effector and vision system have significantly enhanced the robot's performance, achieving a 57.5% harvesting success rate in just 14.9 s.

Review of current mechanical design in agricultural end effector

Agricultural robots have become increasingly crucial in precise agriculture. This paper reviews recent developments in end-effectors applied in the robotic harvesting of fruits and vegetables. The control and harvest function of the end-effector is the focus. Different structures are categorized based on their properties. Advantages and limitations in each category are introduced. For the collection method, the hard-grab method is very popularly for low difficulties and the soft one can protect the fruits adequately and can harvest different fruits of similar size; double-finger type is used widely with the advantages of low cost, while the multi-finger type has the much higher stability. The collection method is divided into the cutting and twisting methods. The former is applied for low difficulties but may spread diseases among plants; therefore, thermal cutting, namely using high temperature to separate stem and crop, and the twisting method is used. Overall, different types of end-effectors show various advantages. With consideration of the target crop physical properties, a proper decision for the harvester can be made.

Apple recognition and picking sequence planning for harvesting robot in a complex environment

In order to improve the efficiency of robots picking apples in challenging orchard environments, a method for precisely detecting apples and planning the picking sequence is proposed. Firstly, the EfficientFormer network serves as the foundation for YOLOV5, which uses the EF-YOLOV5s network to locate apples in difficult situations. Meanwhile, the soft non-maximum suppression algorithm is adopted to achieve accurate identification of overlapping apples. Secondly, the adjacently identified apples are automatically divided into different picking clusters by the improved density-based spatial clustering of applications with noise. Finally, the order of apple harvest is determined to guide the robot to complete the rapid picking, according to the weight of the Gauss distance weight combined with the significance level. In the experiment, the average precision of this method is 98.84%, which is 4.3% higher than that of YOLOV5s. Meanwhile, the average picking success rate and picking time are 94.8% and 2.86 seconds, respectively. Compared with sequential and random planning, the picking success rate of the proposed method is increased by 6.8% and 13.1%, respectively. The research proves that this method can accurately detect apples in complex environments and improve picking efficiency, which can provide technical support for harvesting robots.

Apple recognition and picking sequence planning for harvesting robot in the complex environment

In order to improve the efficiency of robots picking apples in challenging orchard environments, a method for precisely detecting apples and planning the picking sequence is proposed. Firstly, the EfficientFormer network serves as the foundation for YOLOV5, which uses the EF-YOLOV5s network to locate apples in difficult situations. Meanwhile, the Soft Non-Maximum Suppression (NMS) algorithm is adopted to achieve accurate identification of overlapping apples. Secondly, the adjacently identified apples are automatically divided into different picking clusters by the improved density-based spatial clustering of applications with noise (DBSCAN). Finally, the order of apple harvest is determined to guide the robot to complete the rapid picking, according to the weight of the Gauss distance weight combined with the significance level. In the experiment, the average precision of this method is 98.84%, which is 4.3% higher than that of YOLOV5s. Meanwhile, the average picking success rate and picking time are 94.8% and 2.86 seconds, respectively. Compared with sequential and random planning, the picking success rate of the proposed method is increased by 6.8% and 13.1%, respectively. The research proves that this method can accurately detect apples in complex environments and improve picking efficiency, which can provide technical support for harvesting robots.

Accurate detection and depth estimation of table grapes and peduncles for robot harvesting, combining monocular depth estimation and CNN methods

Precision agriculture is a growing field in the agricultural industry and it holds great potential in fruit and vegetable harvesting. In this work, we present a robust accurate method for the detection and localization of the peduncle of table grapes, with direct implementation in automatic grape harvesting with robots. The bunch and peduncle detection methods presented in this work rely on a combination of instance segmentation and monocular depth estimation using Convolutional Neural Networks (CNN). Regarding depth estimation, we propose a combination of different depth techniques that allow precise localization of the peduncle using traditional stereo cameras, even with the particular complexity of grape peduncles. The methods proposed in this work have been tested on the WGISD (Embrapa Wine Grape Instance Segmentation) dataset, improving the results of state-of-the-art techniques. Furthermore, within the context of the EU project CANOPIES, the methods have also been tested on a dataset of 1,326 RGB-D images of table grapes, recorded at the Corsira Agricultural Cooperative Society (Aprilia, Italy), using a Realsense D435i camera located at the arm of a CANOPIES two-manipulator robot developed in the project. The detection results on the WGISD dataset show that the use of RGB-D information (mAP=0.949) leads to superior performance compared to the use of RGB data alone (mAP=0.891). This trend is also evident in the CANOPIES Grape Bunch and Peduncle dataset, where the mAP for RGB-D images (mAP=0.767) outperforms that of RGB data (mAP=0.725). Regarding depth estimation, our method achieves a mean squared error of 2.66 cm within a distance of 1 m in the CANOPIES dataset.

Detection the maturity of multi-cultivar olive fruit in orchard environments based on Olive-EfficientDet

Olive (Olea europaea L.) fruit maturity detection in orchard environments plays a vital role in ensuring the quantity and quality of olive oil. However, it is difficult for olive fruit maturity detection in orchard environments due to the minimal phenotypic difference between neighboring maturity, and the occlusion and overlap of olive fruits. Most existing research pays more attention to olive fruit detection while ignoring olive fruit maturity detection problems. Therefore, a novel Olive-EfficientDet is proposed to detect the maturity of multi-cultivar olive fruits in orchard environments. In Olive-EfficientDet, the convolutional block attention module (CBAM) is rationally embedded into the backbone network for distinguishing different maturity-stage olive fruits with higher accuracy. The improved weighted bi-directional feature pyramid network (Bi-FPN) structure head network is constructed to focus on occlusion and overlap olive fruits, which can fully fuse semantic relationships and location information of different layers. The experimental results showed that the proposed Olive-EfficientDet provides an effective method for olive fruit maturity detection in orchard environments. The mean average precision (mAP) of fruit maturity detection reached 94.60%, 95.45%, 93.75%, and 96.05% for olives of cultivar ‘Frantoio’, ‘Ezhi 8′, ‘Leccino’, and ‘Picholine’; the mean detection time was 337 ms per image; and the model size was only 32.4 MB. In addition, the Olive-EfficientDet exhibits robust adaptability to complex illumination, occlusion, and overlap in uncontrolled and challenging orchard environments. Comparative experiments were conducted using Olive-EfficientDet and other state-of-the-art fruit maturity detection methods. The comparative experiment results showed that the mAP of the olive fruit maturity detection with four cultivars using Olive-EfficientDet was higher than that of SSD, EfficientDet, YOLOv3, and Faster RCNN. Especially, Olive-EfficientDet obtained the highest mAP for olive fruit maturity detection in orchard environments while maintaining an encouraging model size and speed, which can provide a technical foundation for fruit maturity detection in olive harvesting robots.

Strawberry ripeness classification method in facility environment based on red color ratio of fruit rind

Strawberry detection and ripeness classification are important concerns in robotic harvesting, as precondition for efficient and nondestructive picking. However, the different sizes and overlap of strawberries complicate detection, and research on strawberry ripeness classification is lacking. Therefore, this study proposes the red color ratio as a new parameter for strawberry ripeness quantification. This parameter considers the proportion of the strawberry red area as a key factor. First, an adaptive strawberry feature augmentation network (ASFA-net) is proposed to generate masks for the strawberries. ASFA-net uses the shifted window (Swin) Transformer as the backbone network for strawberry feature extraction and employs a feature pyramid network along with the proposed strawberry feature adaptive fusion module, to augment the features. The proposed decoupled head network is then applied to generate the final results. Second, the red region within the mask of ripe strawberries is segmented based on hue, saturation, and value (HSV) to calculate the proportion of the red area of individual strawberries. ASFA-net was verified on a home-made strawberry dataset. The results show that ASFA-net can detect strawberries accurately and efficiently with mean average precision of 95.91 ± 0.64 % and mean intersection over union of 90.15 ± 1.49 %. The strawberry ripeness classification method exhibited good performance with an accuracy of 95.94 ± 1.40 %, false positive rate of 3.09 ± 0.46 %, and false negative rate of 4.22 ± 1.68 %. The purpose of this study is to establish a method for simultaneous strawberry detection and ripeness classification in a facility environment, to provide a new reference method for the vision system of robots.

LACTA: A lightweight and accurate algorithm for cherry tomato detection in unstructured environments

Developing cherry tomato detection algorithms for selective harvesting robots faces many challenges due to the influence of various environmental factors such as lighting, water mist, overlap, and occlusion. To this end, we present LACTA, a lightweight and accurate cherry tomato detection algorithm specifically designed for harvesting robot operation in complex environments. Our approach enhances the model’s generalization ability and robustness by selectively expanding the original dataset using a combination of offline and online data augmentation strategies. To effectively capture the small target featuresof cherry tomatoes, we construct an adaptive feature extraction network (AFEN) that focuses on extracting pertinent feature information to enhance the identification ability. Additionally, the proposed cross-layer feature fusion network (CFFN) preservesthe model’s lightweight nature while obtaining richer feature representations. Finally, the integration of efficient decoupled heads (EDH) further enhances the model’s detection performance. Experimental results demonstrate the adaptability and robustness of LACTA, achieving precision, recall, and mAP values of 94 %, 92.5 %, and 97.3 %, respectively. Compared to the original dataset, the offline-online combined data augmentation strategy improves precision, recall, and mAP by 1.6 %, 1.7 %, and 1.1 %, respectively. The AFEN + CFFN network structure significantly reduces computational complexity by 28 % and number of parameters by 72 %. With a compact size of only 2.88 M, the LACTA model can be seamlessly deployed into selective harvesting robots for the automated harvesting of cherry tomatoes in greenhouses. The code is available at https://github.com/ruyounuo/LACTA.

Role of 5G Technology in Enhancing Agricultural Mechanization

Advanced mechanization strategies such as site specific crop management machinery, Unmanned Aerial Vehicles (UAVs), robotic harvesters relies greatly on the precision technologies and IoT. Present networks (3G/4G), WiFi are in limited availability and bandwidth that cause delays in data transfer (latency) and hinder effective data transfer. But high-speed communication system, reliable and secured data connection is required for precision agriculture, cyber physical management systems and use of IoT in farm operations. Loss of data connection between devices, sensors or data servers’ impacts the precision and effectiveness of technology applied. High data transfer volumes and low latency data exchange were the major advantages of 5G technology, facilitates its use in precision farming and IoT applications in a most efficient way. It also enables the autonomous and robotic machinery for various operations in the aspects of agriculture and livestock production. The major application of 5G in mechanization is control systems for robotic harvesters, drones, autonomous vehicles, surveillance etc. Remote locations such as hill farms has the inherent constraints such as remoteness, marginality, slope etc., remains underutilized, can be developed with advanced communication systems. Even though there are several constraints in adoption of 5G technology, such as shortage in electricity supply, lower range coverage, the technology can ensure raise in agricultural production through mechanization. The chapter discusses the advancements in application of 5G technology in agricultural mechanization.

Vision based fruit recognition and positioning technology for harvesting robots

Fruit harvesting robot operations in unstructured agricultural environments have been facing both accuracy and speed challenges, and vision-based control plays an important role in autonomous harvesting. Visual recognition and harvest positioning technology determine whether the target information can be obtained quickly and accurately and whether fruit picking can be carried out efficiently and orderly. It includes not only the hardware platform technology of the vision system, but also different sensor technologies and the optimization of recognition and positioning algorithms. This paper reviews the application of vision recognition and harvesting localization technologies in fruit harvesting robots, starting with an overview of hardware platform technologies for different vision systems of robots, followed by an analysis of the development of target recognition and positioning technologies in terms of sensor hardware and algorithm design, respectively. In addition, the review discusses the difficulties and challenges encountered by robots in harvesting operations and describes the future trends of vision localization technologies.

Optimizing cotton‐picking robotic manipulator and inverse kinematics modeling using evolutionary algorithm‐assisted artificial neural network

AbstractThis study presents a particle swarm optimization (PSO) algorithm‐assisted neural‐network‐based inverse kinematics solution for a 4‐DoF (degree‐of‐freedom) cotton harvesting robot. A novel setup was developed to measure the three‐dimensional locations of in‐field cotton bolls. Dimensional optimization of the manipulator was conducted using the PSO algorithm to minimize torque requirements at joints. With the optimized links’ lengths, the targeted end‐effector positions were achieved effectively (coefficient of determination (R2) > 99.88). The genetic algorithm optimized the neural network architecture to include three hidden layers with [64 64 32] neurons, identifying the Tanh activation function as the optimal configuration. A custom loss function was used during the training of artificial neural network (ANN). Using angles predicted by the trained ANN, the end‐effector reached targeted positions with positioning errors below 13.0 mm. A hybrid model consisting of an ANN and PSO algorithm was developed to further reduce the error. This trained hybrid model resulted in a positioning error below 1.0 mm with inference time of 6.07 s during simulation phase. As compared to the ANN and PSO algorithm, hybrid model reduced the positioning error and inference time (>40.0%), respectively. For hybrid model, the mean percentage errors of 0.25%, 0.39%, and 0.84% were observed along the x‐, y‐, and z‐axis. A positioning error below 9.0 mm occurred during evaluation of the hybrid model with the fabricated manipulator. Hence, the developed hybrid model precisely determines the joint angles, allowing the end‐effector of the cotton harvesting robot to reach at targeted pose with minimum error.

The Monash Apple Retrieving System: A review on system intelligence and apple harvesting performance

The wide-scale adoption of selective apple harvesting robots is yet to be seen, due to their strict requirement of well-manicured canopies to perform reliably. However, maintaining this standard of canopy is labour and resource-intensive, and an unsustainable practice for farmers in the long term who are already burdened with labour shortages. Without a complete planner and intelligent decision-making, a robotic harvesting platform would struggle to adapt to, and navigate in even typical canopy environments. In this work, we introduce the Monash Apple Retrieving System (MARS), a selective apple harvesting platform that is capable of navigating and harvesting apples in complex canopy environments without the need for canopy simplification. Utilising a four-tier planning process, MARS is able to generate gentle, collision-free harvesting trajectories in complex and unstructured canopies. Extensive trials were conducted over the 2021 and 2022 apple harvesting seasons in Australia, where MARS achieved a harvesting success rate of 62.8% at a cycle time of 9.18 s per apple, with negligible damage to the fruit. Performance increased to 70.8% success rate and 7.91 s of cycle time under ideal visibility conditions. Although performance can be further enhanced with canopy simplification and hardware upgrades, the reported results were achieved under realistic canopy conditions where very limited attempts were made to modify the canopy to create idealistic conditions for robotic harvesting. This baseline performance is promising for farmers who are wanting to utilise robotic harvesting, but do not have the resources to maintain well-manicured canopies. By lowering strict canopy requirements for robotic harvesting, we can improve accessibility of this technology to more farmers who are struggling with labour shortages around the world.

A Comprehensive Review of the Research of the “Eye–Brain–Hand” Harvesting System in Smart Agriculture

Smart agricultural harvesting robots’ vision recognition, control decision, and mechanical hand modules all resemble the human eye, brain, and hand, respectively. To enable automatic and precise picking of target fruits and vegetables, the system makes use of cutting-edge sensor technology, machine vision algorithms, and intelligent control and decision methods. This paper provides a comprehensive review of international research advancements in the “eye–brain–hand” harvesting systems within the context of smart agriculture, encompassing aspects of mechanical hand devices, visual recognition systems, and intelligent decision systems. Then, the key technologies used in the current research are reviewed, including image processing, object detection and tracking, machine learning, deep learning, etc. In addition, this paper explores the application of the system to different crops and environmental conditions and analyzes its advantages and challenges. Finally, the challenges and prospects for the research on picking robots in the future are presented, including further optimization of the algorithm and improvement of flexibility and reliability of mechanical devices. To sum up, the “eye–brain–hand” picking system in intelligent agriculture has great potential to improve the efficiency and quality of crop picking and reduce labor pressure, and it is expected to be widely used in agricultural production.

Application of improved YOLOv7-based sugarcane stem node recognition algorithm in complex environments.

Sugarcane stem node detection is one of the key functions of a small intelligent sugarcane harvesting robot, but the accuracy of sugarcane stem node detection is severely degraded in complex field environments when the sugarcane is in the shadow of confusing backgrounds and other objects. To address the problem of low accuracy of sugarcane arise node detection in complex environments, this paper proposes an improved sugarcane stem node detection model based on YOLOv7. First, the SimAM (A Simple Parameter-Free Attention Module for Convolutional Neural Networks) attention mechanism is added to solve the problem of feature loss due to the loss of image global context information in the convolution process, which improves the detection accuracy of the model in the case of image blurring; Second, the Deformable convolution Network is used to replace some of the traditional convolution layers in the original YOLOv7. Finally, a new bounding box regression loss function WIoU Loss is introduced to solve the problem of unbalanced sample quality, improve the model robustness and generalization ability, and accelerate the convergence speed of the network. The experimental results show that the mAP of the improved algorithm model is 94.53% and the F1 value is 92.41, which are 3.43% and 2.21 respectively compared with the YOLOv7 model, and compared with the mAP of the SOTA method which is 94.1%, an improvement of 0.43% is achieved, which effectively improves the detection performance of the target detection model. This study provides a theoretical basis and technical support for the development of a small intelligent sugarcane harvesting robot, and may also provide a reference for the detection of other types of crops in similar environments.

Tactile-Sensing Technologies: Trends, Challenges and Outlook in Agri-Food Manipulation.

Tactile sensing plays a pivotal role in achieving precise physical manipulation tasks and extracting vital physical features. This comprehensive review paper presents an in-depth overview of the growing research on tactile-sensing technologies, encompassing state-of-the-art techniques, future prospects, and current limitations. The paper focuses on tactile hardware, algorithmic complexities, and the distinct features offered by each sensor. This paper has a special emphasis on agri-food manipulation and relevant tactile-sensing technologies. It highlights key areas in agri-food manipulation, including robotic harvesting, food item manipulation, and feature evaluation, such as fruit ripeness assessment, along with the emerging field of kitchen robotics. Through this interdisciplinary exploration, we aim to inspire researchers, engineers, and practitioners to harness the power of tactile-sensing technology for transformative advancements in agri-food robotics. By providing a comprehensive understanding of the current landscape and future prospects, this review paper serves as a valuable resource for driving progress in the field of tactile sensing and its application in agri-food systems.

Object localization methodology in occluded agricultural environments through deep learning and active sensing

The predominance of branch and leaf shade in agricultural environments presents a barrier for accurate target recognition. Particularly for picking robots, precise localization of the picking object is essential. For this purpose, this paper proposes detection and localization methods based on deep learning and active sensing for harvesting robots in real-world environments with occlusion and varying lighting conditions. Using a deep learning network, the detection method firstly extracts the peduncle and fruit regions; the fruit region is then used to calculate the occlusion rate and the offset distance of the peduncle relative to the fruit. With such information, the robot arm adjusts the camera's field of view to perform multiple recognitions until the confidence is satisfied. Furthermore, to solve the picking problem caused by the peduncle's random tilting, this paper proposes a method to calculate the peduncle's tilt angle for controlling the end-effector to make the corresponding angle rotation. The robot arm and its end-effector are directed to complete the harvesting with the picking point location and tilt angle. In this study, data collection, detection and picking tests were implemented in the field, the results indicated that the method obtained an average successful picking rate of 90% after 300 trials, the error between the estimated occlusion ratio and the genuine value is 16% in average, and the active sensing method has improved the confidence score in occluded situations by over 50%. The proposed active methods have a 33% increase in precision and a 43% increase in efficiency compared to constant methods.

Retracted: Target Recognition and Trajectory Planning of Apple Harvesting Robot considering Color Multimedia Image Segmentation Algorithm

Retracted: Target Recognition and Trajectory Planning of Apple Harvesting Robot considering Color Multimedia Image Segmentation Algorithm

Human-centered approach for an efficient cucumber harvesting robot system: Harvest ordering, visual servoing, and end-effector

This paper presents a human-centered approach for an efficient cucumber-harvesting robot system. Specifically, harvest ordering, visual servoing, and end-effector-based manipulation functionalities were integrated to realize efficient and stable harvesting. The proposed approach involved determining the optimal harvest ordering, guiding the end-effector to the cucumber pedicel through visual servoing, and designing an end-effector to effectively harvest long cucumbers. The performance of the system was evaluated through preliminary and field experiments. The results of the preliminary experiments showed that harvest ordering decreased the harvesting time and travel length and increased the battery efficiency. The visual servoing was robust, and pedicels could be rapidly detected at a speed of 16–23 FPS through computer vision technologies. The pedicel could be accurately positioned within the cutting area of the end-effector. Furthermore, the proposed end-effector could effectively cut thin cucumber pedicels (3–6 mm), with a 100.0% success rate. Field experiments were conducted at three sites in Korea: Green Monsters, Sangju smart-valley, and Fresh-farm. The harvest success rate at the three sites ranged from 50.9% to 60.0%, with an overall value of 56.6%. The overall average harvest time of 56.0 s. The positional accuracy of the system was within the optimal range of 0–30 mm. Furthermore, the primary causes of harvest failure were analyzed, and future research directions to improve the performance of harvest robots were discussed.

Instance Segmentation of Lotus Pods and Stalks in Unstructured Planting Environment Based on Improved YOLOv5

Accurate segmentation of lotus pods and stalks with pose variability is a prerequisite for realizing the robotic harvesting of lotus pods. However, the complex growth environment of lotus pods causes great difficulties in conducting the above task. In this study, an instance segmentation model, LPSS-YOLOv5, for lotus pods and stalks based on the latest YOLOv5 v7.0 instance segmentation model was proposed. The CBAM attention mechanism was integrated into the network to improve the model’s feature extraction ability. The scale distribution of the multi-scale feature layer was adjusted, a 160 × 160 small-scale detection layer was added, and the original 20 × 20 large-scale detection layer was removed, which improved the model’s segmentation accuracy for small-scale lotus stalks and reduced the model size. On the medium-large scale test set, LPSS-YOLOv5 achieved a mask mAP0.5 of 99.3% for all classes. On the small-scale test set, the mAP0.5 for all classes and AP0.5 for stalks were 88.8% and 83.3%, which were 2.6% and 5.0% higher than the baseline, respectively. Compared with the mainstream Mask R-CNN and YOLACT models, LPSS-YOLOv5 showed a much higher segmentation accuracy, speed, and smaller size. The 2D and 3D localization tests verified that LPSS-YOLOv5 could effectively support the picking point localization and the pod–stalk affiliation confirmation.