Table of Content

30 September 2026, Volume 8 Issue 5

Previous Issue   

Topic--Frontier Technology Innovation and Industrial Applications of Agricultural Robots

Autonomous Operation Technologies and Development Trends for Agricultural Robots in Complex Scenarios |

YANG Shuofei, WANG Zhe, YANG Shenghui, MIAO Zhonghua, CHEN Du, ZHENG Yongjun, SONG Zhenghe, LI Hongwen

2026, 8(5):  1-22.  doi:10.12133/j.smartag.SA202609040

Asbtract ( 54 )   HTML ( 11)   PDF (5157KB) ( 13 )  

Figures and Tables | References | Related Articles | Metrics

[Significance] Agricultural robots are increasingly expected to perform precise and timely operations under labor constraints while maintaining agronomic quality. Their operating conditions, however, differ substantially across open fields, orchards, greenhouses, and livestock facilities. Advances in a single algorithm or mechanism therefore do not necessarily translate into reliable operation of an integrated robot. This review examines how task requirements and environmental constraints shape the selection and integration of autonomous operation technologies. Particular attention is given to the relationships between component performance, complete task execution, and the interpretation of experimental evidence, with the aim of supporting equipment design and application decisions. [Progress] The review first distinguishes operation tasks from application environments. Seeding and transplanting, crop protection and weeding, selective harvesting, inspection and monitoring, and livestock operations require different outputs and evaluation criteria, whereas greenhouse production is treated as an environment that can accommodate several of these tasks. Representative research platforms and commercial systems illustrate alternative configurations, while product descriptions are distinguished from experimentally demonstrated performance. Wheeled, tracked, and legged chassis are compared in terms of steering geometry, terrain interaction, payload, maneuvering space, and maintenance demands. Modular configurations can improve adaptability, but changes in structure and load also require corresponding updates to calibration and control parameters. Lightweight design is considered together with traction, stiffness, stability, and tool requirements. Perception methods are examined according to the information needed for navigation and operation. Color and geometric processing, conventional machine learning, deep object detection, and semantic or instance segmentation offer different trade-offs in annotation requirements, computational demand, and sensitivity to changing conditions. Three-dimensional sensing must connect target observations to the robot and tool coordinate systems. Multimodal fusion can combine semantic, geometric, motion, and contact information, provided that calibration, synchronization, and degraded measurements are handled explicitly. Visual foundation models and domain adaptation offer opportunities to reduce labeling effort, but cross-field generalization and task-relevant errors remain essential considerations. Reduced supervision does not by itself establish readiness for crop-contact operation. Navigation, planning, and control are organized into five connected topics: multi-sensor localization and mapping, global and local path planning, coverage planning, manipulator planning and visual servoing, and autonomous control and decision-making. Absolute positioning, relative motion estimation, and path tracking are treated as distinct quantities. The discussion links global navigation satellite system (GNSS) availability, crop-row continuity, repetitive structures, and sensor limitations to localization choices. Path planning must account for vehicle kinematics, implement width, crop exclusion zones, headland access, and the consequences of obstacle avoidance for unfinished work. Manipulator operation further requires coordination of visibility, reachability, collision avoidance, and contact feedback. Research examples show why favorable results under continuous rows or simplified crop conditions require careful interpretation before application to more difficult environments. Execution technologies are analyzed through manipulator configuration, grasping and detachment, precision application, and coordination of contact and material handling. Representative experiments show that reported outcomes depend on crop arrangement, occlusion, positioning infrastructure, and treatment schedules. Accordingly, cross-task numerical rankings are avoided. System evaluation is developed from component testing through integrated closed-loop validation to trials in production environments. Reporting should identify experimental units, target populations, success rate denominators, variability, failure handling, and the time included in productivity estimates. Full operation cycles may include platform motion, retries, unloading, replenishment, and human assistance. Livestock operations require task-specific criteria rather than direct adoption of field-area productivity measures. [Conclusions and Prospects] Sustained agricultural autonomy requires technologies that fit agronomic conditions and work effectively together. Their combined performance needs to be assessed by the work completed to an acceptable standard. Priorities include reliable perception under changing conditions, accurate spatial alignment and time synchronization, feedback on operation quality, maintainable specialized components, and clear evaluation of reliability and cost. Foundation models, vision-language-action methods, synthetic data, and multi-robot collaboration offer further possibilities. Their practical value, however, requires independent testing of task completion, interaction with crops or animals, and recovery from failures. Economic assessments should state their scope and account for equipment utilization, human supervision, maintenance, and downtime. Trials at different locations and in different seasons, with comparable agronomic requirements and clearly recorded human involvement, would help establish whether research gains translate into dependable agricultural operation.

Key Technologies and Research Prospects for Multi-Machine Collaboration of Intelligent Agricultural Machinery |

XIA Xianfei, YAN Haowen, CHEN Yong, XIAO Maohua, CHEN Wei, GUO Peiqi, ZHANG Jinlong, YUE Xinyuan

2026, 8(5):  23-42.  doi:10.12133/j.smartag.SA202608067

Asbtract ( 79 )   HTML ( 10)   PDF (4482KB) ( 23 )  

Figures and Tables | References | Related Articles | Metrics

[Significance] Multi-machine collaboration of intelligent agricultural machinery refers to a coordinated field-operation mode in which multiple agricultural machines jointly perform agricultural production tasks through task allocation, information sharing, and cooperative control. It involves multiple technical aspects, including system architecture, sensing and localization, communication, decision-making and planning, dynamic interaction, and cooperative control. The technical boundaries between multi-machine collaboration, single-machine autonomous operation, and coordination within tractor-implement units must also be clearly delineated. This review summarizes the research progress, key technologies, and typical applications of multi-machine collaborative operation for intelligent agricultural machinery, aiming to provide a reference for subsequent research and engineering implementation. [Progress] Multi-machine collaboration is defined as an agricultural operation system comprising two or more machines with relatively independent mobility or operating capabilities that share inter-machine information and establish coupling in task execution, path planning, motion control, or operation timing. Single-machine autonomous navigation and internal coordination within tractor-implement units are regarded only as supporting technologies for multi-machine collaboration. This review first analyzes multi-machine system architectures and compares the characteristics and applicable conditions of centralized, distributed, and hybrid architectures. Centralized architectures emphasize global information aggregation, unified task allocation, and coordinated path planning, whereas distributed architectures rely more on local sensing, peer-to-peer information exchange, and decentralized decision-making. Hybrid architectures combine global coordination with local autonomy and are more suitable for complex operating conditions that require both overall scheduling and rapid local responses. The key enabling technologies are then reviewed. In multi-source perception and cooperative localization, research has gradually expanded from single-machine multi-sensor fusion to inter-machine relative localization, state sharing, cooperative perception, and shared map construction. In communication, the controller area network (CAN) and the ISO 11783-based agricultural machinery communication protocol (ISOBUS) provide a foundation for intra-machine data exchange, while Wi-Fi, LoRa, 4G/5G, and radio links are commonly used for inter-machine communication. In decision-making and planning, research has progressed from static task allocation and independent path generation toward dynamic task reallocation, task-path joint optimization, spatiotemporal conflict resolution, and global-local collaborative planning. In dynamic modeling and interaction, wheel-soil slip, sideslip, tractor-implement articulation, trajectory deviation, and relative motion among machines are increasingly incorporated into planning and control loops to improve trajectory feasibility and dynamic coordination. In formation control, virtual-structure, behavior-based, and leader-follower approaches have been widely investigated, with growing attention to dynamic obstacle avoidance, disturbance rejection, communication constraints, and safety control. Typical application scenarios are further summarized from the perspectives of homogeneous and heterogeneous systems. Homogeneous systems primarily improve parallel operational capacity through regional decomposition, task allocation, and coordinated coverage, whereas heterogeneous systems emphasize functional complementarity among machines with different sensing, operating, and transport capabilities. Representative applications include multi-machine harvesting, harvester-transport coordination, seeding, spraying, air-ground cooperation, and harvesting-transfer operations. These comparisons reveal differences in coordination requirements, information flows, and control relationships across representative agricultural operation scenarios. [Conclusions and Prospects] Several challenges remain, including perception degradation in complex field environments, communication delays and interruptions, asynchronous multi-source information, task and path conflicts, dynamic disturbances, equipment failures, insufficient interoperability, and limited large-scale field validation. Future research should focus on reconfigurable collaborative architectures, heterogeneous communication networks, multi-machine cooperative perception and shared mapping, dynamic task and path replanning, fault-tolerant formation control, embodied cooperative control, digital twins, and integrated simulation-field evaluation. Long-term, cross-field, and cross-season experiments are also needed to verify system reliability, scalability, safety, and engineering applicability, thereby supporting the transition of intelligent agricultural machinery from single-machine autonomy to continuous, robust, and large-scale multi-machine collaboration.

Research Status and Prospects of Agricultural Machinery Chassis Leveling Technology |

YANG Wencai, CHEN Hongkun, ZHAO Hengliang, ZHU Longtu, SHANG Xiaobiao, ZHANG Hongyang, ZHEN Guangqi

2026, 8(5):  43-63.  doi:10.12133/j.smartag.SA202512012

Asbtract ( 459 )   HTML ( 17)   PDF (2873KB) ( 31 )  

Figures and Tables | References | Related Articles | Metrics

[Significance] Hilly and mountainous areas are characterized by complex topography, rugged roads, steep and narrow slopes, and fragmented, discontinuous farmland plots, making it difficult for traditional agricultural machinery to adapt to such conditions. These terrain-related limitations have become a major bottleneck restricting the development of agricultural mechanization and modernization in such areas. Therefore, the research, development, and promotion of agricultural machinery and equipment suitable for hilly and mountainous areas are urgently needed. As a key technology for improving the terrain adaptability and operation quality of agricultural machinery, chassis leveling has become a research focus in the field of agricultural machinery equipment. [Progress] A comprehensive review of research on evaluation indicators, leveling devices, and control systems for agricultural machinery chassis leveling is presented. This review focuses on summarizing the categories and applicable scopes of static and dynamic evaluation indicators and on analyzing the principles, performance characteristics, and applicable scenarios of two types of leveling mechanisms, namely, three-point support and four-point support. In this review, multisource information fusion schemes, such as attitude sensing, displacement measurement, load monitoring and terrain perception are summarized; the technical characteristics of chassis leveling control strategies, including the angle error control method and the position error control method, are expounded; and the evolutionary path of intelligent algorithms from traditional proportional-integral-derivative (PID) control to fuzzy control, sliding mode control, and neural networks is reviewed. Evaluation indicators, such as roll angle, pitch angle, leveling speed, mean tilt angle, and standard deviation of tilt angle, can provide a relatively comprehensive assessment of the operational performance of chassis leveling systems. Representative three-point support leveling devices include mechanisms based on lateral leveling, hydraulic differential-height adjustment, double guide columns, double parallel four-bar linkages, left-right eccentric wheel swing, and left-right transmission rocker arms. Representative four-point support leveling devices include adaptive balanced rocker suspension, four-point guide-column lifting mechanisms, Y-shaped adjustable suspension, four-hydraulic-cylinder leveling structures, hinged multi-link mechanisms, and articulated three-layer frames. These devices can determine the required compensation according to the inclination state of the chassis and then achieve body attitude adjustment through mechanical structures, hydraulic systems, and other actuating devices. By dynamically adjusting the vehicle body and operating components, they can effectively adjust the center of gravity of agricultural machinery and improve operational stability on sloping terrain. In addition, when integrated with high-precision sensors, intelligent control strategies, and efficient control algorithms, chassis leveling systems can achieve adaptive leveling under complex terrain conditions involving steep and narrow slopes. [Conclusions and Prospects] Existing chassis leveling technologies for agricultural machinery can basically satisfy the leveling requirements of large-scale machinery operating on broad and relatively gentle farmland and have already seen preliminary application. However, in the complex terrain of hilly and mountainous areas, the following limitations still exist: Chassis leveling models and performance evaluation standards have not yet been fully established; the reliability and general applicability of existing leveling devices remain insufficient; the degree of integration of chassis leveling control systems is still limited, and the intelligence and adaptability of control algorithms require further improvement; the false leg phenomenon remains difficult to identify, and engineering verification, practical promotion, and field application remain inadequate. In view of the current research status, existing problems, and future development demands, this paper proposes several priority directions for future study aimed at improving the adaptability of agricultural machinery to hilly and mountainous terrain. These include constructing leveling mechanisms and evaluation systems closely integrated with agronomic practices; developing highly reliable and universal chassis leveling devices; developing low-cost engineering verification and leveling technologies capable of suppressing the false-leg phenomenon. This review is expected to provide reference data for the research and development of advanced agricultural machinery equipment applicable to hilly and mountainous areas of China.

Agricultural Digital Twin: Technical System, Application Progress, and Engineering Implementation |

ZHU Qingzhen, DAI Xu, SUN Baoyi, ZHANG Liyuan, ZHU Yanqiu

2026, 8(5):  64-77.  doi:10.12133/j.smartag.SA202608027

Asbtract ( 22 )   HTML ( 1)   PDF (1926KB) ( 5 )  

Figures and Tables | References | Related Articles | Metrics

[Significance] Agricultural digital twin research spans digital representation, state monitoring, prediction, decision support, and physical execution. These functions correspond to different system capabilities and verification requirements rather than to a fixed developmental sequence. Agricultural objects differ from industrial equipment: crops and animals develop continuously with individual variation, production depends on changing environmental and management conditions, and machinery and robots operate under terrain, communication, and safety constraints. These differences affect what can be observed, how models are updated, and how digital outputs enter production. This review examines the technical system, representative applications, and engineering conditions of agricultural digital twins, focusing on the transition from research prototypes to sustained operation. [Progress] The review draws primarily on the Web of Science Core Collection, using structured information extraction across research objects, states and data, models and tasks, execution interfaces, result feedback, and validation conditions. Digital models, digital shadows, state synchronization, decision support, and physical execution are distinguished by the relationships actually reported in the literature, rather than treated as levels of a fixed maturity ranking; missing descriptions are treated as limits of available evidence rather than as proof that a function does not exist. The technical discussion traces the pathway from sensing and digital representation to model construction, prediction, decision-making, and field execution, and compares process-based, agent-based, and data-driven modeling approaches together with their coupling mechanisms. Model accuracy under one dataset or environment does not establish validity under a different season, location, cultivar, or machine configuration. Four scenario groups—protected agriculture and plant factories, crop phenotyping and growth simulation, livestock and aquaculture, and agricultural machinery and robots—illustrate these differences. In protected agriculture and plant factories, centralized sensing and available environmental control equipment support microclimate monitoring, spatial representation, and increasingly prediction-based control, though reported evidence rarely extends to automatic execution across full production cycles. In crop phenotyping and growth simulation, structural reconstruction and process-based models support growth prediction and cross-location transfer, but such studies typically report model performance rather than field execution or management records. In livestock and aquaculture, digital twins mainly support behavior recognition, state prediction, and structural monitoring; evidence linking these outputs to feeding, environmental control, or other field operations remains limited, and individual differences and long production cycles complicate validation. In agricultural machinery and robotics, digital twins are used for virtual testing and simulation-to-real transfer, with some studies reporting field sowing and harvesting tasks; communication delays, safety, and recovery under continuous operation are less consistently reported than single-trial performance. Across all four groups, an operational loop requires evidence linking a model or decision output to a trigger, an executing actor, an actual action, and a returned result used in subsequent updates; an architecture diagram or control-related terminology alone does not establish this sequence. [Conclusions and Prospects] Existing strengths concentrate in sensing, digital representation, modeling, and prediction; evidence for sustained operation, execution feedback, maintenance, and cross-scenario reuse is reported less consistently. Future work should document object identification, calibration, and model versioning over time; trace execution permissions, commands, and feedback when outputs enter production; and record failures, interventions, and recalibration over extended validation periods. Transfer studies should specify which data, model components, and interfaces can be reused and which require reconfiguration, and comparisons require consistent definitions of tasks, events, and metrics. As this review is limited by its search scope, these priorities indicate directions for further verification rather than a complete account of deployed systems.

Research Progress on Key Technologies and Intelligent Development Trends of Orchard Trenching and Fertilization Equipment |

WEI Daochu, WANG Zhichong, LI Xuecheng, WANG Jingwei, ZHOU Mengchuang, ZHAI Changyuan, ZOU Wei

2026, 8(5):  78-97.  doi:10.12133/j.smartag.SA202604028

Asbtract ( 321 )   HTML ( 3)   PDF (3879KB) ( 19 )  

Figures and Tables | References | Related Articles | Metrics

[Significance] Orchard fertilization is an important field management practice affecting nutrient supply, yield formation, and fruit quality. Ditching combined with fertilization can place fertilizer within the main root distribution zone and is more conducive to improving nutrient use efficiency than surface broadcasting. However, its operational performance is jointly constrained by orchard inter-row spacing, tree structure, terrain conditions, root distribution, and fertilizer properties. With the development of orchard production toward high-density planting, standardized tree forms, mechanization in hilly and mountainous areas, and intelligent management, the limitations of conventional ditching and fertilization equipment have become increasingly prominent in terms of furrow-depth stability, fertilizer placement consistency, fertilizer discharge uniformity, adaptability to complex terrain, and intelligent control capability. This review systematically summarizes the research progress of orchard ditching and fertilization equipment from four aspects: agronomic requirements, key components, whole-machine equipment, and intelligent technologies, to identify existing technical bottlenecks and development directions and to provide a reference for the optimized design and intelligent upgrading of precision orchard fertilization equipment. [Progress] First, the main fertilization methods, including ring-furrow fertilization, strip-furrow fertilization, radial-furrow fertilization, hole fertilization, fertigation, and surface broadcasting, are summarized, and their differences in fertilizer use efficiency, root injury risk, and adaptability to mechanized operation are analyzed. Combined with typical orchard patterns, including dwarf high-density orchards, traditional large-canopy orchards, hilly orchards, and standardized orchards, the requirements for furrow depth, fertilizer placement, fertilizer application rate, and equipment configuration imposed by inter-row spacing, root distribution, and terrain conditions are discussed. Second, the structural characteristics and applicable conditions of ditching mechanisms, including disc-type, auger-type, chain-type, and rotary-blade-type mechanisms, are summarized. Disc-type mechanisms are suitable for shallow-to-medium furrows and narrow-row operations. Auger-type and chain-type mechanisms have stronger deep-ditching capability, but require higher power requirements and maintenance demands. Rotary-blade-type mechanisms have good soil crushing and mixing performance, but soil disturbance and root injury risks should be controlled. In terms of fertilizer metering and conveying mechanisms, star-wheel, screw, centrifugal, scraper-conveyor, and combined fertilizer discharge systems show differing adaptability to different fertilizer forms. Granular fertilizers require greater metering stability, whereas high-moisture organic fertilizers and organic-inorganic mixed fertilizers rely more on anti-blocking, anti-bridging, and forced-conveying capacity. Third, at the whole-machine level, orchard ditching and fertilization machinery has gradually developed from conventional trailed and mounted implements to crawler self-propelled machines, remote-controlled machines, multifunctional combined-operation machines, and variable-rate fertilization equipment. Among them, trailed and mounted implements are suitable for orchards with flat terrain and wide row spacing; crawler-type self-propelled and remote-controlled equipment is more suitable for hilly, mountainous, and complex field plots; and multifunctional combined-operation and variable-rate fertilization equipment can help improve operational integration and fertilization accuracy. Finally, in terms of intelligent technologies, existing studies have gradually introduced information acquisition methods, such as mechanical displacement measurement, attitude sensing, ultrasonic sensing, laser ranging, and light detection and ranging (LiDAR). When combined with hydraulic or electric actuators, these methods have been used for furrow-depth adjustment, fertilizer-rate control, and operation-state monitoring, promoting the transformation of equipment from simple purely mechanical execution to closed-loop control involving sensing, decision-making, actuation, and feedback. [Conclusions and Prospects] Overall, orchard ditching and fertilization equipment still faces several challenges, including difficulty in controlling furrow depth under complex terrain conditions, difficulty in determining fertilizer placement due to non-visible root distribution, strong coupling among parameters such as furrow depth and fertilizer application rate, and pronounced trade-offs between the cost and reliability of intelligent systems. Future research should focus on multi-source information fusion-based perception, coordinated multi-parameter control, modular and standardized design, and green and efficient operation technologies. These developments are expected to promote the upgrading of orchard ditching and fertilization equipment from single mechanical implements to an intelligent, precision-oriented equipment system for smart orchards.

Design and Experiment of a Dual-Screw Propulsion Sliding-Cutting End Effector for Tomato Pruning Robots |

ZHU Mingyang, HE Jinli, MA Zenghong, DU Xiaoqiang

2026, 8(5):  98-109.  doi:10.12133/j.smartag.SA202606025

Asbtract ( 14 )   HTML ( 2)   PDF (2276KB) ( 4 )  

Figures and Tables | References | Related Articles | Metrics

Objective Greenhouse tomato pruning is still done largely by hand, and the dense canopy makes robotic operation difficult. Conventional shearing-type end effectors must be positioned accurately at the lateral-shoot base, and their jaws need considerable lateral room to open; both requirements restrict their use among leaves, inflorescences and fruits. Tomato lateral shoots also have high moisture content and low flexural rigidity, so they tend to bend, slip or tear during cutting, which degrades cut quality and increases infection risk. A dual-screw propulsion and sliding-cut end effector that combines shoot guidance, axial propulsion and sliding-cut removal in one compact structure is proposed to ease the positioning-accuracy requirement, reduce the required operating space, and stabilize the feeding and cutting of flexible shoots. Methods Tomato plants grown under a wide-furrow and narrow-bed system were characterized for morphological parameters (main-stem and lateral-shoot diameters, branching angle and shoot height) and stem mechanical properties, including Poisson's ratio measured in compression, elastic modulus measured by three-point bending, and the friction coefficient between the stem and a silicone finger, to guide the structural design. Force analysis showed that a double-edged blade provided a lower theoretical cutting force and a more balanced lateral force than a single-edged blade, with less eccentric loading and lateral deflection of the flexible shoot; therefore, a double-edged blade was selected. A sliding-cut mode was adopted to reduce the effective edge angle and instantaneous cutting resistance. Using fresh lateral shoots of 11 mm in diameter, a three-factor, three-level Box-Behnken experiment (blade thickness, cutting speed and sliding-cut angle) was conducted to build quadratic response-surface models for maximum cutting force and cutting time. Screw-propulsion and cutting mechanics models were then established to determine the screw geometry and motor torque, accounting for the effective propulsion coefficient and possible slippage. A prototype was built with 3D-printed AlSi10Mg screws and Cr12MoV blades. Single-arm pruning trials (n=50) and comparative geometric capture-tolerance tests (dual-screw guided, dual-screw unguided and conventional scissor-type, at lateral offsets from -10 to 10 mm) were carried out in a commercial greenhouse. Results and Discussions Both response-surface models were significant (P<0.000 1), with non-significant lack-of-fit and adjusted R² values above 0.96. Among the three factors, blade thickness had the strongest linear effect on maximum cutting force, followed by sliding-cut angle and cutting speed, whereas cutting speed dominated cutting time; the interaction between cutting speed and sliding-cut angle was significant for maximum cutting force. The response surface indicated that the lowest cutting-force region lay near a sliding-cut angle of 42°. Optimization gave a blade thickness of 0.30 mm, a cutting speed of 6.7 mm/s and a sliding-cut angle of 42°; validation showed relative errors below 10% for both indicators. In the 50 pruning trials, shoots of 10-12 mm in diameter were cut with 100.0% success rate and 100.0% smooth cuts rate (mean effective pruning time of 3.1 s), whereas the 12-14 mm group reached 95.5% and 86.4%, with a mean effective pruning time of 3.6 s. Overall, the success rate was 98.0%, the smooth-cut rate 94.0% and the mean effective pruning time 3.32 s. In the capture-tolerance tests on 11-mm-diameter specimens, the guided mechanism reached 90% capture success at an offset of ±10 mm, compared with 85% for the unguided version and 65% for the scissor type, indicating that the guide structure compensates for positioning error in a dense canopy. Conclusions The end effector integrates guidance, propulsion and cutting of tomato lateral shoots within a compact working envelope. It handled shoots of 10-14 mm diameter and tolerated geometric positioning error better than conventional shearing mechanisms, providing a reference for end-effector design in greenhouse tomato pruning robots. Future work should examine robustness, screw-spacing adjustability, and adaptive control to improve cut quality on larger shoots.

Visual Localization of Kiwifruit Flower Centers and Intra-Cluster Sweeping Operations: Methods and Experiments |

YANG Zhe, WANG Zimeng, LIU Han, ZHOU Guan, LIU Xingxing

2026, 8(5):  110-121.  doi:10.12133/j.smartag.SA202608007

Asbtract ( 55 )   HTML ( 4)   PDF (10542KB) ( 8 )  

Figures and Tables | References | Related Articles | Metrics

[Objective] To address the challenges posed by the concentrated flowering period of kiwifruit, the relatively low efficiency and inconsistent quality of manual supplementary pollination, and the frequent approach and retraction motions within dense flower distributions in trellis canopies during flower-by-flower robotic operations, a method integrating visual localization of kiwifruit flower centers, three-dimensional (3D) target generation, and intra-cluster sweeping was proposed for continuous pollination operations. The visual perception performance of the proposed method and feasibility of robotic operation were evaluated. [Methods] An RGB-D camera was used to synchronously acquire color images and aligned depth images of the trellis canopy, and a kiwifruit flower detection and flower-center keypoint dataset containing 1 001 images and 11 794 flower targets was constructed. The flower center was defined as the geometric center of the central stigma cluster of a female flower in its two-dimensional image projection. Three visibility levels were defined according to whether the flower-center was clearly visible, occluded but still locatable, and unable to be reliably annotated. With YOLO26n-Pose as the baseline, the original pose-estimation keypoint detection head was replaced by a flower-center offset prediction head that estimates the flower-center position relative to the center of the detected flower bounding box. Center auxiliary supervision, anisotropic Gaussian probabilistic flower-center localization, and a normalized Wasserstein distance (NWD) bounding-box loss were introduced, while the object keypoint similarity (OKS)-based keypoint training loss was excluded from the final model. After the model output flower bounding boxes, flower-center coordinates, and confidence scores, invalid depth values were removed within a local depth window around each flower-center, and the median of the valid depth values was used as the target depth. A depth reliability score was further calculated based on the valid-depth ratio, local depth standard deviation, and center-depth consistency. Flower-center pixels satisfying the quality criteria were back-projected using the camera intrinsic parameters and transformed using the hand-eye calibration extrinsic parameters to generate 3D operation targets in the robot base coordinate system. The system combined cross-frame target identification, target-quality filtering, and state management to prevent completed targets from being repeatedly issued. Spatially adjacent targets were then clustered and ordered using a nearest-neighbor strategy, and centripetal Catmull-Rom splines were employed to generate continuous intra-cluster sweeping paths. The proposed method was evaluated through visual-model ablation and comparison experiments, comparative indoor experiments using artificial flower clusters, and system experiments conducted at four orchard operating positions. [Results and Discussions] The final model achieved bounding-box mean average precision (mAP) values of 84.82% for mAP₅₀ and 54.00% for mAP_50:95, respectively, and keypoint mAP₅₀ and mAP_50:95 values of 88.13% and 87.78%, respectively. The miss rate (Miss@0.5) was 22.94%, and the mean pixel error of flower-center localization for successfully matched targets was 2.56 px. Compared with the YOLO26n-Pose baseline, Miss@0.5 was reduced by 6.71 percentage points, while the mean flower-center localization error remained at 2.56 px. In the comparative indoor experiments using artificial flower clusters, each operation mode included a total of 135 targets. Compared with flower-by-flower operation, the intra-cluster sweeping mode reduced the total robotic end-effector path length from 36.68 m to 19.74 m, and decreased the average operation time per flower from 9.68 s to 5.52 s. In the orchard experiments, 92 targets were distributed across four operating positions. The motion execution success rate, target operation completion rate, and average operation time per flower were 95.80%, 84.78%, and 6.82 s, respectively. Conclusions In the typical orchard scenarios evaluated in this study, the proposed method achieved a continuous closed-loop operation workflow encompassing flower detection, flower-center localization, 3D operation target generation, target scheduling, and robotic manipulator execution. Under the tested indoor conditions with fixed artificial flower clusters, the intra-cluster sweeping strategy reduced the end-effector path length and average operation time per flower while maintaining or slightly improving the target operation completion rate. These results provide a technical basis for visual perception and operation planning in intelligent kiwifruit pollination equipment.

Dexterous Obstacle Avoidance Method for Rigid and Flexible Obstacles of Apple Picking Robots |

HAO Jingtao, YUE Youjun, ZHAO Zhuoqun, ZHAO Hui, LI Tao

2026, 8(5):  122-134.  doi:10.12133/j.smartag.SA202510012

Asbtract ( 258 )   HTML ( 14)   PDF (18879KB) ( 26 )  

Figures and Tables | References | Related Articles | Metrics

Objective China is one of the world's largest apple producers. However, labor costs account for a high proportion of total harvesting expenses, and obstacle avoidance path planning of robotic arms remains a core bottleneck restricting the replacement of manual labor with robots. In orchard environments, rigid obstacles such as traction wires and thick branches are densely intertwined with flexible obstacles such as thin twigs and leaves. Traditional path planning algorithms rely on static geometric modeling, making them difficult to adapt to dynamic geometric boundaries and leading to long planning times and poor adaptability. Although deep reinforcement learning methods have been applied, they still suffer from limited generalization and low training efficiency when rigid and flexible obstacles are handled differently. To address the path planning challenge caused by rigid-flexible hybrid obstacles in unstructured orchards, an obstacle avoidance method based on expert-demonstration-guided deep reinforcement learning is proposed, providing theoretical and technical support for the intelligent harvesting robot operation. Methods First, a high-fidelity simulation environment was constructed based on the MuJoCo physics engine, which modeled a 4-degree-of-freedom robotic arm, a three-fingered gripper, and rigid-flexible hybrid obstacles. The combination of rigid tree trunks and flexible branches modeled using rope-based models, together with the precise configuration of elasticity and damping parameters, enabled realistic physical deformation to be reproduced. The designed 425-dimensional state vector covered key data such as the end-effector position of the robotic arm, fruit position, and attributes of nearly 50 obstacle segments. The action space was discretized into 8 actions, including gripper opening/closing and six-direction displacement of the end-effector with a step size of 0.01 m. The harvesting task was decomposed into three phases: approaching, grasping, and placing. The simulation environment could automatically complete phase switching and task judgment and distinguish rigid and flexible obstacles using labels—brown for rigid branches and green for flexible branches. Collision detection was incorporated to monitor interference between the gripper and obstacles, providing realistic scenario support for algorithm training. In terms of algorithm design, an expert-demonstration-guided soft actor-critic (EG-SAC) algorithm was proposed. A hierarchical reward function was designed that assigned rewards according to the distance variation between the end-effector and the target point in each phase, and imposed penalties for collisions, timeouts, and other abnormal situations. A total of 600 expert trajectories were generated by analytical calculation. For flexible occlusion scenarios, branch-poking guide points were designed to guide the robotic arm to approach the target accurately. The policy network adopted a three-layer fully connected neural network. After pre-training via behavioral cloning, it was optimized using the hindsight experience replay (HER)-SAC algorithm. Kullback-Leibler (KL) divergence regularization was introduced to constrain policy deviation, thereby balancing the exploitation of expert experience and the optimization potential of reinforcement learning. Orchard field tests were conducted using the self-developed multi-arm harvesting robot developed by the research team. The test site was a standard high-density dwarf orchard. The vision system acquired image and depth information via an RGB-D sensor, detected fruits using YOLOv8 (You Only Look Once), and generated 3D point clouds. Three typical scenarios were selected, with 20 fruit samples in each scenario. Scenario 1: simple flexible occlusion, Scenario 2: rigid-flexible hybrid occlusion, and Scenario 3: rigid occlusion. The initial poses of the robotic arm were kept consistent within the same round of tests and randomized across different rounds to verify generalization ability. Results and Discussions In simulation experiments, a system comprising a robot mobile platform, a multi-degree-of-freedom robotic arm, and a rigid-flexible hybrid fruit tree model was built based on MuJoCo. This system could randomly generate branch poses and fruit positions and synchronously randomize the initial poses of the robotic arm to ensure training-sample diversity. Comparative experiments were set up with an EG-SAC experimental group and a standard SAC baseline group. The results showed that EG-SAC converged rapidly andoutperfored the standard SAC, with a 75% improvement in convergence speed, and could drive the robotic arm to implement differentiated obstacle avoidance behaviors. The field test results showed that: the success rates of the EG-SAC algorithm in the three scenarios were 80%, 60%, and 75%, respectively, with an average success rate of 71.7%, approximately twice that of the RRT algorithm; the average collision rate was only 13.3%, much lower than the 81.7% of the RRT algorithm; and the operation time ranged from 12.11 to 12.62 s, comparable to that of the RRT algorithm. Failure mode analysis indicated that the main cause of RRT algorithm failure was obstacle misgrasping, accounting for 42.0%, while the failures of the EG-SAC algorithm were mostly attributed to fine-grained control issues. Conclusions In conclusion, this study addresses the obstacle avoidance challenge in unstructured orchards through a differentiated interaction mechanism for rigid-flexible obstacles, an expert demonstration-guided reinforcement learning framework, and high-fidelity simulation and field verification. The research findings provide a promising reference for the practical application of harvesting robots. Future work could further improve the response speed through model optimization and hardware acceleration. In addition, the expert-guided framework could be extended to other complex manipulation tasks, which may help promote the practical application of agricultural robots.

Hand-Eye Configuration and Chassis Stepping Optimization for a Dual-Arm Grape-Picking Robot |

ZHANG Chenghao, YANG Shenghui, XIA Chuandi, ZHANG Kangyang, ZHENG Yongjun, XU Liming, LI Jieyun, XUE Shaohua, LIU Xingxing

2026, 8(5):  135-145.  doi:10.12133/j.smartag.SA202606010

Asbtract ( 155 )   HTML ( 7)   PDF (2764KB) ( 22 )  

Figures and Tables | References | Related Articles | Metrics

[Objective] During field harvesting operations, dual-arm robots often face challenges such as insufficient workspace coverage of the dual arms, substantial overlap between collaborative working areas, and fixed-step movement patterns that are poorly adapted to spatial variations in cluster distribution. These issues can lead to excessive workspace overlap, between the two manipulators, low workspace utilization, and missed harvests due to rigid step sizes, thereby limiting the robot's continuous operation capability and overall harvesting efficiency. To address these problems, the optimization of hand-eye configuration and an adaptive chassis movement strategy for dual-arm picking robots are investigated in this study, aiming to improve workspace utilization and enhance continuous operation performance and overall harvesting efficiency. [Methods] A dual-arm grape picking robot equipped with an onboard storage and buffering system was designed. The robot consisted of a mobile chassis, a dual-arm picking system, a vision perception system, a storage and buffering system, and a control system. Based on operational requirements of the dual-arm collaborative picking system, an effective workspace model was established by intersecting the reachable workspaces of the robotic arms, the depth camera's field of view, and the grape cluster distribution zones. A hand-eye layout optimization model was developed by considering occlusion relationships, using camera row spacing and distance from the camera to the trellis wall as optimization variables. Optimal layout parameters were determined through a two-parameter uniform grid search. Sensitivity analysis was performed for workspace utilization and workspace overlap under different layout configurations, revealing how layout parameters affected workspace coverage and dual-arm picking performance, thereby enabling quantitative optimization rather than empirical parameter selection. To address difficulties in detecting small peduncles and acquiring accurate depth information under long-range "eye-to-hand" visual configurations, the vision system combined YOLOv8 object detection with geometric modeling of clusters to detect grape clusters and estimate peduncle cutting points. Target tracking was further employed to associate clusters with their respective cutting points, providing reliable target coordinates for robotic picking. Additionally, an adaptive chassis movement strategy based on the spatial distribution of peduncle cutting points was proposed to reduce missed harvests caused by fixed-step motion. This strategy dynamically switched between forward and compensatory steps according to whether cutting points fell within uncovered regions of the manipulator workspace, helping to ensure continuous coverage of the picking area and reduce missed clusters due to uneven distribution. Field trials were conducted across five scenarios with different cluster distributions in a standardized horizontal trellis vineyard, comparing the adaptive movement strategy against the fixed-step approach. [Results and Discussions] The optimized hand-eye layout achieved a workspace utilization rate of 89.7%, while the workspace overlap between the two manipulators was reduced to 3.5%, thereby improving coverage of the grape-cluster harvesting region. Sensitivity analysis further revealed that both camera row spacing and distance from the camera to the trellis wall markedly influenced workspace utilization and arm overlap; the optimal parameter settings balanced workspace coverage and dual-arm collaboration performance. Vision experiments reached 100% cluster detection accuracy and 96% cutting-point localization accuracy, meeting the positioning requirements of the robot. Field picking tests demonstrated that the robot achieved an average picking success rate of 94%, representing an 11% improvement over the fixed-step strategy and a reduction of approximately 12% in the missed-harvest rate. [Conclusions] The proposed robot enables stable and continuous field picking operations. The proposed hand-eye layout optimization method improves workspace utilization. The proposed adaptive chassis motion strategy reduces the missed-harvest rate and improves the picking success rate and operational stability. The research results provide a theoretical basis and engineering reference for optimizing the hand-eye system layout, designing chassis motion strategies, and achieving continuous and efficient grape picking in horizontal trellis systems.

A Liftable Omnidirectional Mobile Bait-Feeding Robot for Three-Dimensional Eel Aquaculture |

CHEN Yibo, YUAN Quan, LIU Guanghua, ZHANG Chunyan, XU Zhen, ZHOU Wenzong

2026, 8(5):  146-156.  doi:10.12133/j.smartag.SA202606031

Asbtract ( 18 )   HTML ( 0)   PDF (1867KB) ( 3 )  

Figures and Tables | References | Related Articles | Metrics

[Objective] In the three-dimensional rack layout used in industrialized eel (Monopterus albus) aquaculture systems, the breeding space extends substantially in the vertical direction, while the horizontal channels offer limited room for turning and maneuvering. Under such spatial constraints, manual bait feeding imposes considerable labor intensity, and operations at elevated rack levels are especially difficult. Moreover, existing general-purpose agricultural and aquacultural equipment is poorly suited to these narrow, vertically extended breeding environments characterized by restricted horizontal clearances and multi-tier structures. To address the pressing need for automated and reliable bait delivery in such specialized aquaculture facilities and to reduce the heavy reliance on manual labor in factory-based eel farming, a dedicated three-dimensional bait-delivery robot for eel aquaculture was designed, developed, and systematically integrated. [Methods] The proposed robot integrated several key mechanical and electronic subsystems into a unified platform. The mobility subsystem comprised a dual-steering-wheel omnidirectional mobile chassis, which provided flexible translational and rotational movement capabilities essential for navigating narrow breeding channels with limited turning space. The vertical access subsystem featured a nested gantry lifting mechanism driven by timing-belt pulley groups arranged in series, providing substantial vertical travel range to access multiple tiers of the three-dimensional breeding racks. The manipulation subsystem consisted of a six-degree-of-freedom robotic arm capable of precise positioning and orientation adjustment for accurate bait placement at designated feeding points. These mechanical components were governed by a comprehensive electronic control system organized with a master-slave architecture. Specifically, an RK3588 industrial core board served as the master controller, responsible for high-level task scheduling, navigation planning, and coordination among the various subsystems, while a programmable logic controller (PLC) performed real-time, low-level scheduling and control of the fork mechanism and the lifting motors. For autonomous navigation and localization within the facility, two 2D Light Detection and Ranging (LiDARs) were deployed for simultaneous localization and mapping (SLAM), enabling the robot to construct accurate environmental maps and navigate reliably within the complex breeding facility layout. For robotic arm manipulation, the MoveIt 2 framework together with the Open Motion Planning Library (OMPL) was employed for collision-free path planning, ensuring safe and efficient arm movements within the constrained 3D workspace. Through this integrated electromechanical and software design, the robot addressed the bait-delivery challenges posed by three-dimensional eel aquaculture environments and performed fixed-point bait feeding operations at various elevated levels of the breeding racks. [Results and Discussions] Field tests were conducted at the Zhuanghang Comprehensive Experimental Station in Shanghai to rigorously evaluate the robot's navigation and manipulation performance under realistic operational conditions. The results showed that the navigation-to-point repeatability was better than 2.0 cm, with a maximum deviation radius of 2.14 cm, indicating highly reliable positioning performance that satisfied the precision requirements for mechanical arm-based bait delivery. In terms of manipulation reliability and task execution capability, across 192 actual bait-delivery attempts conducted during the field trials, the single-attempt success rate reached 98.44%, reflecting the robustness of the grasping and placement operations in the actual aquaculture environment. Furthermore, when evaluated against a comprehensive baseline of 200 planned operations encompassing the entire workflow from autonomous navigation to precision bait delivery at designated rack positions, the overall system success rate was 94.50%, demonstrating consistent and robust operational capability under extended field conditions and validating the practical feasibility of the integrated robotic system for industrialized aquaculture applications. [Conclusions] The robot meets the accuracy and efficiency requirements for automated bait-feeding in industrialized eel aquaculture and can reduce the labor cost and physical burden associated with manual bait-feeding operations. These results may support the transition of eel aquaculture from labor-intensive manual operation toward automated and unmanned intelligent farming and may help improve the economic viability of factory-based aquaculture practices.

A Laser Targeting Weed Control System Based on a Parallel Robotic Arm: Design and Experiment |

ZHAO Xueguan, DENG Long, ZHAI Changyuan, WU Huarui, MIAO Zhonghua, ZHENG Shenyu

2026, 8(5):  157-168.  doi:10.12133/j.smartag.SA202605037

Asbtract ( 191 )   HTML ( 12)   PDF (2632KB) ( 40 )  

Figures and Tables | References | Related Articles | Metrics

[Objective] Laser weeding, as an emerging non‑chemical and environmentally friendly solution, offers high precision and immediate action without causing soil disturbance or herbicide residues. However, existing laser systems based on galvanometer scanners have limited scanning ranges; covering wide working widths requires multiple parallel laser units, which substantially increases costs. In addition, variations in weed height and uneven terrain further compromise targeting accuracy. To overcome these challenges and achieve precise and efficient weed control in open‑field vegetable production, a novel dynamic laser targeting system integrating an improved deep‑learning detector, binocular vision, and a parallel robotic arm was proposed. [Methods] Three common weed species-Amaranthus retroflexus, Eleusine indica, and Chenopodium album-which are prevalent in open‑field vegetable fields in the Beijing-Tianjin-Hebei region were selected as targets. A weed image dataset was constructed from 4 200 field‑collected images captured under diverse weather and lighting conditions and was augmented to 16 800 images via rotation, brightness adjustment, blurring, and contrast enhancement; all images were manually annotated in YOLO format. An improved HAD‑YOLO model was developed for real‑time weed detection, and trained on an NVIDIA RTX 3090 GPU using stochastic gradient descent. A hardware platform was established comprising an Intel RealSense D455 binocular camera, an OY‑D‑S‑350 parallel robotic arm, a blue‑laser emitter, an NVIDIA AGX Orin edge computing device, and a robotic arm controller. Custom host‑computer software was developed for model weight selection, confidence and IoU threshold adjustment, and TCP communication with the arm. Coordinate transformation from pixel to robot base coordinates was achieved using camera intrinsic/extrinsic parameters, while depth information was actively acquired through the binocular camera's infrared structured‑light system. Kalman filtering and the Hungarian algorithm were employed for multi‑frame target tracking, and a virtual line was set within the camera's field of view to trigger coordinate transmission when weeds crossed that line. A FIFO‑based path‑planning strategy was implemented to schedule laser tracking sequences, and real‑time encoder pulse counts were used to estimate target positions after they left the camera's field of view, enabling dynamic targeting on the moving conveyor belt. Static positioning accuracy experiments and dynamic laser weeding trials were conducted, with weed mortality assessed at 5 and 7 days post‑treatment based on meristem necrosis and absence of new shoots. [Results and Discussions] Static positioning experiments showed that HAD‑YOLO achieved a mean average precision (mAP) of 94.2% across the three species, outperforming the baseline YOLOv5 model. The average relative error in the X and Y directions was 1.73%, meeting the precision requirement for targeted laser weeding. In dynamic weeding trials, the system achieved a detection rate of 97% and an overall laser hit rate of 90% across all weeds. Post‑hit mortality rates were 86.67% for Amaranthus retroflexus, 89.13% for Chenopodium album, and 81.82% for Eleusine indica. Overall weed elimination rates (hit and subsequently dying) were 78%, 82%, and 72% for the three species, respectively, yielding a combined elimination rate of 77% for all weeds. Observations over seven days showed progressive necrosis, tissue carbonization, and loss of meristem activity in targeted weeds, indicating effective physiological damage. [Conclusions] The proposed dynamic laser targeting system, combining HAD‑YOLO detection, binocular vision localization, and parallel robotic arm control, showed reliable performance with high detection accuracy, acceptable positioning precision, and effective weed elimination. This study provides a feasible technological pathway for intelligent, chemical‑free weed control and may help reduce herbicide dependence and environmental pollution. The findings provide a technical reference for laser‑targeted weeding.

Research on Tractor Trajectory Tracking Control Based on a Dual-Motor Steer-by-Wire System |

LI Lei, PAN Liang, DONG Jiwei, CAO Zhonghua, LUO Xingfa, ZHAN Xiaomei, LI Yali, SUN Zhiqiang

2026, 8(5):  169-181.  doi:10.12133/j.smartag.SA202507027

Asbtract ( 1126 )   HTML ( 10)   PDF (2325KB) ( 25 )  

Figures and Tables | References | Related Articles | Metrics

[Objective] Conventional steering systems for autonomous tractors often suffer from insufficient precision and robustness, making it difficult to meet the demands of high-accuracy trajectory tracking. Although single-motor steer-by-wire (SBW) systems proposed in previous studies can improve tracking precision, they are limited by restricted output torque and poor fault tolerance, especially under complex conditions such as low-speed and heavy-load operations. To address these issues, this study aims to develop a high-performance trajectory tracking control method based on a dual-motor SBW system, thereby improving tracking accuracy, robustness, and operational safety under various working conditions. [Methods] A hierarchical control architecture was constructed in this study. At the upper level, a trajectory tracking controller was designed. Based on a two-wheel bicycle dynamic model of the tractor, a linear quadratic regulator (LQR) controller was developed to solve the optimal control problem and compute the desired front-wheel steering angle for trajectory tracking. At the lower level, a steering execution controller was developed for the dual-motor SBW system. First, a mathematical model incorporating both electrical and mechanical components of the system was established. To address nonlinearities and uncertainties, a sliding mode controller (SMC) was designed for each motor to achieve robust control. To mitigate the inherent chattering problem of SMC, a disturbance observer (DOB) was incorporated. The DOB continuously estimated the total disturbance in real time, including both external interferences and internal parameter variations, and provided feed-forward compensation. This compensation enabled a reduction in the switching gain of the SMC, thereby effectively suppressing chattering while maintaining robust performance. For dual-motor coordination, an independent control strategy was adopted, leveraging the identical parameters of both motors. Under this scheme, the two motors received the same steering angle command and operated in parallel, maintaining structural simplicity while enhancing system fault tolerance. The overall control strategy was structured as a hierarchical system, comprising an upper-level LQR trajectory tracking module and a lower-level steering execution module based on SMC+DOB. [Results and Discussions] To validate the effectiveness of the proposed method, both simulations and field experiments were conducted. Simulations were performed on a MATLAB/Simulink–CarSim co-simulation platform, specifically evaluating steering angle tracking performance and the disturbance rejection capability of the DOB. The results showed that, compared with the SMC controller without DOB, the SMC+DOB controller reduced the front-wheel steering angle tracking error by 13.12% while significantly suppressing chattering. Field experiments were carried out on a T954 tractor platform to evaluate the control strategy under realistic operating conditions. A representative working path, consisting of both straight and curved segments, was used, and tests were performed at two speeds: low speed (10 km/h) and high speed (30 km/h). To benchmark performance, three control strategies were compared: LQR+SMC, MPC+SMC+DOB, and the proposed LQR+SMC+DOB, to assess tracking accuracy, robustness, and real-time control effectiveness. The field experimental results showed that: 1) All SBW-based methods significantly outperformed traditional steering in terms of lateral tracking error and heading error. 2) Under low-speed conditions, the proposed LQR+SMC+DOB method reduced the average lateral tracking error, heading error, sideslip angle, and front-wheel steering angle error by 20.0%, 40.0%, 6.5%, and 14.3%, respectively, compared to LQR+SMC. Under high-speed conditions, the corresponding reductions were 23.8%, 52.2%, 4.5%, and 27.3%. 3) Compared with MPC+SMC+DOB, the proposed LQR+SMC+DOB achieved similar trajectory tracking performance while significantly reducing the computation time per iteration. The results showed that the adopted dual-motor SBW system effectively enhanced steering output torque and system fault tolerance, thereby improving trajectory tracking accuracy and thus providing a hardware foundation for resolving the power and safety challenges in tractor heavy-load operations; moreover, the SMC+DOB controller combined strong robustness with smooth control action, effectively suppressing external disturbances and markedly reducing chattering, thereby improving the accuracy and stability of steering execution. [Conclusions] This research proposed and validated a high-precision trajectory tracking control strategy for autonomous tractors. By integrating an upper-level LQR trajectory tracking controller with a lower-level dual-motor SMC+DOB steering controller, a robust and efficient hierarchical control framework was established. This control strategy showed good trajectory tracking performance and real-time capability across different speeds, offering a reliable technological solution for high-performance operations of autonomous tractors. Compared with the MPC control strategy, the control strategy proposed in the study demonstrated exceptionally high computational efficiency, rendering it highly suitable for real-time control applications.

Information Processing and Decision Making

Rapeseed Silique Detection and Counting Based on YOLOv11n-A³ |

ZHOU Lin, WANG Yan, LI Haixin, ZHENG Jishu, LI Mai, CAO Zhonghua, WU Zhaojuan

2026, 8(5):  182-193.  doi:10.12133/j.smartag.SA202603005

Asbtract ( 145 )   HTML ( 15)   PDF (3161KB) ( 17 )  

Figures and Tables | References | Related Articles | Metrics

[Objective] Rapeseed is the largest oilseed crop in China, and increasing its yield is the major objective of breeding efforts. As a key component of rapeseed yield, the number of effective siliques per plant is an important phenotypic trait in the phenotyping process. Rapid and accurate counting of this trait is crucial for breeding high-yield and superior varieties. Although object detection and counting research has made significant progress in the agricultural field, it still faces some challenges in actual detached phenotyping scenarios. Therefore, aiming at the accurate counting difficulties caused by the widespread dense clustering, local overlapping of branches, and tiny silique beak targets in detached rapeseed phenotyping scenarios, a rapeseed silique detection and counting method is proposed based on YOLOv11n-A³ (YOLOv11n with Area Attention and Aligned Fusion). This method improves detection accuracy and feature extraction efficiency while maintaining a lightweight design, providing technical support for edge-device applications in rapeseed phenotyping and the breeding of high-yield, high-quality varieties. [Methods] First, the proposed method adopted the lightweight convolutional structure of YOLOv11n as its backbone. To enhance the global representation capability of silique features, the cross stage partial with spatial attention (C2PSA) module was embedded in the high-level layers of the backbone for joint channel-spatial modeling. This approach improved the representation of multi-scale features while retaining the receptive-field-expansion advantage of the spatial pyramid pooling fast (SPPF) module, without substantially increasing computational costs. Second, the Area Attention module was introduced to locally enhance the neck feature fusion network. By replacing specific convolutions with the Area Attention module, the network leverages local region perception to improve feature extraction and anti-interference capabilities for densely overlapping small targets, while keeping training and deployment risks manageable. Finally, a bidirectional feature alignment and fusion architecture consisting of an aligned feature pyramid network (Align-FPN) and an aligned path aggregation network (Align-PAN) was constructed to replace the standard feature fusion path. In the Align-FPN stage, high-level semantic features were effectively transmitted to lower levels through the area-attention enhanced cross-feature (A2C2f) module and tiny multi-scale context (TMC) block (P3), and the soft nearest neighbor interpolation (SNI) module was introduced to alleviate spatial inconsistencies among multi-scale features. In the Align-PAN stage, the extended spatial-window down-sampling (ESD) module was further combined for feature separation and enhancement, highlighting key scale features through an adaptive weighted fusion mechanism and improving the interaction and consistency of multi-scale features. [Results and Discussions] The precision, recall, and mAP_0.5-0.95 of YOLOv11n-A³ reached 92.6%, 85.1%, and 47.1%, respectively, increases of 1.8, 2.5, and 12.5 percentage points over the baseline model. The computational load, parameter count, and detection speed were 6.39 GFLOPs, 2.58 M, and 625.3 f/s, respectively. In the silique counting task, the coefficient of determination (R²) reached 0.937; the mean absolute error (MAE) was 2.61; the root mean square error (RMSE) was 3.7; and the mean absolute percentage error (MAPE) was 11.37%. Furthermore, to meet the application requirements of real-time field phenotyping, the lightweight model was encapsulated and deployed in a custom-developed artificial intelligence (AI) tool to achieve fast and accurate counting of rapeseed siliques. This effectively verified the engineering application potential of the proposed lightweight algorithm on edge devices. [Conclusions] The YOLOv11n-A³ model improves counting accuracy while maintaining a lightweight design and detects small targets in densely distributed scenarios. The rapeseed breeding AI tool developed based on this model achieves the rapid and accurate detection and counting of rapeseed siliques, providing technical support for rapeseed phenotyping and the breeding of high-yield, high-quality varieties. Overall, YOLOv11n-A³ achieved a favorable balance among detection accuracy, counting precision, and computational overhead, demonstrating its lightweight and efficient design.

Appearance Defect Detection Algorithm of Euryale Ferox Based on Improved YOLOv11n |

ZHANG Kun, ZHANG Chunyu, CHEN Longmei, LIU Qicheng, LI Yongkang, LIU Kai, ZENG Wenhao

2026, 8(5):  194-206.  doi:10.12133/j.smartag.SA202602012

Asbtract ( 156 )   HTML ( 18)   PDF (2862KB) ( 23 )  

Figures and Tables | References | Related Articles | Metrics

[Objective] To address the problems of high labor costs, low efficiency, and poor consistency in manual sorting during post-harvest processing of Euryale ferox, and to develop a high-precision, lightweight, and real-time sorting model that provides technical support for intelligent processing, an appearance defect detection algorithm based on an improved YOLOv11n model is proposed. [Methods] YOLOv11n was improved in three aspects: feature extraction, downsampling mechanism, and loss function. First, the Universal Perception Large‑Kernel ConvNet Block (UniRepLKNetBlock) was integrated into the C3k2 structure of the neck network to construct a novel feature extraction module named CURK (C3k2‑UniRepLKNetBlock). This module used a depthwise large‑kernel convolution as the main branch, with four parallel convolutional branches. After training, these branches were merged into a single convolutional layer through structural reparameterization, which significantly enlarged the effective receptive field and enhanced the model's representation capability. Second, a Depth Lightweight Adaptive Extraction module (DLAE) replaced the standard convolutional downsampling layer in the 8th layer of the backbone. DLAE adopted a parallel two‑branch design: a feature extraction branch based on depthwise separable convolution (DWConv) to capture local texture details, and an attention branch that generated spatial attention weights through global average pooling and 1×1 DWConv, followed by Softmax normalization. The attention weights were then multiplied channel‑wise with the features, reducing computational load while adaptively enhancing key defect regions and suppressing background interference. Third, the Scale Dynamic IoU Loss (SDIoU) was introduced to replace the original loss function. In the BBox branch, SDIoU calculated a dynamic coefficient based on the ratio of the target bounding box area to the preset maximum target area, combined with the feature-map compression ratio (ROC), and a threshold \delta =\mathrm{0.5} was used for clipping, automatically adjusting the weights of scale loss and location loss. A self‑built Euryale ferox appearance defect image dataset was constructed. A total of 2 537 original images were collected, and after data augmentation, the dataset was expanded to 5 354 images covering five categories: qualified, surface scratch, broken, dark (overripe), and shell. The dataset was randomly divided into training, validation, and test sets in a 7:1:2 ratio. [Results and Discussions] Ablation experiments showed that the combination of CURK, DLAE, and SDIoU achieved the best overall performance while maintaining lightweight advantages: precision was 95.4%, recall was 92.8%, and mAP50 reached 97.4%. Compared with the baseline YOLOv11n, recall increased by 2.9 percentage points and mAP50 by 0.4 percentage points. The model weight file size was reduced to 4.9 MB, parameters to 2.31 M, and the computational cost to 6.1 GFLOPs. The inference speed reached 189.2 f/s, meeting real‑time detection requirements. In comparative experiments with mainstream models, the proposed model achieved the highest mAP50 (97.4%) and the lowest parameter count and computational cost. Heatmap visualization analysis indicated that after integrating CURK, DLAE, and SDIoU, the model's focus on Euryale ferox target regions became more concentrated, background interference was significantly suppressed, and missed and false detections effectively decreased. A physical visual detection validation platform was built, and 402 samples each from Tianchang City and Funan County were tested. The overall accuracy was 94.5% for Tianchang samples and 90.0% for Funan samples, with an average of 92.25%, indicating the model's favorable generalization capability and engineering adaptability under different imaging conditions and geographical origins. [Conclusions] Through the synergistic effects of the CURK large‑kernel re‑parameterization module, the DLAE lightweight adaptive downsampling module, and the SDIoU scale‑dynamic loss function, the improved YOLOv11n model maintains high detection accuracy while balancing model lightweighting and real‑time performance, indicating potential for engineering application. The research can provide an efficient, accurate, and lightweight technical reference for intelligent sorting of Euryale ferox.

Tomato Ripeness and Appearance Defect Detection Model Based on Improved YOLOv10s |

LI Yanmei, YAN Meng, WANG Yu, HANG Qi, WANG Sheng

2026, 8(5):  207-218.  doi:10.12133/j.smartag.SA202605006

Asbtract ( 210 )   HTML ( 22)   PDF (1415KB) ( 27 )  

Figures and Tables | References | Related Articles | Metrics

[Objective] Tomato ripeness directly determines harvest timing, storage and transportation cycle, and market pricing, while appearance defects not only reduce commercial value but also may cause cross-infection during storage, making them key indicators for automated equipment. Traditional manual detection suffers from low efficiency, strong subjectivity and high miss rate, while existing deep learning models generally have shortcomings of insufficient grading granularity, incomplete functional coverage and poor deployment adaptability, hardly balancing accuracy and real-time performance on low-power edge devices. The aim of this study is to propose a lightweight high-precision integrated tomato detection model to provide reliable visual technical support for automated harvesting and intelligent sorting. [Methods] An improved YOLOv10s tomato detection model named Tomato-YOLO was proposed, which simultaneously performed fine-grained grading of 4 ripeness categories and integrated identification of 4 appearance defect categories. A tomato detection dataset containing 3 807 images was constructed for the study, integrating 2 221 public samples and 1 586 self-collected field samples and covering multiple complex field scenarios. The dataset was hierarchically and randomly split into training, validation and test sets at a 7:2:1 ratio, and five data augmentation strategies were adopted during training to improve generalization, with a total of 8 detection labels (4 ripeness + 4 appearance defects) set. Four improvements were implemented on the YOLOv10s baseline: First, the scale-aware context module (SACM) was introduced to replace the C2f module in the backbone network, enhancing the feature capture capability for subtle ripeness differences and micro defects while suppressing field background interference. Second, a cross-layer multi-scale bidirectional feature fusion strategy was constructed and extended to a four-detection-head architecture, with a P2 ultra-small-target detection layer added to fill the scale- coverage gap of the native architecture and boost perception robustness for small and occluded fruits in complex scenes. Third, the Shape Intersection over Union (SIoU) loss function was adopted to replace the native Complete Intersection over Union (CIoU) loss, paired with an adaptive weighting strategy for small targets, adding angle constraints to optimize bounding box regression accuracy for nearly circular tomatoes and raising the learning priority of hard samples. Fourth, layer-adaptive magnitude pruning (LAMP) technology was introduced to prune redundant convolutional layers, and the optimal pruning ratio was determined via gradient comparison to achieve lightweight performance with negligible accuracy loss. [Results and Discussions] Experiments were conducted on the RTX3070 platform, and the results showed that the precision, recall and mean average precision (mAP@0.5) of Tomato-YOLO reached 92.1%, 92.0% and 95.0% respectively, 6.9, 11.6 and 6.8 percentage points higher than those of the original YOLOv10s, with significant performance improvements for both defect and ripeness categories. Compared with mainstream models including faster region-convolutional neural network (Faster R-CNN), single shot multibox detector (SSD), YOLOv5s, YOLOv8s, YOLOv11s, YOLOv12s, Tomato-YOLO achieved significantly higher detection accuracy, while its model size and inference speed were comparable to those of native YOLOv10s, balancing accuracy and efficiency. Ablation experiments verified the independent effects and synergistic gains of the four improvements, and pruning experiments identified 15% as the optimal pruning ratio. At this ratio, the weight file size was reduced to 15.1 MB and computational load to 27.9 GFLOPs, and accuracy decreased by only 1.2 percentage points, achieving nearly lossless lightweighting. When the pruned model was deployed on the Jetson Nano edge platform, the inference speed reached 30.4 f/s, and the overall mAP@0.5 remained at 93.7%, far exceeding the ≥20 f/s real-time detection threshold for agricultural scenarios, meeting the continuous operation requirements of harvesting equipment. [Conclusions] The Tomato-YOLO proposed in this study effectively addresses key limitations of existing tomato detection models, achieving a balance of high accuracy and real-time performance for fine-grained ripeness grading and integrated appearance defect identification. It provides reliable technical support for intelligent harvesting and refined sorting in the tomato industry.

FLPNet: A Lightweight Real-Time Apple Leaf Disease Detection Model Based on Frequency-Spatial Dual-Domain Decoupling |

SONG Xiangshuo, WANG Jun

2026, 8(5):  219-237.  doi:10.12133/j.smartag.SA202605024

Asbtract ( 22 )   HTML ( 2)   PDF (3654KB) ( 6 )  

Figures and Tables | References | Related Articles | Metrics

[Objective] In complex imaging environments, dynamic lighting variations and foliage occlusion can degrade the visibility of disease spot features, making the accurate identification of early-stage micro-lesions difficult for existing models. To address the challenges associated with apple leaf disease detection in natural environments characterized by significant lighting interference, an enhanced real-time detection transformer (RT-DETR) model, termed the frequency-lighting-aware perception network (FLPNet), is proposed for disease detection under complex imaging conditions. [Methods] FLPNet was developed to overcome the limitations of traditional convolutional neural networks in spatial-domain feature extraction by incorporating a frequency-domain decoupling and enhancement module (FDEM). By leveraging the complementary properties of wavelet transform and Fourier analysis, FDEM effectively decoupled and reconstructed fine-grained texture features of occluded disease spots. Furthermore, a physics-driven Retinex layer, designed as a feature enhancement module based on Retinex theory, was introduced to mitigate the effects of uncontrolled illumination. A low-resolution global attention module was also incorporated to reduce computational redundancy while preserving global contextual information. [Results and Discussions] FLPNet achieved an mAP@0.50 of 85.7%, representing an improvement of 5.0 percentage points, while reducing the number of parameters by 3.3 M and the computational cost by 9.1 GFLOPs. Compared with YOLOv12n, FLPNet achieved a 4.1 percentage points improvement in mAP@0.75. The results showed that FLPNet outperformed state-of-the-art models, including YOLOv12n and DEIM, in overall detection performance. The proposed modules significantly reduced computational complexity while enhancing disease detection capability under complex imaging conditions. [Conclusions] FLPNet improves apple leaf disease detection under complex natural imaging conditions while substantially reducing computational complexity. The model achieves a favorable balance between detection performance and computational efficiency, indicating potential for deployment on resource-constrained agricultural edge devices. These characteristics provide essential theoretical support and technical guidance for the practical application of apple leaf disease detection.

RT-DETR-DPF: A Lightweight Kiwifruit Leaf Disease Detection Algorithm |

HE Yating, LUO Youlu, PAN Yonghao, ZHAO Taotao, MOU Guoyu, WU Pengju

2026, 8(5):  238-250.  doi:10.12133/j.smartag.SA202605015

Asbtract ( 18 )   HTML ( 0)   PDF (2012KB) ( 4 )  

Figures and Tables | References | Related Articles | Metrics

[Objective] Kiwifruit leaf diseases are difficult to identify automatically because different diseases can exhibit highly similar visual characteristics, while lesion size, shape, and distribution vary considerably under natural growth conditions. In particular, small and early-stage lesions are frequently missed by conventional object detection models. Manual disease diagnosis is also labor-intensive and strongly dependent on the experience of agricultural technicians. Therefore, this study aims to develop an accurate and lightweight deep-learning-based method for automatic kiwifruit leaf disease detection, improving the recognition of small lesions while reducing model parameters and computational complexity. [Methods] A kiwifruit leaf disease dataset was constructed using images collected from a kiwifruit production orchard in Deyang, Sichuan province, China. Three common diseases, namely bacterial canker, brown spot, and anthracnose, were selected as detection targets. After annotation, 25 453 lesion instances were obtained, with an average of 10.47 lesions per image. The dataset was divided into 2 136 training images, 147 validation images, and 147 test images. To address the limited size of the training set, offline data augmentation was applied exclusively to the training images. The training set was expanded to 3 372 images, whereas the validation and test sets remained unchanged. Based on RT-DETR-r18, an improved model named RT-DETR-DPF was developed through three key modifications. First, the BasicBlock in the backbone was replaced with a detail-enhanced convolution (DEConv) module, which enhances the extraction of lesion edges, textures, and fine-grained structural information through multi-branch differential convolution. Second, the attention-based intra-scale feature interaction (AIFI) module was replaced with a polarity-aware linear attention (PolaFormer) module to strengthen the discrimination of disease-related regions through polarized query-key interactions. Third, a feature-focused diffusion pyramid network (FDPN) was designed to replace the conventional feature pyramid network-path aggregation network (FPN-PAN), improving the interaction and fusion of multi-scale lesion features. The model was further lightweighted by reducing feature dimensions, repeated blocks, and decoder queries. [Results and Discussions] Ablation experiments showed that DEConv, PolaFormer, and FDPN produced complementary improvements in detection performance. The complete RT-DETR-DPF model achieved a precision of 93.1%, a recall of 85.7%, an mAP50 of 92.3%, and an mAP50:95 of 75.6%. Compared with the baseline RT-DETR-r18, precision and mAP50 increased by 5.0 and 2.6 percentage points, respectively. Meanwhile, the parameter count decreased from 19.9 M to 13.2 M, representing a reduction of approximately 33.7%, while GFLOPs decreased from 56.9 to 37.7, corresponding to a reduction of approximately 33.7%. The inference speed was 67.0 f/s. Comparative experiments with other detection models showed that RT-DETR-DPF achieved higher precision than the evaluated RT-DETR variants, RetinaNet, YOLOv8m, YOLOv11l, YOLOv12m, and faster region-based convolutional neural network (Faster R-CNN), while maintaining relatively low model complexity. Compared with YOLOv11l, RT-DETR-DPF improved precision by 4.4 percentage points while reducing the parameter count and computational complexity by 47.8% and 56.5%, respectively. Compared with Faster R-CNN, RT-DETR-DPF improved mAP50 by 11.5 percentage points and reduced the parameter count by 68.0%. Feature visualization further showed that DEConv generated stronger responses around lesion boundaries and texture regions, PolaFormer produced more concentrated responses in disease-related areas, and FDPN enhanced feature interaction across different scales. Although the inference speed was slightly lower than that of the baseline model, the substantial reductions in parameter count and GFLOPs demonstrate a more favorable balance between detection accuracy and computational cost. [Conclusions] The proposed RT-DETR-DPF achieves a coordinated balance between detection accuracy and model complexity through complementary improvements in feature extraction, attention modeling, and multi-scale feature fusion. These results indicate the potential for accurate kiwifruit leaf disease detection in computationally constrained environments. Nevertheless, the current dataset exhibits limited diversity in geographic regions and environmental conditions, and the model requires further validation across different kiwifruit varieties and extreme weather conditions.

Lightweight Pest Detection for Goji Berry Fields via Collaborative Semantic Enhancement of YOLOv11 |

WANG Mei, ZHOU Yuanqing, WANG Wenhui, LIU Ping, ZHU Meilin

2026, 8(5):  251-262.  doi:10.12133/j.smartag.SA202606016

Asbtract ( 43 )   HTML ( 7)   PDF (2748KB) ( 8 )  

Figures and Tables | References | Related Articles | Metrics

[Objective] Goji berry (Lycium barbarum L.) is an important medicinal and edible crop with considerable nutritional, health-care, and economic value. However, pest detection in complex field environments remains challenging because pests exhibit small target sizes, weak fine-grained features, partial occlusion, and strong background interference. These factors can cause incomplete target features and inaccurate localization and classification. Therefore, an object detection model that enhances fine-grained features, strengthens cross-scale semantic information transmission, and maintains low computational complexity is needed for accurate and efficient goji berry pest detection. [Methods] Based on the you only look once (YOLO) framework, a collaborative semantic feature enhancement YOLOv11 (CSFE-YOLOv11) model was proposed. The model performs collaborative optimization at three successive stages: fine-grained feature enhancement, cross-scale semantic fusion, and lightweight prediction optimization. First, an attention-embedded feature enhancement module (AFE) was introduced into the backbone to enhance local fine-grained pest features and suppress irrelevant background responses by combining local convolution-based feature extraction with global attention-based contextual modeling. Second, a cross-scale semantic fusion module (CSF) was constructed in the neck to strengthen nonlinear interactions and semantic information transmission among different scales through a depthwise convolution branch (DWB), a dual-branch nonlinear interaction structure (SC), and a channel mapping and reconstruction component (SSA). Finally, a lightweight decoupled prediction head (LDP head) was designed to separate classification and localization tasks and reduce redundant computation. A 17-class goji berry pest dataset was used for training and evaluation. Small, medium, and large targets accounted for 4.01%, 24.09%, and 71.90%, respectively, while targets with different degrees of occlusion accounted for 38.41%. Comparative experiments, ablation studies, multi-scale analysis, visualization, cross-dataset evaluation, and local field validation were conducted. [Results and Discussion] On the 17-class goji berry pest dataset, CSFE-YOLOv11 achieved a precision of 85.27%, recall of 71.49%, mean Average Precision at IoU=0.5 (mAP50) of 75.29%, and mAP50-95 of 48.25%. Compared with YOLOv11, mAP50 and mAP50:95 increased by 4.47 and 2.36 percentage points, respectively. The number of parameters decreased from 2.59 M to 2.28 M, while GFLOPs decreased from 6.3 to 5.2, corresponding to reductions of 11.97% and 17.46%, respectively. Ablation experiments showed complementary effects among AFE, CSF, and the LDP head. The average precision (AP) of small, medium, and large targets increased from 1.6%, 18.2%, and 50.4% to 5.4%, 32.5%, and 52.8%, respectively. Visualization results showed that the proposed model improved the detection of small, weakly salient, and partially occluded pests while reducing responses to complex background regions. On the IP102 dataset, three independent experiments with random seeds of 42 123, and 456 were conducted. CSFE-YOLOv11 achieved an mAP50 of 63.17%±0.29%, compared with 61.07%±0.68% for YOLOv11, demonstrating improved cross-dataset performance with relatively stable results. Local Ningxia field validation further showed a 4-percentage-point improvement in mAP50 over YOLOv11. [Conclusion] CSFE-YOLOv11 provides a collaborative detection framework integrating fine-grained feature enhancement, cross-scale semantic fusion, and lightweight decoupled prediction. By assigning complementary functions to AFE, CSF, and LDP head at different stages of the detection network, the proposed model improves the representation and prediction of difficult pest targets while reducing model parameters and computational complexity. The results indicate its potential for accurate and efficient goji berry pest monitoring in complex field environments and provide a feasible technical reference for subsequent lightweight deployment in smart agriculture.

Holstein Cattle Pose Estimation Using YOLOv11 with Integrated SimAM and Four-Scale Detection |

LIU Siyi, LI Zhenbo

2026, 8(5):  263-278.  doi:10.12133/j.smartag.SA202604052

Asbtract ( 127 )   HTML ( 11)   PDF (1757KB) ( 13 )  

Figures and Tables | References | Related Articles | Metrics

[Objective] In large-scale Holstein farms, accurate pose estimation remains difficult because images are affected by complex illumination, occlusions caused by facilities and neighboring individuals, and local similarity between black-and-white coat patterns and barn textures, which weaken feature discrimination and reduce keypoint localization reliability. Distal keypoints, especially hoof keypoints, are particularly affected. Existing lightweight models meet real-time requirements but have insufficient representational capacity for these small distal keypoints under occlusion and background interference. To improve distal keypoint localization while keeping the model compact and real-time, a lightweight Holstein cattle pose estimation algorithm, YOLOv11n-AS4D, was proposed based on YOLOv11n-Pose by integrating adaptive attention with multi-scale perception. [Methods] Two complementary improvements were introduced. First, an asymmetric four-scale detection head based on P2 high-resolution features was constructed to compensate for the spatial information loss of small distal keypoints during repeated downsampling. Compared with conventional three-scale structures, the added P2 branch (stride 4) provided a higher-resolution feature layer that retained shallow spatial details for final keypoint prediction. To control the extra cost of this high-resolution branch, three coordinated strategies were adopted: a unidirectional feature flow, in which the P2 branch only received deep semantic features for prediction and was not propagated back into the bottom-up aggregation path, thereby suppressing shallow noise transmission; a progressive channel configuration, with 64, 128, 256, and 512 channels for P2 to P5, which kept the main convolutional cost at each scale comparable level; and depthwise separable convolution in the aggregation path to further reduce computation. Second, an Adaptive SimAM module was proposed to improve feature discrimination under occlusion and texture interference. Unlike the original SimAM with a fixed regularization coefficient and a single-scale response, it introduced three extensions: adaptive regularization, which generated a sample-level coefficient from a global feature descriptor to adjust the attention response to the current feature distribution; multi-scale energy modeling, which combined multi-receptive-field pooling with a Sobel-based edge prior to jointly capture local residual texture and larger-scale context; and hierarchical deployment, which inserted the module after the P2, P3, and P4 stages of the backbone and at the P3 scale of the neck according to their functional roles. The model was trained and evaluated separately on a self-constructed Holstein cattle pose dataset and the public CMBN dataset to examine scene adaptability and generalization. [Results and Discussions] On the self-constructed and CMBN datasets, YOLOv11n-AS4D achieved 97.6% and 98.0% precision, 97.6% and 98.7% recall, and 84.6% and 83.8% mAP@0.5:0.95, respectively. Compared with YOLOv11n-Pose, mAP@0.5:0.95 increased by 1.2 and 1.4 percentage points, precision increased by 0.4 and 0.3 percentage points, and recall increased by 0.4 and 0.5 percentage points, on the two datasets, respectively, indicating improved localization under stricter thresholds. The number of parameters decreased from 2.90/2.91 M to 2.27/2.28 M, while computational cost increased from about 7.3/7.2 GFLOPs to 9.8 GFLOPs and the end-to-end inference speed reached 83.1 f/s, showing that accuracy was improved at a moderate computational cost while real-time inference was maintained. Ablation experiments showed that the two modules were complementary rather than redundant. Their joint gains in mAP@0.5:0.95 were 1.2 and 1.4 percentage points on the self-constructed and CMBN datasets, respectively, exceeding the summed individual gains on the self-constructed dataset and remaining close to their sum on the CMBN dataset. This indicated that the four-scale detection head and Adaptive SimAM enhanced different aspects of feature representation: fine-scale spatial localization of distal keypoints and feature discrimination under occlusion, respectively. In occlusion robustness tests, average precision (AP) under severe occlusion increased from 73.5% for YOLOv11n-Pose to 75.9% for YOLOv11n-AS4D, and on continuous video sequences, the keypoint missing rate decreased from 8.73% for the baseline to 1.94% for the proposed model. Per-keypoint analysis showed reduced hoof localization errors, supporting the benefit of high-resolution feature compensation for small distal keypoints. [Conclusions] The proposed YOLOv11n-AS4D improved Holstein cattle pose estimation accuracy under complex farming conditions while reducing the number of parameters, with only a moderate increase in computation. By combining an asymmetric four-scale detection head with Adaptive SimAM, it mitigated spatial information loss of distal keypoints and improved feature discrimination under occlusion, while maintaining real-time inference, thereby providing more reliable skeletal perception for downstream behavior recognition, lameness assessment, and intelligent herd management.

Dairy Cow Feeding‑Behavior Recognition Based on Jaw‑Movement Features Optimized via Model‑Performance Feedback |

DENG Xiongyang, ZHANG Jiahao, LIU Jie, ZOU Zeshi, ZHANG Wenkuan, CAO Jianpeng, CHEN Guipeng

2026, 8(5):  279-292.  doi:10.12133/j.smartag.SA202605016

Asbtract ( 82 )   HTML ( 9)   PDF (2098KB) ( 15 )  

Figures and Tables | References | Related Articles | Metrics

[Objective] Feeding behavior is widely used to monitor the physiological status of dairy cows and supports early disease warning in herd management. Among various behavioral indicators, jaw movement stands out as the most direct external sign of feeding, providing a physically interpretable signal for automated recognition. Traditional recognition models for dairy cow feeding behavior, however, suffer from low accuracy and high hardware overhead. These limitations make them difficult to deploy on resource-constrained edge devices commonly found in pasture environments. To address these issues, a classification and monitoring approach is proposed that pairs a custom wearable device with lightweight ensemble machine-learning models. The goal is to build a lightweight model that balances recognition accuracy with on-device deployability, so that the system can run in real time on the hardware available at typical livestock farms. [Methods] A compact wearable unit integrating a tension sensor and a six-axis inertial measurement unit was fixed to the lower jaw of each cow. This device enabled synchronous acquisition of jaw opening-closing dynamics and head posture variation throughout natural feeding bouts. The feature extraction process was organized around the physical properties of the captured signals. Specifically, features were computed across four complementary dimensions. Time-domain statistics characterized amplitude and temporal distribution. Frequency and time-frequency representations captured rhythmic chewing patterns. Spatial correlation metrics encoded the inter-axis coordination of head motion. Trajectory-synthesis descriptors summarized the geometric path of jaw movement over time. The resulting 110-dimensional feature space preserved biologically meaningful signal structure across all four dimensions. Because many of the 110 features were redundant, a performance-feedback-driven restricted backward elimination strategy (PF-RBE) was designed and validated for dimensionality reduction. Conventional recursive feature elimination (RFE) ranked features by a single importance score and discarded the lowest-ranked ones in bulk, which risked removing features with limited individual contribution but meaningful joint effects. PF-RBE addressed this problem by using feature importance only as a search heuristic. The actual retention or elimination of each feature was determined by measuring the change in classification performance after each ablation step. Through recursive evaluation, this strategy retained key features more precisely and effectively balanced classification performance against feature dimensionality. The process yielded a compact 59-dimensional core feature subset. Four ensemble algorithms, Random Forest, XGBoost, LightGBM, and CatBoost, were then evaluated under multiple time-window lengths to determine how temporal context length affected recognition accuracy across different model architectures. [Results and Discussions] Recognition performance improved in a stepwise fashion as the time window was extended. The XGBoost model was trained on the PF-RBE-selected 59-dimensional feature subset with a 60-second time window. Through systematic comparison across different time windows, the 60-second window was established as the optimal behavioral perception window size. Notably, this window was able to completely describe the behavioral habits of dairy cows. Full feeding cycles including transitions between sub-behaviors were captured, and recognition accuracy improved substantially. Among all candidate models, CatBoost and XGBoost achieved the highest accuracies of 93.71% and 93.64%, respectively. Further comparison showed that, under the 60-second window configuration, the XGBoost algorithm maintained competitive classification performance while requiring fewer computational resources. Its model file was only 4.48 MB and its total inference time was merely 3.925 6 s. XGBoost offered an advantage over the other algorithms by combining competitive classification accuracy with low computational cost. The F₁-Score reached 92.81%, and was adopted as the core evaluation metric under class-imbalanced conditions. [Conclusions] Under conditions of limited computing power and storage, XGBoost still maintained high recognition accuracy. Its compact model size and low inference cost were well suited to the hardware constraints typical of edge-wearable devices deployed on livestock. Overall, the proposed approach provides a feasible technical reference for the deployment of edge computing and real-time behavioral monitoring in smart pasture applications.

Intelligent Management and Control

Dynamic Energy-Saving Supplementary Lighting Control Method for Protected Grapes Based on Optimal Photosynthetic Efficiency Point |

JIN Meng, LI Hualong, LI Miao, HE Jin, LIU Xianwang, MA Zhirun, XIE Miaomiao, CHENG Binbin, XUAN Lei, YE Fei

2026, 8(5):  293-305.  doi:10.12133/j.smartag.SA202603010

Asbtract ( 26 )   HTML ( 2)   PDF (4478KB) ( 5 )  

Figures and Tables | References | Related Articles | Metrics

[Objective] In protected grape cultivation, supplementary lighting can improve crop photosynthetic efficiency, regulate plant growth, and improve fruit yield and quality. However, most existing supplementary lighting studies focus on optimizing instantaneous photosynthetic rate or yield-related indicators through fixed or empirical lighting strategies, ignoring the complex trade-off between photosynthetic performance and energy consumption under variable greenhouse conditions. As energy costs rise and sustainable agriculture becomes more important, a dynamic energy-saving supplementary lighting control strategy for greenhouse Shine Muscat grapes based on the optimal photosynthetic efficiency point (OPEP) is proposed in this study, with plant physiological response modeling and multi-objective optimization integrated, to develop an intelligent, energy-saving lighting regulation system that balances crop physiological demands and energy conservation. [Methods] First, a thorough dataset was built by collecting crucial environmental and physiological factors at various stages of the grape veraison phase, such as temperature, photosynthetically active radiation (PAR), and leaf-level photosynthetic rate. The choice of the veraison period is especially crucial, since this stage signifies a key transitional stage in grape growth, when both the demand for carbon assimilation and the sensitivity to the environment are elevated. A modified rectangular hyperbola model was used to describe the nonlinear response relationship between PAR and photosynthetic rate, allowing for precise depiction of light saturation and the law of diminishing returns under high-irradiance conditions. In addition, a D-Sigmoid function was employed to simulate the temperature–photosynthetic rate response, depicting the asymmetric and threshold-dependent physiological characteristics of grape leaves under different thermal conditions. Based on these two sub-models, a light-temperature interaction model was built to reflect the joint effects of environmental factors on photosynthetic performance, enabling more realistic simulation of greenhouse microclimate changes. A multi-objective optimization system was then established with two conflicting aims: (1) maximizing the net photosynthetic rate to ensure adequate carbon assimilation for grapevine growth, and (2) minimizing the energy consumption of supplemental lighting to enhance energy efficiency and cut down operational expenses. To address this issue, the strength pareto evolutionary algorithm 2 (SPEA-2) was employed because of its strong ability to preserve solution variety and approach the Pareto-optimal front in intricate nonlinear search domains. The algorithm repeatedly produced a Pareto frontier that showed the trade-off between photosynthetic performance and energy consumption. Afterward, a sequential search algorithm was utilized to precisely determine the OPEP from the Pareto set. [Results and Discussions] Experiments were conducted under controlled greenhouse conditions to evaluate the efficacy of the proposed approach. The findings showed that, in the grape veraison stage, the suggested dynamic supplementary lighting approach notably enhanced energy utilization effectiveness when compared with traditional fixed-intensity lighting plans. In particular, the consumption of lighting energy was decreased by an average of 34.27%, whereas the decline in photosynthetic rate was restricted to merely 6.60%. This shows that the system effectively prevents excessive compromise of physiological performance when attaining considerable energy savings. A deeper examination of the temporal response curves indicated that the suggested method enabled dynamic adjustments to the lighting intensity based on environmental variations, avoiding needless energy consumption during times of light saturation or when temperature conditions were favorable. This adaptable behavior is especially advantageous in greenhouse settings where microclimatic circumstances can vary rapidly and non-linearly. Moreover, the Pareto frontier produced by SPEA-2 showed a smooth and evenly-spread trade-off surface, validating the robustness of the optimization framework. The recognized OPEP point was found to be situated close to the inflection area of the response curve, where marginal photosynthetic benefits start to decrease significantly in relation to energy input. This shows that the proposed method can approach the physiological efficiency limit of grape leaves when there is combined light and temperature control. [Conclusions] The proposed approach is able to cut down the energy consumption of supplementary lighting while maintaining a relatively high photosynthetic capacity for greenhouse-cultivated Shine Muscat grapes. By combining physiological response modeling with multi-objective evolutionary optimization, this method provides a systematic and interpretable framework for precise lighting control. It efficiently enhances the utilization rate of energy and attains an ideal equilibrium between production advantages and energy usage in protected horticultural systems. Furthermore, this framework may also be applicable to other high-value greenhouse crops and provides a general strategy for smart agricultural energy management under future climate and sustainability limitations. Future research could further integrate other environmental factors, such as humidity and CO₂ concentration, and investigate real-time adaptive control tactics based on reinforcement learning to improve model responsiveness in dynamic greenhouse environments.

Digital Economy

The Impact of Agricultural Mechanization on Farmers' Income: A Causal Inference Approach Using Double Machine Learning |

CAO Yangchun, LI Yuanjie, ZHANG Jing

2026, 8(5):  306-318.  doi:10.12133/j.smartag.SA202604002

Asbtract ( 95 )   HTML ( 4)   PDF (944KB) ( 13 )  

Figures and Tables | References | Related Articles | Metrics

[Objective] Agricultural mechanization is an important foundation of agricultural modernization, but the simultaneous constraints of land fragmentation and rural labor outflow have made its income-increasing effect increasingly uncertain. Most previous studies rely on a single indicator, such as total machinery power, however this cannot reflect the multi-dimensional character of mechanization. In addition, the mediating routes of mechanization's impact on income have not been fully elucidated. Many empirical analyses are based on parametric linear models, which may be subject to high-dimensional control variable biases and misspecification of nonlinear interactions. Therefore, it is important to elucidate the causal effects, transmission channels, and contextual dependencies of agricultural mechanization on farmers' income for the formulation of evidence-based policies to support sustained income growth and high-quality agricultural development. [Methods] The double machine learning (DML) framework with support vector machines was used as the basic learner, and used panel data from 31 Chinese provinces during 2008–2023. The main explanatory variable was the comprehensive mechanization level (CML). The main outcome was rural per capita disposable income, while wage income and net operating income were also conducted for structural decomposition. A wide range of control variables were incorporated to account for agricultural output, fiscal support, catastrophe risk and socio-economic factors. For robustness, winsorization was adopted, sample splits (1:2 and 1:7) were altered, municipalities were excluded; the learning technique was updated using random forest and gradient boosting; and instrumental variable estimation with non-agricultural earnings was applied. Mediation models were employed to examine if the strategy works through improving the agricultural operation level (AOL). Moderation models were used to test whether agricultural machinery socialized services, cooperative density, equipment structure, and urban-rural income gap strengthened the income-enhancing effect. Heterogeneity analyses were also carried out by grain-producing region, geography, and income level. [Results and Discussions] The benchmark estimation results showed that the level of agricultural mechanization had a significant positive impact on total income. For every 1% increase in the comprehensive level of agricultural mechanization, the per capita disposable income of farmers increased by approximately 0.18%. When income sources were broken down, mechanization had a considerable positive impact on net operating income, but a negative impact on pay income. The gain in operating income far exceeded the loss in wages. The mediation analysis showed that the agricultural operation level was an important route for mechanization to improve resource efficiency and reduce unit costs. The heterogeneity analysis showed three patterns: The income effect was significantly negative in major grain-producing areas but positive in non-grain-producing regions; positive in high-terrain relief areas but negative in low-relief plains; and statistically significant only in low-income provinces. The moderation tests showed that socialization services of agricultural machinery, cooperative density, equipment structure and urban-rural income gap positively moderated the income effect. Additional interaction analyses demonstrated that agricultural labor productivity, fiscal expenditure, grain production, financial development, and GDP had a positive moderating effect, however the crop disaster rate dramatically reduced the effect, emphasizing the restrictive role of natural risks. [Conclusions] The study confirms that agricultural mechanization remains an effective means for increasing farmers' income, but its benefits depend heavily on operational improvement, development of the service market, organizational density, equipment upgrading, and factor price incentives. The policy should be changed from expanding quantity to the structural optimization in grain-producing and plain areas, while promoting mechanization in non-grain-producing and mountainous areas. Full income gains will be unlocked through expanding socialized services, strengthening equipment structures and responding to wage signals. Further research should go deeper into interaction impacts of land transfer and machinery service markets and the micro-level mechanisms at the household or county levels, providing a more comprehensive picture of the diverse conditions that allow mechanization to benefit farmers.