Over the last decade, agricultural robotics and mechatronics have advanced significantly. Notable examples include the commercialization of unmanned agricultural vehicles for cereal cultivation and the development of artificial intelligence applications in agricultural image analysis rooted in field science.

However, food-security-related challenges in agricultural robotics and mechatronics cannot be effectively addressed through simple extensions or combinations of existing solutions alone. Meanwhile, the importance of food security is rapidly increasing under changing environmental, economic, and societal conditions. Nevertheless, some of these challenges have not yet been recognized or formally defined, and thus remain latent.

This special issue has two primary purposes.

(I) It reveals previously overlooked challenges not only within agricultural fields but also in off-field infrastructures and agricultural facilities, thereby broadening the scope of agricultural robotics and mechatronics.

(II) It presents the potential of “new seeds,” including novel scientific understandings of real agricultural components and systems as well as principled integrations of existing solutions with artificial intelligence and data science methodologies.

Beginning with this editorial, this special issue comprises 20 papers, including five development reports and one review.

We express our sincere gratitude to all the authors and reviewers. We hope that this special issue will stimulate further research and development addressing food-security-related challenges in agricultural robotics and mechatronics.

Agricultural robots that are operated in open fields have different mechanisms, sensing technologies, and functionalities depending on the crop and task. In this review, we present the classification of such agricultural robots and describe the mobile mechanism, power source, implementation, navigation, and sensing technology for each task. Finally, conclusions are suggested based on future perspectives.

This paper discusses desired disturbance response model + desired reference response model for simultaneously satisfying an impedance control and a position orientation tracking control for nonlinear hydraulic arms. First, we review hydraulic cylinder dynamics with nonlinear pressure dynamics. Second, through the reviewed hydraulic cylinder dynamics and an existing controller design procedure, we discuss two desired models for achieving an impedance control and a position orientation tracking control simultaneously. Finally, the effectiveness of the proposed desired models is confirmed by a preliminary indoor experiment including a simulated open channel.

This study presents a novel map-based energy consumption prediction model for agricultural electric vehicles operating in real orchard environments. Traditional static models often overlook resistance factors caused by varying terrain and soil conditions. To address this, we introduce an unknown resistance component F_u, mapped spatially to reflect local environmental influences such as slope and soil hardness. Field experiments were conducted in a vineyard in Hokkaido, Japan, using GNSS and battery data collected at 10 Hz during uphill and downhill runs. The proposed model achieved a maximum mean absolute percentage error of 2.3%, significantly outperforming conventional models. A notable negative correlation between F_u and soil hardness was observed, confirming that softer soils increase vehicle resistance. Simulations of continuous operations across adjacent routes further demonstrated reduced cumulative prediction errors, supporting applications in route planning and battery management. F_u(x,y) is currently treated as static, and future work will expand it to a spatiotemporal parameter F_u(x,y,t) to incorporate dynamic environmental changes. Online learning and validation across diverse terrains are also planned. This approach enhances model adaptability, offering a reliable tool for energy-efficient and sustainable operation of electric vehicles in agriculture.

This study proposes an electric crawler-type robot for autonomous weeding operations on steep slopes. We designed a coverage weeding path for autonomous travel for robots using a global navigation satellite system. The machine was designed to adjust the height of both crawlers to address variations in slope. In addition, the path following the deviation error of the robot with different roll angles was measured. The proposed robot demonstrated optimal performance when the roll angle was set to 17.7°, resulting in an average lateral deviation error of 0.02 m and a weeding area coverage of 99.3%.

Recently, consumers have shown a high level of interest in food safety and security, as well as in reducing environmental impacts. Therefore, for rice, which is the staple food of the Japanese people, its cultivation without the use of herbicides or chemical fertilizers is desired. This paper proposes a two-wheeled weeding robot that floats on a water surface and is equipped with a movable mechanism in which large-diameter wheels contact the soil under their own weight in response to water level height and soil unevenness. The robot body was designed to float on water by placing floats on both sides and spanning the rice row, enabling it to continue moving even as the rice grows. The movable system utilizes a simple yet robust four-bar linkage, which is designed to accommodate soil adhesion. This study first clarified the design method for a four-bar linkage mechanism. Next, it experimentally investigated the effect of the angle of paddles attached to the wheels on the propulsive force. Finally, it experimentally confirmed the effectiveness of the proposed weeding robot.

This study proposes a generalizable and extensible framework for task allocation among multiple agricultural machines. Although several previous studies have focused on specific aspects, such as route planning and task scheduling under constrained conditions, few have addressed the combined challenges of task division, variability in farmland scale, and algorithm selection with hyperparameter tuning in an integrated manner. To fill this gap, we formulate the problem as a split delivery vehicle routing problem, which enables flexible division of field tasks across machines. Based on this formulation, we construct a unified framework that incorporates farmland modeling, machine modeling, and farmer-specific preferences. The proposed framework is designed to accommodate multiple optimization algorithms such as simulated annealing, local search, genetic algorithm, and ant colony optimization under a common structure, allowing flexible applications across diverse agricultural scenarios. We evaluated the performance and sensitivity of the algorithm to the hyperparameters using simulations for varying farmland sizes and computation times. The results demonstrate that the framework effectively supports algorithm selection and parameter tuning according to situational needs. This approach offers a versatile foundation for optimizing agricultural tasks, and can be extended to dynamic and real-time environments using real farmland data.

This study proposes a costmap generation method for orchard navigation that integrates both semantic and geometric information from point clouds acquired using unmanned aerial vehicle (UAV)-based light detection and ranging (LiDAR). Conventional approaches often rely on aerial imagery, which cannot capture the internal structures of tree crowns, or on ground-based mapping, which is inefficient and typically limited to height-based costmaps. In this study, orchard-scale three-dimensional point clouds were acquired using UAV-LiDAR, and RandLA-Net was applied for semantic segmentation to classify tree trunks, crowns, and ground. Based on this classification, we constructed a semantic costmap that incorporates obstacle height and shape and integrated it into the Navigation2 framework for unmanned ground vehicle (UGV) navigation. Simulation experiments (20 trials) achieved an 85% success rate, significantly higher than that of conventional methods (60%–65%). Furthermore, field experiments (15 trials) achieved a 93% success rate, demonstrating safe and efficient path planning even in densely canopied environments.

This study investigates a method for collision detection caused by soft and deformable obstacles, such as soil and plants (soft obstacles). Field robots must operate while interacting with the soft obstacles in their environments. However, previous studies on collision detection have primarily focused on hard obstacles, such as rocks and artificial structures, and research addressing soft obstacles has been limited. Because soft obstacles absorb impact forces, collision detection using conventional methods is challenging, and determining the collision time is even more difficult. To address these challenges, this study proposes a method that estimates the states, including unknown collision forces, using an extended Kalman filter. A wavelet transform is then applied to the estimated values to detect collisions. The effectiveness of the proposed method was validated using experimental data obtained from a water tank.

This study presents a real-time fruit monitoring system that integrates light detection and ranging (LiDAR) and RGB camera data for 3D fruit counting and spatial density mapping in horizontal trellis pear orchards. The system employs instance-level sensor fusion, combining YOLO-based 2D fruit detection with SLAM-generated 3D point clouds to localize and track individual fruits. A customized temporal tracking algorithm mitigates duplicate counts, while center-based spatial filtering improves detection accuracy. Among the four evaluated YOLO models, YOLOv11s was selected based on its F1-score, lowest false negatives (FN) count, and real-time performance. Field validation in a 6 m × 70 m orchard plot demonstrated high counting accuracy (96.2%) and reliable spatial density estimation, with a mean absolute error of 0.64 fruits/m². The system effectively identified yield variations across different orchard regions. These findings support the use of LiDAR–camera fusion for scalable, high-precision fruit monitoring in orchard environments, particularly in labor-intensive horizontal trellis systems.

In recent years, the damage to humans and crops caused by bears and other vermin has become increasingly serious across Japan. Although smart agricultural monitoring systems have shown some promise, they are still limited by issues such as specificity to certain species, high expenses, and a lack of adaptability. This study focused on creating and testing a zero-shot system for vermin detection using a multimodal large language model. A total of 1,073 images were collected using cameras installed at three locations in Nanae-cho, Hokkaido, Japan, between May and September 2025. Twenty-two images showed the target animals, including 12 bears, nine deer, and one crow. A comparative evaluation of GPT-4o, LLaVA, YOLO-World, and Grounding DINO showed that GPT-4o had promising recall in our preliminary deployment (recall =1.00), although 17 false detections occurred in images without animals.

In this study, we addressed agricultural labor shortages by developing a smart farming sensor module that integrated low-cost environmental sensors with a multipoint soil moisture sensor to predict broccoli growth, plant height (PH), and leaf count (L_n). Multivariable regression confirmed that integrated solar radiation (S) was the most dominant factor, although broccoli growth involved a complex interplay of solar radiation, optimal temperature, humidity, and soil moisture. More importantly, the analysis revealed that the middle layer soil moisture (u_m) exhibited the strongest positive contribution to PH. This finding indicated that water availability in the main root zone was essential for vertical growth and highlighted the indispensability of multipoint sensing over conventional single-depth measurements to accurately model the intricate relationship between soil moisture and crop development. Moving forward, we aim to leverage the superiority of multipoint data to construct a sophisticated growth prediction model, thereby contributing to the optimization of irrigation and temperature management in smart farming systems.

Automated potato harvesting requires vision systems meeting stringent operational constraints: near-zero potato misclassification (PMR <1%) and high impurity detection rate (IDR >60%). Although YOLOX supports real-time processing, it exhibits critical performance limitations for small object detection. This paper introduces a unified optimization strategy combining training-level modifications (P3 feature enhancement, SimOTA parameter optimization, and size-aware loss weighting) with inference-level threshold optimization. Experimental results based on 10,000 images containing 232,000 annotations demonstrate that all four evaluated approaches achieved the defined operational constraints. SimOTA emerged as the optimal configuration, delivering the highest small object recall (18.18%) while maintaining PMR of 0.06% and IDR of 95.05%. P3 feature enhancement achieved the lowest PMR (0.05%), SALW provided balanced performance (PMR 0.06%, IDR 96.11%), and the P3+SimOTA combination failed critically (PMR 9.90%), revealing fundamental incompatibility between optimization components. All successful configurations exceeded real-time processing requirements (45–61 fps), confirming suitability for deployment in resource-constrained agricultural automation systems.

Orchard trees are primarily cultivated for food production, but they also provide environmental and aesthetic benefits. Proper maintenance, particularly through pruning, is crucial; however, manual pruning is labor-intensive, time-consuming, and dependent on expertise, limiting its consistency on a large scale. Automated pruning reduces labor demands and enhances scalability. In another context, automated pruning encounters additional challenges owing to the complex and inconsistent geometry of trees (branch size, position, and orientation) and their hierarchy (parent-child relationships), which play a vital role in pruning decisions. To address these challenges, a pipeline for estimating tree geometry and hierarchy was proposed. Branches and trunk from a single RGB image were segmented using a custom YOLOv8 model, and key geometric and distance features were extracted through principal component analysis, which captured over 99% of the geometric variation. A genetic algorithm then infers hierarchical relationships, assisting in the recognition of branch levels and supporting biological pruning decisions. The experimental results demonstrated distinct features across the hierarchical levels, achieving an F1 score of approximately 80% and a Jaccard index exceeding 70% during hierarchical validation. These findings demonstrate the potential of the proposed method to transform visual perception into geometric and hierarchical representations of tree structure, thereby providing essential structural information to support autonomous and biologically informed pruning decisions.

Berry thinning is essential for producing high-quality grape varieties such as Shine Muscat because it directly impacts fruit size and quality. To address the labor-intensive nature of this task, this study presents an autonomous robotic arm system that integrates depth sensing with a learning-based transformation method and is implemented using a ResNet-18 convolutional neural network to predict berry coordinates and execute cutting actions. Its performance was compared with a geometric transformation method based on Robot Operating System 2 (ROS2) coordinate transformations in both indoor and outdoor environments. In indoor trials, the learning-based transformation approach achieved an approach accuracy of 96.8% and cutting accuracy of 78.5%, outperforming the geometric transformation approach, which achieved 94.6% for approach and 69.6% for cutting. On outdoor slopes, environmental challenges degraded the performance of both the approaches; however, the learning-based transformation method maintained higher accuracies, achieving 75.6% for approach and 60.3% for cutting, compared with the geometric transformation approach, which achieved 63.1% approach accuracy and 44.1% cutting accuracy. The complete thinning cycle required an average of 3.67 min to process 10 berries, confirming its feasibility for practical use. Limitations in the curved scissor end-effector reduced cutting effectiveness, highlighting the need for improved blade design. This study demonstrates the potential of combining geometric and learning-based transformation methods for artificial intelligence-driven robotic thinning to achieve efficient vineyard management.

Spraying agricultural chemicals is vital for maintaining crop yield and quality; however, it comes with its own set of challenges such as spray drift and rising costs. Despite advancements in autonomous vehicles that reduce the reliance on manual labor, issues related to pesticide spray volume and spray drift persist in robotic application systems. Hence, efficient spraying technology is essential for the uniform application of pesticides, especially in hedge-trained crops where dense canopies limit access to abaxial surfaces. Although conventional air-assisted spraying systems enhance coverage through high-velocity airflow and spray pressure, there is a practical limit to the spray volume reduction as insufficient spray volume can reduce penetration into dense and complex canopies. In this study, we propose a novel tractor-mounted two-stage spraying system with a leaf-turning mechanism designed for hedge-trained crops, targeting both adaxial and abaxial surfaces. Experiments were conducted under controlled laboratory conditions and real vineyard conditions, and the results showed that the proposed spraying system used only 60% of the spray volume compared with a conventional air-assisted spraying system, while achieving equivalent coverage on adaxial surfaces and a 6% improvement in coverage on abaxial surfaces. These findings suggest that the proposed spraying system can reduce spray drift while maintaining or improving spraying performance, primarily due to the reduced spray volume, minimized air–pesticide-droplet interactions, lower operating pressure, and limited air assistance range.

Regular mowing operations are required on ridge paths; however, ridge paths are often narrow and sloped, making mechanization necessary from the perspectives of work efficiency and safety. In this study, the performance of currently available commercial mowers was first organized and analyzed, and a conceptual design of a mower specialized for ridge-path operations was proposed. First, a remote-control operation system was adopted to reduce the operator’s workload. Next, a hybrid drive system was employed in which an engine generates electricity to drive electric motors. In addition, a hammer-knife mechanism, which provides high durability and excellent mowing performance, was adopted as the mowing mechanism. Furthermore, in order to enable operation in narrow environments such as ridge paths, the machine width was set to approximately 600 mm. To evaluate the performance of the developed machine, traveling tests, mowing performance tests, and fuel consumption tests were conducted. In the mowing performance tests, evaluations were carried out under different cutting-height conditions, and a high mowing performance of over 96% on average was confirmed under the cutting height condition of 80 mm. In contrast, under the cutting height condition of 180 mm, grass was pressed down beneath the machine body during traveling and did not reach the cutting blades, resulting in a tendency for approximately 15% of the grass to remain uncut on average. In addition, turning tests under slope conditions were conducted on both asphalt and grass surfaces. The results showed that although sufficient performance was not obtained on the asphalt surface under the slope angle condition of 20°, stable traveling and turning performance were confirmed on the grass surface, which simulates actual ridge-path environments. Furthermore, the evaluation of fuel consumption characteristics demonstrated that the developed machine is capable of long-duration continuous operation. From these results, it was confirmed that the developed machine has sufficient adaptability to narrow environments such as ridge paths and exhibits good traveling and mowing performance.

Harvesting and thinning in orchards involve intensive fruit transport, which is inefficient and burdensome, particularly in mountainous and hilly areas. Heavy vehicles can damage soils and roots, whereas manual transport is labor-intensive. We developed a quadruped robot system capable of stable walking on uneven terrain. Equipped with detachable harvest baskets, an electric basket for discarded fruit, and voice-command control, the system enhances efficiency and enables sustainable orchard management in challenging landscapes.

This study developed and experimentally evaluated an automatic boom-height control system for boom sprayers for large-scale field crop farming by integrating a compact position-sensitive detector (PSD)-based boom-height detection device with a hydraulic control mechanism. The detection device, equipped with three PSDs arranged at a 6° outward tilt in a three-dimensional configuration, suppressed optical interference and contributed to system miniaturization. A two-step signal processing algorithm combining minimum-value selection and median filtering enabled stable measurement of crop canopy height, with detection errors within ±5 cm for potato, soybean, and sugar beet. Larger errors (+7.4 cm) were observed for wheat, indicating the need for crop-specific correction. Step response experiments of the hydraulic control system were conducted using a container 65 cm high as a reference object. The results showed rise and fall times of 0.95 and 2.41 s, respectively, demonstrating stable operation without overshoot. These findings highlight both the effectiveness of PSD-based sensing for boom-height detection and the necessity of implementing more sophisticated control strategies to achieve accurate regulation in long, flexible agricultural booms.

This study presents a comprehensive tomato harvesting robot system addressing three critical technical aspects: 1) optimal manipulator configuration design, 2) robust environmental recognition, and 3) efficient end effector control. For the manipulator configuration design, four different mounting configurations were systematically evaluated, with the vertical configuration featuring an offset end effector achieving the highest target reachability of 97.7%. For environmental recognition, a multi-sensor system that combines RGB and depth (RGBD) cameras and light detection and ranging (LiDAR) was implemented, utilizing depth filtering to suppress outliers. The end effector integrates suction and cutting mechanisms, employing a suction pad with conforming motion and Bowden cable-driven scissors. A bunch model was developed based on actual fruit bunches to create a testing environment with diversity and reproducibility. Field experiments conducted in a commercial greenhouse demonstrated continuous harvesting operations with a 68% suction success rate and a 45% overall harvesting success rate across 159 target fruits from 200 bunches. Additionally, the fruit position distribution in the field was measured, which can be utilized for layout optimization. This study contributes to advancing practical agricultural robotics by providing validated solutions for the three fundamental challenges in robotic crop manipulation.

In this study, we propose a sparse recurrent neural network model with regularization to identify the environmental factors contributing to strawberry growth evaluation. The learning data consisted of eight environmental variables, including carbon dioxide concentration and solar radiation, measured by observation devices in a greenhouse. The target data included shipment volume, quality, growth status, and the farmer’s intuitive evaluation. Data obtained during three periods between February and March 2025 were used in this study. Model training was performed using backpropagation through time with the mean squared error as the loss function. To induce sparsity, L1 and L2 regularization were applied, suppressing moderately influential weights and yielding a more interpretable model structure. Unlike conventional black-box models that rely on post-hoc explanation techniques, the proposed method constructs an intrinsically interpretable learning model in which the influence of each environmental variable is directly reflected in the learned network structure. Rather than improving prediction accuracy, this study aimed to clarify the dominant environmental factors through a transparent and structurally constrained learning framework. The results suggested that growth evaluation was influenced by the carbon dioxide concentration around observation device 1, atmospheric pressure measured across multiple devices, and other environmental variables. These findings demonstrate that important environmental factors in strawberry cultivation can be effectively visualized, supporting transparent AI-based analysis and practical decision-making in agricultural production.

In this paper, we propose a sensorless collision-detection method that exploits static redundancy and achieves robustness against unknown manipulation forces. The proposed method does not require additional sensors other than joint position or motor current sensors commonly used in industrial manipulators and collaborative robots. This method enables the detection of collisions at the terminal link under unknown manipulation force conditions that are difficult to detect using conventional methods. By extracting the null-space component of the Jacobian matrix from the joint driving torque, the method achieves low-latency detection with simple computation. Moreover, by limiting the task space in which the manipulation forces are applied according to the task, the proposed method can be applied to general-purpose manipulators. In this study, the effectiveness of the proposed method was validated through simulations using a planar 2-degree-of-freedom (2-DOF) manipulator and experiments using a vertical 6-DOF manipulator.

In this study, we propose a mobile robot that can start moving by utilizing the rotational movements of its two arms. The robot has two rotating arms and a body. It is equipped with a device that can fix its body to a platform constructed on a wall or floor. When the body is fixed to the platform, the robot can store the kinetic momentum of its center of mass by rotating its arms. When the body is released from the platform, the robot begins to move along the kinetic momentum of its center of mass. In our previous study, we demonstrated the feasibility of a mobile robot that hopped up and down the steps under Earth’s gravity, and we assessed its performance of the robot by performing numerical simulations. In the present study, we evaluated the performance of the mobile robot as it moved on a flat, low-friction floor, and we assessed the implementability of a device capable of fixing the robot’s body to a platform, by conducting a round-trip movement experiment. Furthermore, we proposed a modified mobile robot equipped with an oscillation mechanism and verified the feasibility of the modified mobile robot through comparative experiments.

In recent years, educational support robots that assist learning have attracted attention. This study focuses on robots that teach driving techniques. In driving schools, instructors teach drivers using a method that gradually teaches them about situations and actions. This teaching method is known as the GROW model. Previous studies have shown that collaborative learning with robots using the GROW model is an effective way of teaching driving behavior to university students. However, in today’s society, the number of accidents caused by elderly drivers is increasing, and the demand for teaching driving behaviors to the elderly is high. Therefore, it is important to verify the effectiveness of this approach for elderly drivers. This study investigated the effects of collaborative learning with a robot using the GROW model on elderly people. A comparative experiment was conducted using three groups: a robot with the GROW model, a conventional robot without the GROW model, and a learning system. The experimental results showed that compared to the conventional robot and learning system, the robot equipped with the GROW model was more memorable for the elderly.

Animals exhibit a variety of gait patterns that vary with locomotor speed and energy efficiency. Although a variety of gait patterns have been studied, the mechanism of gait generation has remained vague. In this study, we performed ground reaction force equalization for stable locomotion by developing a one-input, two-output leg-equipped robot with a differential gear. The developed robot can tolerate angular velocity differences and equalize the torque transmitted to the legs. To investigate the possibility of rhythm generation based on ground reaction force equalization, hardware experiments and mathematical analysis were conducted using a nonlinear system. The robot experiments and theoretical analysis show that the robot can spontaneously generate rhythm by equalizing the ground reaction force.

Screw propellers, while commonly used in underwater propulsion, suffer from issues such as entanglement, noise, and reduced visibility. In contrast, fish tail fin propulsion offers minimal environmental impact, superior obstacle avoidance, and high maneuverability in confined spaces, thus exhibiting excellent compatibility with underwater environments. Consequently, fish-like robots are considered promising for underwater exploration in disaster zones and ecological surveys. Although various fish-like robots have been developed, replicating the smooth, continuous streamlined shape of real fish using multi-link structures remains challenging. While flexible materials have been employed to create continuous structures, issues concerning body shape and surface integrity during swimming persist. Therefore, this study aims to develop a robot that maintains a continuous streamlined body shape using a silicone-based outer skin. The robot’s external shape is designed based on 3D scan data of Japanese horse mackerel, and a wire-driven bending mechanism with a flexible outer skin that prevents wrinkling during bending is developed. Swimming experiments demonstrated that the developed robot maintained a smooth and continuous body structure without wrinkles during bending, successfully replicating carangiform swimming, particularly the coordinated movement of the tail fin and body. Furthermore, the relationship between the robot’s swimming speed and tail fin frequency closely matched that of a real horse mackerel, confirming the achievement of efficient swimming.

Robotics is applied in various fields and thus robot components with various shapes and stiffness values are required. We previously developed a variable-stiffness and deformable link using a shape-memory alloy and the jamming transition phenomenon. The link can be fixed in an arbitrary shape and then restored to its initial shape via the shape memory effect. We previously attached a prototype link to a robot arm and evaluated its pick-and-place motion for various objects with different shapes and weights. However, as we used a lubricant to facilitate deforming the link, objects sometimes slipped off the link. Therefore, in this study, we propose a method for deforming the link without a lubricant. Another robot arm is added to press a mold onto the link to increase operational efficiency. We compare three deformation methods in terms of the time required to change molds, weight capacity, structural change during repeated motion, force required to deform the link, and positioning accuracy. The experimental results show that the weight capacity increased when the link was deformed without a lubricant. Moreover, similar to our previous study, changing the link shape to suit the target object improved positioning accuracy. Using the two-robot-arm system, the time required to change molds decreased by 94% compared to that in our previous study. Furthermore, the pressing force of the link during the fixing of the shape affected the contact length between the link and the object and the positioning accuracy.

In crop detection, ridge structures provide crucial cues for classifying crops and weeds. However, it is difficult to obtain ridge structures for unmanned ground vehicles which can capture images only within a narrow field of view. This study proposes a lightweight algorithm that enables a model to implicitly infer the ridge structure from plant-to-plant spatial relationships and sizes. An object detector first detects each plant. The resulting bounding boxes are treated as pairwise features in the nodes. Metainformation indicating whether two nodes share the same ID is combined with their geometric relationships and encoded as edge features. A graph attention network addresses these relationships to infer and propagate ridge-aware regularities. By understanding the structure only from object relationships, the method compensates for the information lost to the limited field of view without any explicit edge structure input. In the experiments wherein we deliberately introduced a domain shift between the training/validation sets and test set, the proposed method increased the baseline mAP₅₀ from 30.6% to 44.4%. This amounts to an increase of up to 13.8 percentage points. In addition, the proposed method requires only approximately 10 ms/frame on a Jetson AGX Orin to classify plants. This method acquires ridge structures internally without relying on external sensors or hand-tuned thresholds. Thus, it displays potential for in-field agricultural applications such as autonomous weeding.

In this study, we present a novel autonomous excavation method that achieves high efficiency under varying soil conditions. This method consists of two main steps, including first estimating the density of the soil and then generating an optimal excavation path based on the estimated density. The proposed method estimates soil density by taking advantage of the bulking phenomenon, which refers to an increase in the volume of excavated soil. This estimation relies solely on 3D point-cloud data obtained before and after excavation. Using the estimated soil density, an optimal excavation path is generated by applying a genetic algorithm in a physics simulator that replicates both the hydraulic excavator and the target ground. The algorithm explores a range of paths over multiple generations to find one that maximizes efficiency. The effectiveness of the proposed method was verified through simulations and field experiments. In particular, field experiments conducted in soft soil showed that the proposed method improved excavation efficiency by 27.7% compared with a baseline method using fixed parameters.