Picking point localization method based on semantic reasoning for complex picking scenarios in vineyards

The study addresses the challenges faced by agents in dynamic and uncertain environments, where decision-making is easily disrupted, task structures are difficult to maintain, and strategies lack continuous adaptability. It proposes an adaptive task decomposition and strategy updating method grounded in large model reasoning. The approach first introduces state modeling and semantic context encoding mechanisms that capture environmental non-stationarity, allowing the agent to acquire and integrate temporal information throughout long-term interactions. Building on this foundation, an adaptive task decomposition module dynamically generates hierarchical task structures through semantic reasoning, enabling the agent to preserve coherent execution even when task goals change, disturbances intensify, or feedback becomes incomplete. In parallel, the strategy updating mechanism adjusts decision distributions based on real-time feedback, allowing rapid recovery and stable behavior when action deviations, sudden scene changes, or state noise occur. These components are integrated into a unified closed-loop reasoning framework that equips the agent with structural understanding, behavioral adjustment, and robust execution capabilities in complex scenarios. Systematic evaluation across multiple key metrics demonstrates that the method improves task completion, planning consistency, error recovery, and decision stability, highlighting its potential value for complex tasks in diverse application domains.

Current infusion monitoring methods primarily rely on two technological approaches: non-visual sensor technology and visual sensor technology, for real-time monitoring of the remaining liquid volume in infusion bottles within infusion scenarios. However, non-visual sensor-based methods often suffer from complex installation procedures and are prone to external interference, while visual sensor-based methods tend to exhibit low detection accuracy in complex infusion environments involving small targets, low contrast, tilted objects, and partial occlusions, making it difficult to accurately monitor the remaining liquid. To address these challenges, we propose a high-precision and lightweight object detection algorithm-SDM-YOLO11n-based on an improved version of YOLO11n. Specifically, a lightweight Split-Parallel Convolution (SPConv) module is introduced to enhance the backbone network's spatial modeling capabilities and improve feature extraction efficiency; the traditional upsampling operation is replaced with a Dynamic Sampling (DySample) module for more adaptive feature reconstruction and multi-scale information fusion; and a Mixed Local Channel Attention (MLCA) mechanism is incorporated to strengthen attention to key regions of infusion bottles and their internal liquids, thereby further improving detection accuracy. In addition, a method based on the ratio of geometric parameters of oriented bounding boxes is proposed to precisely estimate the remaining liquid volume in infusion bottles. Experimental results show that SDM-YOLO11n improves mAP@0.5:0.95 by 0.6 percentage points compared to YOLO11n, with a model size of only 5.1 MB. The proposed algorithm achieves high-precision detection of infusion bottles and their internal liquids in complex scenarios and enables real-time monitoring of the remaining liquid volume in multiple infusion bottles.

Visual Question Answering (VQA) is a method of answering questions in natural language based on the content of images and has been widely concerned by researchers. The existing research on the visual question answering model mainly focuses on the point of view of attention mechanism and multi-modal fusion. It only pays attention to the visual semantic features of the image in the process of image modeling, ignoring the importance of modeling the spatial relationship of visual objects. We are aiming at the existing problems of the existing VQA model research. An effective spatial relationship reasoning network model is proposed, which can combine visual object semantic reasoning and spatial relationship reasoning at the same time to realize fine-grained multi-modal reasoning and fusion. A sparse attention encoder is designed to capture contextual information effectively in the semantic reasoning module. In the spatial relationship reasoning module, the graph neural network attention mechanism is used to model the spatial relationship of visual objects, which can correctly answer complex spatial relationship reasoning questions. Finally, a practical compact self-attention (CSA) mechanism is designed to reduce the redundancy of self-attention in linear transformation and the number of model parameters and effectively improve the model's overall performance. Quantitative and qualitative experiments are conducted on the benchmark datasets of VQA 2.0 and GQA. The experimental results demonstrate that the proposed method performs favorably against the state-of-the-art approaches. Our best single model has an overall accuracy of 71.18% on the VQA 2.0 dataset and 57.59% on the GQA dataset.

The plant leaf veins coupling feature representation and measurement method based on DeepLabV3+ is proposed to solve problems of slow segmentation, partial occlusion of leaf veins, and low measurement accuracy of leaf veins parameters. Firstly, to solve the problem of slow segmentation, the lightweight MobileNetV2 is selected as the extraction network for DeepLabV3+. On this basis, the Convex Hull-Scan method is applied to repair leaf veins. Subsequently, a refinement algorithm, Floodfill MorphologyEx Medianblur Morphological Skeleton (F-3MS), is proposed, reducing the burr phenomenon of leaf veins’ skeleton lines. Finally, leaf veins’ related parameters are measured. In this study, mean intersection over union (MIoU) and mean pixel accuracy (mPA) reach 81.50% and 92.89%, respectively, and the average segmentation speed reaches 9.81 frames per second. Furthermore, the network model parameters are compressed by 89.375%, down to 5.813M. Meanwhile, leaf veins’ length and width are measured, yielding an accuracy of 96.3642% and 96.1358%, respectively.

An intelligent diagnosis model for missing substation alarms based on an adaptive rule engine introduces dynamic threshold adjustment and feedback learning mechanisms on top of traditional rule matching and deep learning frameworks, enabling multi-source data fusion and real-time inference. The system architecture comprises four layers: data acquisition, rule engine, deep diagnosis, and visualization decision-making, constructed using attention networks and incremental weight update strategies. Experimental results demonstrate that under conditions of concept drift and noise disturbance, the model achieves a Recall_miss of 91.6%, representing an approximately 28% improvement over static rules. The F1 score increases to 92.1%, with a P95 latency of only 345 ms. Even under 800 TPS concurrency, the availability rate remains at 98.9%, significantly outperforming traditional methods.

Edge-aware and spectral–spatial information aggregation network for multispectral image semantic segmentation

Fruit picking in complex orchard environments is limited by the low detection and recognition accuracy due to clustered background, illumination variation, and partial occlusion. To improve the picking performance, a hierarchical visual grasping architecture based on You Look Only Once (YOLO) algorithm and adaptive error compensation is proposed. The upper layer uses YOLOv5 and inverse kinematics to recognize and localize the target. Ant colony optimization is specifically used for hyperparameter tuning of YOLOv5 to improve the detection accuracy. The middle layer dynamically compensates for the output torque of the joint actuator through the feedback linearization method. The lower layer finishes precise grasping through nonlinear mapping model between the pulse width modulation signal and the servo angle. Experimental results validate that the proposed architecture outperforms traditional methods by reaching 98.2% and 97.7% recognition accuracy in obstacle-free scenarios and complex environments, a higher grasping success rate, and a lower positioning deviation.

Liver disease being a major global health problem, with millions of patients progressing to final-stage liver failure and limited pool of donor organs[6]. Existing models for liver transplant decision-making, such as the MELD score, rely on a small number of biomarkers, do not fully capture disease complexity, and do not integrate imaging or automated donor– recipient matching[7]. This leads to delayed decisions, fragmented workflows, and potential inequities in organ allocation.This paper presents Orgdan, a human-centered, web-based platform that integrates liver risk assessment, automated medical image analysis, and intelligent donor–recipient matching into a single decision-support system optimized for practical clinical use.Orgdan combines three core components: (1) An eight-biomarker risk scoring model built from the Indian Patient Liver Disease (IPLD) dataset using logistic regression-derived weights[3] [5] (2) A lightweight computer vision pipeline for semi- automatic liver segmentation from CT and MRI scans using classical image processing [4] and (3) An AI-assisted donor matching module based on a ternary tree structure and Grey Wolf Optimization (GWO) tuned Random Forest, SVM and XGBoost [14]. Performance was evaluated using risk classification metrics, segmentation overlap scores, donor matching accuracy, and comparison with the MELD score. The biomarker-based risk model achieved an overall accuracy of 92% in classifying patients into low, medium and high liver risk categories, with a low false negative rate for high-risk patients. The liver segmentation module reached a mean Intersection over Union (IoU) of 0.87 and Dice coefficient of 0.93 across 50 CT/ MRI scans, with average processing time of 2.3 seconds per image on a standard workstation. The GWO optimized donor matching framework attained an Area Under the ROC Curve (AUROC) of 0.92, with 85–90% accuracy across tuned models, outperforming the traditional MELD-based baseline (AUROC 0.68). Orgdan demonstrates that a humanized, interpretable, and computationally efficient AI platform can augment clinical decision making in liver disease by bringing together risk scoring, imaging, and donor matching [13]. Compared with MELD, Orgdan uses twice the number of biomarkers, provides transparent explanations for predictions, and automates several currently manual steps. Future work will focus on validation using real transplant registry data, deeper integration into hospital information systems, and enhancement of the imaging module with advanced segmentation networks [11].

An Integrated GIS, optimization and simulation framework for optimal PV size and location in campus area environments

Surface defects directly affect the mechanical properties of industrial strip steel products. To evaluate the integrity of the strip steel surface, a channel-wise global Transformer-based dual-branch network (CGTD-Net) for strip steel surface defect detection, dubbed CGTD-Net, is proposed in this study. First, the strip steel surface images are preprocessed using saturation adjustment and random flipping strategies to remove unnecessary background information and improve network generalization. Second, the Swin Transformer is employed at the end of the backbone network and the negative impacts of a single channel are then mitigated by using the multichannel feature pyramid networks via Transformer, which improves the extraction ability of the global semantic information for tiny or narrow defects. Third, an edge detection branch network is constructed with a spatial–channel global attention (SCGA) module to further enhance the feature extraction on both spatial and channel information. Finally, the CGTD-Net is compared with 11 state-of-the-art methods on the NEU-DET dataset, and ablation experiments are also implemented. The comparison results, conducted on a single 3090Ti GPU, reveal that the CGTD-Net achieves a mean intersection over union (mIoU) of 75.16% at 178 frames/s, outperforming other methods. The ablation experiment demonstrates that the CGTD-Net improves the mIoU by 7.83% and the <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">${F}$ </tex-math></inline-formula> -score by 6.3% compared to the baseline.

It is critical for fire-rescuers to rapidly identify flammable and explosive materials around a burning building and accurately locate trapped individuals within the golden rescue period. To address this need, this paper proposes and develops FireSeer—an automated system for fire detection and unit-level localization. Deploying aerial vehicles, FireSeer assists rescue teams by quickly identifying burning units and their floor levels. Technically, the system determines the specific location of fire-affected floors through a three-stage processing pipeline consisting of a Finder, a Detector, and a Localizer. The Finder extracts key information, such as fire signatures and building corners, from onboard camera images, and isolates the building foreground via perspective transformation. Subsequently, the Detector identifies individual units on fire. In the Localizer stage, a floor classification method categorizes the detected units and outputs the specific unit number where the fire is located. Additionally, a prediction and optimization module corrects missed detections, false positives, and occlusions. Experiments using both synthetic and real-world data confirm the algorithm’s real-time feasibility on embedded devices. Results from typical scenarios show that the FireSeer prototype achieves a row prediction accuracy (RA) of over 97% and a column prediction accuracy (CA) of over 86%, with an average processing time of 2.72 ms—demonstrating the system’s high precision and lightweight efficiency.

In complex, semi-structured orchard environments, mobile inspection robots often suffer from excessive turning points, low search efficiency, limited trajectory-tracking accuracy, and poor adaptability to dynamic obstacles. To address these issues, this study proposes an integrated autonomous navigation method that employs an improved A* algorithm for global path planning, a Fuzzy-Weighted Dynamic Window Approach (FW-DWA) for local path optimization, and a model predictive control (MPC)-based trajectory-tracking controller. First, a dynamic heuristic-weight adjustment strategy is introduced into the conventional A* algorithm, in which a correction factor adaptively tunes the heuristic weight; a two-stage node optimization procedure then removes hazardous and redundant nodes to improve path smoothness and safety. Second, the FW-DWA, grounded in fuzzy control theory, uses goal distance and obstacle distance to update the weights of the heading, clearance, and velocity evaluation functions in real time, thereby enhancing obstacle avoidance in dynamic environments. Finally, a discrete kinematic model is established to design the MPC Controller, which achieves high-precision tracking through receding-horizon optimization and feedback correction. Experiments conducted in real orchards demonstrate that the proposed method reduces path length by 5.79%, shortens planning time by 3.64%, and increases the minimum safety distance by 50%. Comparative results further show that the MPC Controller attains a mean position error of 0.032 m and a mean heading error of 3.14°, clearly outperforming a conventional Proportional–Integral–Derivative (PID) controller. These findings provide an effective solution for reliable autonomous navigation of orchard inspection robots and offer a valuable reference for smart agricultural robotics applications.

Traditional DeepLabV3+ image semantic segmentation methods face challenges in pitaya orchard environments characterized by multiple interference factors, complex image backgrounds, high computational complexity, and extensive memory consumption. This paper introduces an improved visual navigation path recognition method for pitaya orchards. Initially, DeepLabV3+ utilizes a lightweight MobileNetV2 as its primary feature extraction backbone, which is augmented with a Pyramid Split Attention (PSA) module placed after the Atrous Spatial Pyramid Pooling (ASPP) module. This improvement enhances the spatial feature representation of feature maps, thereby sharpening the segmentation boundaries. Additionally, an Efficient Channel Attention Network (ECANet) mechanism is integrated with the lower-level features of MobileNetV2 to reduce computational complexity and refine the clarity of target boundaries. The paper also designs a navigation path extraction algorithm, which fits the road mask regions segmented by the model to achieve precise navigation path recognition. Experimental findings show that the enhanced DeepLabV3+ model achieved a Mean Intersection over Union (MIoU) and average pixel accuracy of 95.79% and 97.81%, respectively. These figures represent increases of 0.59 and 0.41 percentage points when contrasted with the original model. Furthermore, the model’s memory consumption is reduced by 85.64%, 84.70%, and 85.06% when contrasted with the Pyramid Scene Parsing Network (PSPNet), U-Net, and Fully Convolutional Network (FCN) models, respectively. This reduction makes the proposed model more efficient while maintaining high segmentation accuracy, thus supporting enhanced operational efficiency in practical applications. The test results for navigation path recognition accuracy reveal that the angle error between the navigation centerline extracted using the least squares method and the manually fitted centerline is less than 5°. Additionally, the average deviation between the road centerlines extracted under three different lighting conditions and the actual road centerline is only 2.66 pixels, with an average image recognition time of 0.10 s. This performance suggests that the study can provide an effective reference for visual navigation in smart agriculture.

In this article we introduce new approach for human body poses and movement sequences recognition. Our concept is based on syntactic description with so called Gesture Description Language (GDL). The implementation of GDL requires special semantic reasoning module with additional heap-like memory. In the following paragraphs we shortly describes our initial concept. We also present software and hardware architecture that we created to test our solution and very promising early experiments results.KeywordsPose recognitionmovement sequences recognitionsyntactic descriptionsemantic reasoningnatural interface

Path planning is a core technology for mobile robots. However, existing state-of-the-art methods suffer from issues such as excessive path redundancy, too many turning points, and poor environmental adaptability. To address these challenges, this paper proposes a novel global and local fusion path-planning algorithm. For global path planning, we reduce path redundancy and excessive turning points by designing a new heuristic function and constructing an improved path generation method. For local path planning, we propose an environment-aware dynamic parameter adjustment strategy, incorporating deviation and avoidance dynamic obstacle evaluation factors, thus addressing issues of local optima and timely avoidance of dynamic obstacles. Finally, we fuse those global and local path-planning improvements to form our fusion path-planning algorithm, which can enhance the robot's adaptability to complex scenarios while reducing path redundancy and turning points. Simulation experiments demonstrate that the improved fusion path-planning algorithm not only effectively addresses existing issues but also operates with higher efficiency.