To address the challenges of navigation path planning in unmanned aerial vehicle (UAV)-based pipeline inspection missions arising from factors including long pipeline distances and dynamically varying installation environments, this paper proposes a visual navigation method for UAVs based on pipeline self-contour reconstruction. First, a mirror-based field-of-view conversion device is utilized to perform perspective transformation of the forward-facing camera's navigation view. The body-eye transformation is then computed using the camera's structural parameters. Second, a fully convolutional network is employed for pipeline region segmentation and edge contour extraction in complex backgrounds. Third, leveraging the extracted pipeline contour features as input, pipeline self-contour reconstruction is realized via the constant diameter pipeline perspective projection model. Finally, UAV navigation control is accomplished by mapping virtual pipeline axes to real-world coordinates through the body-eye transformation. Experimental results from computer simulations and indoor/outdoor experiments validate the proposed method. It facilitates ready-to-deploy autonomous vision-guided navigation for pipeline inspection UAVs without the need for pre-planned trajectories, offering improved operational efficiency and flexibility in UAV-based pipeline inspection scenarios.

In recent years, The Vision-Language-Action (VLA) models have shown significant promise for applications in embodied agricultural tasks. To handle the complexities of deploying harvesting robots in orchards with a variety of crop types, we aim to address several key challenges, such as the one-fruit-one-strategy development dilemma, poor generalization capability, high hardware costs, and insufficient operational stability under complex unstructured disturbances. We propose a compact and easily deployable general dexterous harvesting framework based on the integration of VLA model and Reinforcement Learning. The framework takes the lightweight SmolVLA model as its backbone and introduces the Proximal Policy Optimization (PPO) algorithm to perform online finetuning on the initial imitation learning policy. We systematically explores the transfer value of pre-trained vision-language knowledge to agricultural manipulation tasks. With a focus on the pick-and-place harvesting tasks, the model's generalization capabilities are assessed across three dimensions: Vision, Semantic, Execution. We conducted experiments under various challenging conditions, including new backgrounds, dynamic lighting, unseen fruit types, multi-object scenes, and variations in initial poses. Result show that the framework improves the adaptability and stability of harvesting robots under the above conditions. In real-world trials, the model achieves a task success rate over 80%, validating its practical applicability and in dexterous fruit picking. This method in this paper effectively reduces the hardware requirements and deployment costs of harvesting robots. It provides a practical high-generalization and low-cost solution for resource-limited small-scale orchards.

Designing and analyzing compact micro rotation stages are key issues in precision engineering. The serial connection of a compliant motion-scaling mechanism to a rotary microdriving component enhances the actual rotation resolution. However, the existing compliant rotation-reduction mechanism (CRRM) requires the rotation centers of multiple flexure hinges within the mechanism to remain consistently aligned in a straight line, which adversely affects miniaturized applications. In this study, a novel CRRM that does not require the consistent collineation of multiple rotation centers was designed, analyzed, and tested. Based on the deflection principles of cantilevers, a conceptual design was first outlined. Small-deflection-based models for the output and input rotations of the CRRM and its rotation gain were established, and the conditions for rotation reduction were derived. Geometrically nonlinear models of the CRRM were developed using the beam constraint model. Subsequently, a nonlinear parametric optimization was performed by considering a dynamic constraint. Finally, the proposed CRRM, its static and dynamic models, and its optimization were validated via finite-element analysis and experiments.

Accurate perception of the environment is paramount in robotic manipulation, driving an increasing reliance on multimodal sensing systems. Vision and tactile sensing emerge as two pivotal modalities; however, prevailing fusion strategies, often based on direct feature concatenation, fail to adequately model the complex spatiotemporal interdependencies between them. This limitation becomes pronounced in dynamic interaction scenarios, where the inability to capture fine-grained visuotactile correlations severely restricts the perception of critical events like object slippage. To address this challenge, we propose a novel multimodal fusion framework based on Tucker decomposition. Our approach introduces three key innovations: independent factor matrices that decouple visual and tactile features into separate latent spaces to mitigate modal confusion; a cross-core tensor fusion strategy leveraging a shared core tensor to explicitly model high-order cross-modal interactions; and an efficient tucker tensor decomposition fusion (TTDF) module, built upon a Transformer architecture, which incorporates hierarchical fusion and dynamic gated attention to adaptively recalibrate modal contributions, improving generalization and robustness. Extensive experiments demonstrate superior performance, achieving state-of-the-art accuracy in robotic slippage detection. The framework enhances visuo-tactile fusion precision, reduces object damage rates in production, increases operational efficiency, and provides crucial technical support for advanced flexible robotic manipulation.

With the rapid development of intelligent manufacturing technologies, factors such as the rapid updating of products and the complexity of humanoid robot collaborative scenarios have put forward higher requirements for the dynamic obstacle avoidance and adaptive control of industrial assembly robots. Although deep reinforcement learning offers an effective solution for robot assembly, it typically requires long training time and massive data to achieve the desired performance. However, transfer learning can enhance sample efficiency yet suffers from insufficient generalization in new scenarios. To tackle the above issues, this study proposes a pre-trained policy in the source environment to accelerate target policy training by sharing action-space features and policy network structure during transfer. Additionally, the observation space in the target environment is adjusted by concatenating robot state information with obstacle state information, capturing more critical information to address the mismatch between the observation spaces of the source and target environments. The core innovation is to first train a suboptimal policy in a simple, obstacle-free source environment for basic assembly, then transfer and fine-tune it in highly dynamic target environments involving humans and humanoid robots. This method reduces accidental collisions and enables reliable assembly in complex dynamic scenes. Experimental results demonstrate that our approach significantly improves policy learning efficiency, accelerates adaptation to new obstacle-avoidance tasks, and achieves faster convergence and stronger generalization compared with existing methods.

Cable-driven hyper-redundant manipulators (CDHRMs) have significant potential for application in narrow space exploration and minimally invasive surgery owing to their high degrees of freedom and flexibility. However, existing designs generally suffer from insufficient stiffness, limited load capacity, and poor positioning accuracy. To address these challenges, we propose a bio-inspired CDHRM modeled after the biomechanical properties of the human spine. The modular design integrates bionic spinal segments and spherical joints to balance compliance with a high payload capacity and enhanced stiffness. For accurate control, we developed a hybrid kinematic model combining a Piecewise Constant Curvature (PCC) approach with quaternion representation. Additionally, to solve high-dimensional inverse kinematics (IK), we propose an improved Particle Swarm Optimization (PSO) algorithm incorporating a k-dimensional tree (KDT). Simulations and prototype experiments validate the proposed system. Compared to conventional methods, our KDT-guided PSO algorithm reduces the average IK solving error by 57

Grasp-induced damage during mechanical harvesting critically compromises the postharvest quality and market value of table grapes. To address this, this study developed a multiscale finite element model that explicitly incorporates the mechanical heterogeneity and elastoplastic behavior of grape skin and flesh, enabling high-fidelity simulation of damage initiation and propagation. The model was calibrated and validated through independent grasping experiments, establishing a reliable equivalent stress threshold of 0.05 MPa for damage onset, with a relative prediction error below 2 %. Systematic analysis using the validated model elucidated the coupling effects of grasping speed, contact force, and contact material properties on internal damage. Results indicate that contact force is the dominant influencing factor, followed by grasping speed via strain rate effects, while the elastic modulus of the contact material regulates stress distribution and energy dissipation, with silicone rubber exhibiting optimal damage mitigation due to its compliant load-dispersing nature. Furthermore, a high-precision regression model (R-2 > 0.91) was developed based on simulation data to predict damage susceptibility, providing a quantitative tool for the optimization of harvesting end-effector design and operational parameters. This work enhances the understanding of grasp-induced damage mechanisms and provides directly applicable insights for developing low-damage robotic harvesting systems.

Stable grasping operations play a pivotal role in enabling robotic dexterity. To achieve a level of proficiency akin to humans, robots must effectively integrate visual and tactile information throughout their grasping tasks. However, relying solely on visual information may not enable robots to achieve dexterous operations. To address this issue, this study explores the application of visual–tactile fusion in robot grasping operations and proposes the YGC model. Specifically, we decompose the grasping operation into three subtasks: target detection, pose generation, and grasping state detection. Firstly, in order to enhance the feature extraction capabilities of the YOLOv5s backbone network, we propose the integration of a multi-scale feature fusion (MSFF) module, which replaces the existing C3 module. In pose generation, we compare various input channels to identify the optimal input method. For grasping state detection, we evaluate the performance of single-modality and multi-modality inputs. Secondly, experimental results show that by replacing the MSFF module, the mean average precision (mAP) value improves by 3.59

View of the inadequate performance of YOLOv8 in detecting small-sized and low-sample targets in industrial connector inspection scenarios, this paper proposes an optimized YOLOv8 model that integrates the C2fSE and CBAM dual-attention mechanisms. Firstly, the original C2f structure is replaced with the C2fSE module to achieve adaptive calibration of channel features across the entire network. Secondly, the CBAM module is embedded at the end of the Neck to enhance the joint spatial and channel expressive capabilities of deep-level features. Experimental validation is conducted using a self-made industrial connector dataset, and the results show that the optimized model achieves a 4.82% improvement in mAP50-95 compared to the original YOLOv8 baseline. While maintaining overall detection accuracy, it achieves stable and high-precision detection for the RJ45 port category, which has the fewest samples. Furthermore, compared to a failed model with CA attention inserted across all layers, the optimized model achieves a 5.12% improvement in mAP50, effectively addressing the performance collapse caused by the insertion of a single attention mechanism across all layers.

Enabling robots to understand natural language and locate referred objects for grasping remains a key challenge. Language-guided visual grounding connects visual perception and language understanding. As a fine-grained setting, Referring Image Segmentation (RIS) further provides pixel-level masks, which are particularly useful for precise grasping. However, existing RIS methods still face difficulties in robotic scenarios. Convolution- and transformer-based models are often limited by restricted receptive fields or quadratic complexity. Meanwhile, Mamba-based models provide efficient long-sequence modeling, but can still suffer from information decay and loss of spatial details. To address these issues, we propose MambaGLR, an efficient Mamba-centered hybrid framework for language-guided visual grounding, instantiated on RIS and tailored for robotic grasping. MambaGLR adopts a stage-specific cross-modal design: a global-local fusion module is applied in early high-resolution stages to capture both long-range dependencies and local spatial details, while a detail-guided cross-modal refinement module explicitly introduces cross-attention in later low-resolution stages to compensate for information decay and strengthen vision-language alignment. In addition, we construct RefGrasp, a grasp-oriented RIS dataset, and establish a unified benchmark with OCID-VLG and RoboRefIt to support future research in robotic visual grounding. Extensive experiments show that MambaGLR achieves strong grounding accuracy with a favorable efficiency-accuracy trade-off, while laboratory tabletop robot experiments demonstrate its feasibility under the evaluated service-oriented grasping settings. The dataset is publicly available at [https://github.com/xiaozheng-liu/MambaGLR]

In real orchard environments, harvesting robots often operate in dense foliage, narrow inter-row spaces, and under complex, dynamically changing illumination. These conditions limit the robot's ability to obtain wide, continuous viewpoints, with common visual features between sequential frames often obscured by environmental disturbances. This leads to inherently weak-covisibility inputs, characterized by unstable spatial features due to alternating direct sunlight and shadows, small overlapping regions caused by hierarchical occlusions from branches and leaves, and low consistency in morphological and spatial feature distributions across different fruit species. Consequently, robots struggle to build a coherent understanding of orchard scenes and fruit targets, which limits performance in fruit volume estimation, yield assessment, fruit localization, and harvesting path pl anning.To address these challenges, this study analyzes the limitations of existing point cloud registration methods, particularly their reliance on initial poses, feature alignment, and error convergence. We propose a novel point cloud reconstruction framework that integrates iterative semantic cues and geometric priors into a unified optimization mechanism. By jointly constraining geometric priors and semantic segmentation, the proposed method establishes a dynamic iterative optimization process that significantly enhances the stability and accuracy of point cloud registration and surface reconstruction under weak-covisibility conditions. Quantitative comparisons demonstrate that the proposed method achieves a registration RMSE of 1.02 mm under weakcovisibility conditions. Field experiments in real orchard environments further demonstrate robust reconstruction under illumination variations, occlusions, and robotic motion, achieving a fruit reconstruction error of 8.34% and an average reconstruction time of 1.87 s per fruit.This work provides effective support for 3D modeling and autonomous operation of agricultural robots.

During the cultivation and growth of green plums, various defects frequently occur, potentially affecting their overall quality and economic value. Accurate classification and identification of these defects have become essential components of the harvesting process, particularly when employing smart agricultural equipment. These defects pose significant challenges to the yield and quality of green plums, making their precise detection crucial for ensuring optimal output and economic efficiency. However, most contemporary research on fruit defect classification and grading using artificial intelligence techniques primarily focuses on accuracy, often neglecting the constraints imposed by limited resources. This study addresses the aforementioned challenges by employing knowledge distillation techniques to optimize the performance of a lightweight model. Specifically, during the knowledge distillation process, the vision transformer model, known for its robust recognition capabilities, was selected as the teacher model. The lightweight MobileNetv3 model, chosen for its ease of deployment, served as the student model and was trained using the Lion optimizer. In addition, the dual guidance learning module was designed to enhance knowledge transfer between the teacher and student models, thereby improving the overall capability of the student model. Experimental validation demonstrated that the proposed method excels in the green plum defect recognition task, with the student model, MobileNetv3, achieving an accuracy of 99.17% and exhibiting high performance in key metrics such as precision, recall, and F1-score. Notably, MobileNetv3 not only delivers exceptional performance but also features a low parameter count and computational complexity, facilitating its efficient deployment in practical applications. This study provides an effective and practical solution for the automatic identification and sorting of green plum defects, significantly advancing the development and application of smart agricultural technologies.

Fresh table grapes are typically harvested by cutting the stem. However, due to the small size of the fruit stem and its frequent occlusion by leaves, robots often struggle to locate the cutting point, leading to harvesting failures. To address this issue, the paper proposes a view planning method based on self-supervised deep reinforcement learning. The method enables the robot to dynamically adjust camera viewpoints through sequential actions generated by a policy network, ultimately locating the fruit stem. Unlike methods that rely on time-consuming 3D reconstruction, a reward function is designed directly on 2D images by incorporating the occluded regions’ area and position. To improve the training efficiency of the action policy network, a learning strategy based on a Self-Supervised Convolutional Network (SSCN) is introduced. Under this strategy, the effectiveness of actions produced by the policy network is evaluated by the SSCN, and the immediate reward is dynamically updated to guide the policy network learning. The output and target of the self-supervised learning loss are predicted online by the SSCN and automatically generated by the Action Evaluation Algorithm (AEA), respectively. To verify the performance of the proposed method, comparative experiments are conducted on both a multi-view grape dataset and in real-world outdoor environments. Experimental results show that the proposed method significantly outperforms baseline methods regarding grape stem detection success rate and network training efficiency, demonstrating its effectiveness and feasibility.

Detecting grape stems is essential for the autonomous operation of grape-picking robots. In natural orchards, the complex and irregular positioning of grape bunches, along with frequent occlusions caused by obstacles and leaves, create a dynamic and unpredictable environment that substantially affects the robot’s perception quality and harvesting efficiency. Inspired by human active observation mechanisms, this study introduces an end-to-end active visual perception framework using deep reinforcement learning (DRL) to enhance the detection of grape stems under complex occlusion. An instance segmentation network is trained to obtain 2D masks, which are integrated into an octree volumetric grid to produce detailed volumetric observations for DRL training. Different occupancy weights are assigned to both explored and newly discovered regions, contributing to a novel reward function based on information gain, which subsequently drives the network to optimize camera movements, guiding the robotic arm towards the best viewpoint for efficient grape stem detection. The superior performance of the proposed method has been validated in both laboratory and outdoor orchards settings. The primary contribution of this work lies in presenting a novel fully end-to-end detection framework for occluded grape stems. Compared to existing methods, it enables the system to directly learn optimal viewpoint strategies through interactions between the robotic arm and the environment, effectively handling feature extraction and environmental modeling without the need for manually designed information gain metrics. This advancement provides foundational support for the next generation of highly autonomous picking robots, capable of operating adaptively in complex, unstructured environments.

Understanding the rheological deformation behaviors of clustered fruits during harvesting through contact force modeling remains a critical challenge. This study presents a dynamic model that combines the viscoelastic constitutive properties of fruit berries with multi-body contact interactions, developed through biomechanical experiments and multi-scale computational approaches. Axial and radial compression tests (loading rate: 40-80 mm/min) were used to quantify the mechanical damage thresholds for grapes (longitudinal: 18.9-44.1 N, transverse: 10.0-31.4 N). The stress-deformation characteristics of fruit berries were systematically analyzed using the generalized Maxwell viscoelastic model. Based on Hertz's contact theory, a multi-level contact chain model for manipulator-to-fruit and fruit-to-fruit interactions was established, and finite element simulations showed that the rheological deformation induces significant asymmetric stress transmission characteristics and hierarchical transmission laws. The gradient decay rate of stress transmission from the directly contacted area to the indirectly contacted area was 47.3 %, which accumulates non-linearly with contact force and contact time. This study overcomes the limitations of traditional single-fruit models, providing new theoretical foundations for precise force control strategies in the harvesting of clustered fruits such as grapes.

The USB-A interface is widely used in electronic devices, however, the current USB cable quality inspection process heavily relies on manual labor, in order to achieve automated testing of USB cables, the first step is to achieve USBA insertion assembly. This paper proposes a USB-A insertion assembly strategy that combines visual, force feedback, and pose information to address the uncertainty of initial contact pose. The entire assembly process uses a hybrid force/position control scheme based on admittance control to realize the interaction between the robot and the environment. First, The method for adjusting the position between the peg and hole was determined by analyzing the assembly clearance, and the search operation was completed by combining the search trajectory generated by vision. Then, by attempting to insert action perception orientation and using a “lift-rotate” action for orientation adjustment. Finally, an assembly experimental platform is built to verify the effectiveness of the proposed assembly strategy. The experimental results show that the success rate of the USB-A assembly strategy proposed in this paper is 100 %.

In the complex orchard environment, precise picking point localization is crucial for the automation of fruit picking robots. However, existing methods are prone to positioning errors when dealing with complex scenarios such as short peduncles, partial occlusion, or complete misidentification, which can affect the actual work efficiency of the fruit picking robot. This study proposes an enhanced picking point localization method based on semantic reasoning for complex picking scenarios in vineyard. It innovatively designs three modules: the semantic reasoning module (SRM), the ROI threshold adjustment strategy (RTAS), and the picking point location optimization module (PPOM). The SRM is applied to handle the scenarios of grape peduncles being obstructed by obstacles, partial misidentification of peduncles, and complete misidentification of peduncles. The RTAS addresses the issue of low and short peduncles during the picking process. Finally, the PPOM optimizes the final position of the picking point, allowing the robotic arm to perform the picking operation with greater flexibility. Experimental results show that SegFormer achieves an mIoU (mean Intersection over Union) of 84.54%, with B_IoU and P_IoU reaching 73.90 % and 75.63 %, respectively. Additionally, the success rate of the improved fruit picking point localization algorithm reached 94.96 %, surpassing the baseline algorithm by 8.12 %. The algorithm's average processing time is 0.5428 +/- 0.0063 s, meeting the practical requirements for real-time picking. (c) 2025 The Authors. Publishing services by Elsevier B.V. on behalf of KeAi Communications Co., Ltd. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).

Defect detection based on computer vision is a critical component in ensuring the quality of industrial products. However, existing detection methods encounter several challenges in practical applications, including the scarcity of labeled samples, limited adaptability of pre-trained models, and the data heterogeneity in distributed environments. To address these issues, this research proposes an unsupervised defect detection method, FLAME (Federated Learning with Adaptive Multi-Model Embeddings). The method comprises three stages: (1) Feature learning stage: this work proposes FADE (Feature-Adaptive Domain-Specific Embeddings), a framework employs Gaussian noise injection to simulate defective patterns and implements a feature discriminator for defect detection, thereby enhancing the pre-trained model’s industrial imagery representation capabilities. (2) Knowledge distillation co-training stage: a multi-model feature knowledge distillation mechanism is introduced. Through feature-level knowledge transfer between the global model and historical local models, the current local model is guided to learn better feature representations from the global model. The approach prevents local models from converging to local optima and mitigates performance degradation caused by data heterogeneity. (3) Model parameter aggregation stage: participating clients utilize weighted averaging aggregation to synthesize an updated global model, facilitating efficient knowledge consolidation. Experimental results demonstrate that FADE improves the average image-level Area under the Receiver Operating Characteristic Curve (AUROC) by 7.34% compared to methods directly utilizing pre-trained models. In federated learning environments, FLAME’s multi-model feature knowledge distillation mechanism outperforms the classic FedAvg algorithm by 2.34% in average image-level AUROC, while exhibiting superior convergence properties.

Smart agriculture leverages information technology and IoT to achieve automation, intelligence, and precision in agricultural production. Harvesting robots, as crucial components, must accurately understand orchard environments and crop damage to optimize crop management and harvesting strategies, enabling efficient, low- loss, and large-scale operations. However, current semantic segmentation models often fail to meet the high-efficiency requirements in orchards, particularly in segmenting small-pixel targets, resulting in misidentification, missed detection, and segmentation distortion. To address these issues, we propose an improved segmentation network, BoT-PSPNet, for precise segmentation of orchard roads, trunks, fruits, and their damage. Specifically, we use PSPNet as the primary framework and integrate the Multi-Head Self-Attention (MHSA) module from Transformer into the ResNet50 Bottleneck. This approach overcomes PSPNet's limitations in handling long-range dependencies. The MHSA module provides robust global feature extraction capabilities and, through its self-attention mechanism, effectively captures long-range dependencies, addressing PSPNet's deficiencies in multi- scale information fusion, computational efficiency, and flexibility. We collected a dataset of 1,000 orchard images, including road conditions and various fruit damages, and conducted validation experiments. The results demonstrate that the proposed method outperforms traditional PSPNet models in segmentation accuracy, particularly in handling complex backgrounds and detailed features. This study provides a new methodological support for enhancing the visual perception accuracy of scenes and targets in agricultural environments.

Feature extraction is a critical step in point cloud registration networks, determining how effectively point cloud data is represented. The feature extraction module in RPMNet uses fixed-range local features as input, whereas DGCNN dynamically constructs adjacency graphs and uses graph convolutions for feature updates, better capturing both local and global features of the point cloud. This study aims to explore the application of the Dynamic Graph Convolutional Neural Network (DGCNN) in RPMNet's feature module to address the inadequacies in local and global feature extraction and improve its registration accuracy in point cloud data processing, particularly on the ModelNet40 dataset. The study conducted training and testing on the ModelNet40 dataset, comprising 5124 training point clouds, 1198 validation point clouds, and 1245 test point clouds. Through performance comparison analysis, various evaluation metrics such as rotation error and translation error were used to assess the model's performance. The results show that the DGCNN-enhanced RPMNet reduced the isotropic rotation error from 0.056 to 0.053. This indicates that applying DGCNN to RPMNet can dynamically capture local and global features of point clouds, improving feature representation accuracy and model robustness. These findings are significant for the field of point cloud data processing, validating the effectiveness of DGCNN and providing new directions for future research. This advancement promotes the application of graph neural networks in practical problems, enhancing the technical level of the related field.