Accurate detection of aircraft skin surface defects is critical for aviation safety. However, conventional networks suffer from high-frequency edge and texture degradation during downsampling when detecting small-object defects, and achieving structural compactness remains challenging. To address these issues, a lightweight dual-domain wavelet feature-enhanced defect detection algorithm, DEDW-YOLO, is proposed. First, a Dual-Domain Wavelet Downsampling (DDWD) module synergistically couples frequency-domain Haar wavelet transforms with spatial-domain space-to-depth convolutions to mitigate small-object information loss. Second, the C3K2_DE module integrates Distribution Shifting Convolution and Efficient Channel Attention to minimize computational redundancy while preserving feature extraction capabilities. Furthermore, the network architecture is reconfigured alongside the introduction of a Lightweight Shared Convolutional Detection Head (LSCD) to optimize small-object localization precision. Finally, the WIoUv3 loss function enhances bounding box regression quality. Experimental results on a self-constructed skin defect dataset demonstrate that DEDW-YOLO achieves an mAP@0.5 of 87.5% (a 2.6% improvement over the baseline), while reducing parameters, model size, and computational cost by 83.8%, 71.7%, and 37.5%, respectively, successfully balancing high-precision small-object detection and model lightweightness.

With the booming development of the air transportation industry, manual visual detection can no longer meet the requirements of rapid and accurate aircraft skin detection. Specifically, manual visual detection is highly subjective, with high risk of omission and low efficiency. The inspectors are susceptible to fatigue and environmental factors that lead to safety hazards. To address these issues, utilizing aircraft skin inspection robots to replace manual labor is a viable option to improve detection efficiency and accuracy. Firstly, this review describes several typical aircraft skin defects types and non-destructive testing (NDT) methods. Secondly, it reviews the application of unmanned aerial vehicles (UAVs), ground mobile robots, and wall-climbing robots in aircraft skin detection, and analyzes their respective advantages and disadvantages. Next, the vision detection technology in practical applications is analyzed and summarized. Finally, the development prospects of aircraft skin inspection robots are discussed from the perspectives of robotics technology and NDT methods for skin defects. In the future, the enhanced integration of multi-modal information fusion, large-scale model, and digital twin technology into aircraft skin maintenance practices will significantly advance the automation of aircraft skin defects detection.

Gait recognition, which identifies individuals by walking patterns from afar, stands as a key video surveillance technology. Methods generally fall into silhouette-based and skeleton-based types, though progress has been made, recognizing camouflaged individuals in complex environments remains a challenge. Silhouettes are susceptible to environmental or clothing influences and contain only external contours. Skeletons comprise internal structure information but ignore body shape. To address this, a novel gait recognition framework based on multimodal feature learning is proposed in this paper to learn rich gait representation. Firstly, a multi-branch network based on the hybrid fusion strategy is introduced to leverage the complementarity of two modalities. Secondly, the silhouette branch utilizes multi-scale convolution with diverse kernel sizes for appearance feature extraction across various receptive fields, while a pyramidal attention mechanism is integrated to highlight critical information within these features. Moreover, a residual graph convolutional network is utilized to obtain hierarchical information on body joints in the skeleton. Considering the strengths and weaknesses of each modality, this paper utilizes the cross-modal transformer to obtain comprehensive gait features in the fusion branch. Ultimately, multi-branch gait features are integrated to enhance the overall recognition performance. Additionally, we build a novel gait dataset of camouflaged individuals captured by the unmanned aerial vehicle for research. It achieves a Rank-1 accuracy of 89.2 % on GCI-150, 96.5 % on CASIA-B, and 90.6 % on OUMVLP respectively. Particularly for clothing changes in complex environments, our method obtains performance gains compared to other methods. For instance, it achieves 94.5 % accuracy for coat-wearing walking conditions on CASIA-B, thereby proving its potential for military applications.

Skin defect detection is critical to aircraft safety and maintenance, as its accuracy directly impacts flight safety and aircraft lifespan. Traditional automated detection methods depend on substantial labeled data. However, the scarcity of real defect samples, high collection costs, and lack of diversity easily lead to model overfitting, which is difficult to meet the engineering needs. To address this problem, this paper proposes CHPDNet, enabling accurate defect segmentation with just one labeled sample. Firstly, the feature masking enhancement module (FMEM) is introduced to alleviate the loss of fine-grained spatial information and improve the segmentation accuracy for small defects by dual-channel mask complementation. Moreover, we generate reliable priori masks based on CLIP and use the optimized correlation matrix to enhance the mask quality, which enhances the model's adaptability and robustness to skin defects. In addition, the hierarchical purification decoder (HPD) is designed to utilize a progressive strategy from coarse to fine segmentation to reduce the missed and false detections. Extensive comparative experiments demonstrate that CHPDNet achieves superior performance on both ASD-2i and MVTec-5i datasets under 1-shot settings, with mIoU scores of 60.13 % and 62.31 %, respectively. The code is available at: https://github.com/lhpzfsc/CHPDNet.

Autonomous surface inspection of large-scale aircraft presents a unique challenge: balancing high-resolution coverage with stringent safety constraints imposed by complex, non-convex 3D geometries. Conventional multi-UAV coverage path planning (CPP) methods typically treat the aircraft as a generic mesh, overlooking structural regularities and resulting in suboptimal and erratic trajectories. To address this, we propose the Hierarchical Coordinated Path Optimizer (H-CPO), a hierarchical framework that explicitly leverages aircraft geometric priors—specifically symmetry and curvature saliency—to guide the planning process. The framework comprises three hierarchical levels: (1) Saliency-Aware View Planning, which employs a curvature-weighted sampling strategy to generate a compact, high-coverage viewpoint set; (2) Symmetry-Guided Coordination, where a Symmetry-Aware Collaborative Task Allocation (SA-CTA) algorithm exploits the airframe's bilateral symmetry to decompose the search space and balance workloads; and (3) Corridor-Constrained Trajectory Refinement, utilizing a Geometry-Guided Variable Neighborhood Search ($\boldsymbol{G}^{2}$-VNS) for topologically consistent sequencing, followed by a corridor-constrained convex optimization (CG-MSR) to enforce safety clearance constraints. Simulations on a C919 model show H-CPO achieves 99.3 % surface coverage and reduces jerk by 33 % (via CG-MSR) over unconstrained minimum-snap baselines, significantly enhancing trajectory smoothness and flight stability for imaging.

In this paper we propose a Twin Delayed Deep Deterministic Policy Gradient (TD3) algorithm with a prioritized buffer and memory that uses Long Short-Term Memory Networks (LSTMs) and use it as a dynamic controller for a differential drive mobile robot to accomplish a trajectory tracking task. We accomplish this by combining LSTM-TD3 with the Loss Adjusted Approximate Actor Prioritized Experience Replay (LA3P) to form LA3P-LSTM-TD3. We train the dynamic controller while varying the dynamic parameters and introducing a variety of random disturbances to improve the robustness and generalization ability of the dynamic controller. The reinforcement learning based dynamic controller operates alongside a well-established kinematic controller that provides reference velocities, with the dynamic controller compensating for dynamic effects, unmodeled dynamics, parameter variations, and disturbances. Simulation results demonstrate that our approach improves control accuracy and outperforms vanilla TD3.

Accurate identification of skin defects is crucial to the aircraft skin inspection. However, datasets are scarce, which significantly hampers the effectiveness of deep learning (DL) models for detecting aircraft skin defects. To tackle this issue, a generative model named Fourier GAN is proposed, which can generate realistic skin defect images under limited samples to optimize the detection model’s performance. Firstly, the transformer-based high frequency spectral discriminator (HFSD) is designed to fully utilize the frequency-domain information, aiming to minimize the high-frequency discrepancy and improve model stability. Furthermore, the high frequency calibration (HFC) modules utilize high frequency information of real images to improve the generator’s ability to generate correct high frequency features, which helps generate vivid defect images. The experimental results indicate that Fourier GAN can synthesize high-quality and diverse skin defects images. The precision of the detection model is significantly enhanced after using Fourier GAN as the data augmentation method.

This paper investigates the problem of adaptive tracking control for a class of uncertain multi-input and multi-output nonlinear systems in the presence of asymmetric input constraints and external disturbance. In order to address the different action ranges of input signals in asymmetric dead-zone and saturation models, an adaptive backstepping control method related to asymmetric parameters is designed. A state-dependent upper bound of uncertainty is proposed instead of a constant upper bound. This avoids the problem of state constraints caused by the boundness of uncertainty before obtaining closed-loop characteristics. Certain positive results are emerged in this paper where an innovative adaptive control methodology is demonstrated to cope with system uncertainty. The proposed controller does not require a priori knowledge on the bound of them. By means of the Lyapunov stability theory, the close-loop system is proven to be uniformly ultimately bounded, the system states converge to a domain containing the origin, and the output tracks the reference signal commendably. Simulation examples are presented to show the effectiveness of the proposed control method.

Mobile robot path planning in complex environments remains a significant challenge, especially in achieving efficient, safe and robust paths. The traditional path planning techniques like DRL models typically trained for a given configuration of the starting point and target positions, these models only perform well when these conditions are satisfied. In this paper, we proposed a novel path planning framework that embeds Large Language Models to empower mobile robots with the capability of dynamically interpreting natural language commands and autonomously generating efficient, collision-free navigation paths. The proposed framework uses LLMs to translate high-level user inputs into actionable waypoints while dynamically adjusting paths in response to obstacles. We experimentally evaluated our proposed LLM-based approach across three different environments of progressive complexity, showing the robustness of our approach with llama3.1 model that outperformed other LLM models in path planning time, waypoint generation success rate, and collision avoidance. This underlines the promising contribution of LLMs for enhancing the capability of mobile robots, especially when their operation involves complex decisions in large and complex environments. Our framework has provided safer, more reliable navigation systems and opened a new direction for the future research. The source code of this work is publicly available on GitHub.

To address the problems of poor dynamic performance and tracking overshoot of existing control strategies for industrial robot arms, this paper proposed a feedforward compensation-based nonlinear model predictive control (FNMPC) framework to achieve better tracking performance. First, the nonlinear frictional force of a six-degree-of-freedom robot manipulator is obtained by utilizing the least squares parameter identification method, and a more accurate dynamic model of the manipulator is established via the Newton-Euler method. Second, an FNMPC strategy is designed to achieve accurate tracking control of the robot manipulator. Finally, the performance of the proposed FNMPC strategy is verified by conducting a semi-physical simulation comparison experiment. Semi-physical simulation results show that the proposed control strategy exhibits superior performance in terms of dynamic response stationarity and cumulative tracking error.

Inspection of skin defects plays an important role in flight safety. Accurate detection of skin defects is critical for aircraft skin defect detection robots. The performance of deep learning-based defect detection models is hindered by the scarcity of aircraft skin defect images, thereby posing challenges for aircraft skin defect detection. In this paper, Fourier generating adversarial network (GAN) is proposed, which generates high quality images of aircraft skin defects under limited data. Firstly, the high frequency discriminator is crafted to minimize the high frequency discrepancy between the synthetic and real image. Furthermore, a high frequency calibration module is introduced, which uses the discriminator to guide the generator to synthesize the correct high frequency signals to make the synthesized defect images more realistic. By augmenting the defect dataset with the Fourier GAN, the precision of the detection models can be significantly improved. Finally, an aircraft skin defect dataset is constructed. After augmenting the dataset with Fourier GAN, the mAP of YOLOv8-l can reach 94.28%, which is 3.89% higher than the original dataset. By augmenting the dataset with Fourier GAN can effectively improve the detection quality of aircraft skin detection robots.

As a unique behavioral feature that can be obtained at distance, gait plays an essential role in special target surveillance, security prevention, and more. If gait recognition can be deployed on unmanned aerial vehicles (UAVs), it will open novel research perspectives for special operations. However, existing researches are dominated by ground view video recognition in constrained scenes rather than aerial view. In addition, gait recognition in complex environments remains challenging due to the similarity of targets and backgrounds, especially in scenarios where individuals intentionally hide identities through camouflage. To address these problems, we utilize the UAV carrying the infrared camera for gait acquisition and propose a novel gait recognition framework of camouflaged people. Firstly, an infrared gait dataset captured by the UAV flying at different altitudes and perspectives is constructed in this paper, named IR-150. Then, a novel hierarchical framework that integrates the advantages of CNN and Transformer is designed for gait recognition, named CTGait. Specially, depth-wise convolution and residual structures are applied to the local aggregation unit and feed-forward networks to introduce the locality mechanism. By means of diagonal masking and learnable temperature parameters, the attention mechanism is redesigned to enhance the recognition accuracy further. Various experiments prove that the CTGait is superior to the state-of-the-art methods, which achieves an average Rank-1 accuracy of 97.74 % and 82.08 % on the datasets CASIA-C and IR-150, respectively. Tests on the public datasets CASIA-B and GREW demonstrate that the CTGait exhibits significant robustness even in complex scenarios.

This article studies an adaptive backstepping barrier Lyapunov function-based control scheme for a class of multi-input and multi-output nonlinear systems in the presence of time-varying asymmetric full-state constraints and external disturbances. Simultaneously, a prominent feature of these systems is the unknown time-varying control direction. To stabilize such systems, Nussbaum gain technique is constructively framed to overcome the unknown control direction problem. To prevent that the constraints are overstepped, the time-varying asymmetric barrier Lyapunov functions are employed in each step of the backstepping design. The design involves an adaptive-based online approximator to cope with unknown dynamics of the system. A simulation example on aircraft skin inspection robot with three-dimensional trajectory is given to show the effectiveness of the proposed control scheme.

Aircraft skin defect detection is one of the crucial steps in the regular maintenance. However, the performance of defect detection algorithms is limited by insufficient data, which makes a challenge for automated defect detection of aircraft skin. Although image generation techniques relieve the problem of insufficient defect data, they are not effective under the condition of limited original data. To tackle this issue, this article proposes an image generation method that can obtain high-quality images of aircraft skin defects under limited data. First, high-frequency skip connection (HSC) is performed in the generator to enhance the utilization of frequency information. Moreover, high-frequency features fusion blocks (HFFBs) are utilized to effectively improve the high-frequency awareness of the discriminator. Meanwhile, adaptive high-dimensional diffusion (AHD) is introduced into the discriminator, which can automatically adjust the diffusion intensity according to the degree of discriminator overfitting. Finally, an aircraft skin defect dataset is constructed and a Generation-Poisson blending data augmentation method is proposed. The mAP of you only look once (YOLOX) trained using the augmentation dataset can reach 78.55%, which is 5.27% higher than the original data. The proposed method is capable of generating realistic and diverse images of aircraft skin defects, which significantly enhances the precision of defect detection networks in various scenes.

This study proposed a novel approach for mobile robots path planning and avoiding collisions by using Collision Probability (CP) along with the Soft Actor-Critic Lagrangian (SACL-L) framework. Our approach enables the mobile robot to dynamically deal with static and dynamic environments while ensuring safety and efficiency. The proposed SAC-L (CP) aims to minimize the total costs, which is the combination of both negative rewards and collision occurs. This dual focus strategy ensures trajectory planning inherently safer and providing a robust solution for complex dynamic obstacles environments. The framework’s efficiency is validated through extensive simulations on the Gazebo platform involving three increasingly difficult scenarios, demonstrating superior performance, adaptability and safety of our approach compared to traditional Deep Reinforcement Learning (DRL) methods. Our results showcase significant improvements in social and ego safety scores, contributing to the advancement of autonomous navigation in complex environments. This framework marks a step towards safer, more reliable mobile robot navigation and opens new avenues for future research in mobile robot path planning. A supplementary video further demonstrates the effectiveness of our framework. 1 .

PurposeAutonomous flight of unmanned aerial vehicles (UAVs) in global position system (GPS)-denied environments has become an increasing research hotspot. This paper aims to realize the indoor fixed-point hovering control and autonomous flight for UAVs based on visual inertial simultaneous localization and mapping (SLAM) and sensor fusion algorithm based on extended Kalman filter.Design/methodology/approachThe fundamental of the proposed method is using visual inertial SLAM to estimate the position information of the UAV and position-speed double-loop controller to control the UAV. The motion and observation models of the UAV and the fusion algorithm are given. Finally, experiments are performed to test the proposed algorithms.FindingsA position-speed double-loop controller is proposed, by fusing the position information obtained by visual inertial SLAM with the data of airborne sensors. The experiment results of the indoor fixed-points hovering show that UAV flight control can be realized based on visual inertial SLAM in the absence of GPS.Originality/valueA position-speed double-loop controller for UAV is designed and tested, which provides a more stable position estimation and enabled UAV to fly autonomously and hover in GPS-denied environment.

Fasteners defects in aircraft skin can seriously threaten the operational safeness of the aircraft. Therefore, the periodic detection of defects on aircraft skin fasteners is necessary. Due to the small size of aircraft skin fastener defects, traditional defect detection algorithms have lower detection precision. Therefore, we propose YOLO-FDD, an effective network for detecting aircraft skin fastener defects. Firstly, four detection layers and residual spatial pyramid pooling (R-SPP) module are used in the network. Secondly, an attention fusion feature pyramid networks (AF-FPN) is designed to adaptively fuse features and purify semantic conflicts. In addition, the attention head with swin transformer block (STB) and coordinate attention (CA) is designed as the feature detection module to enhance the location precision of defects. Finally, a dataset comprising of four types of aircraft skin fastener defects is constructed. According to the experimental findings, YOLO-FDD achieved a mAP of 83.08%, and the detection speed is 30.44 FPS, the YOLO-FDD network has great potential for application in aircraft skin fastener defect detection.

This article considers the flexible switching control problem of a two-frame aircraft skin inspection robot (TFASIR) with full-state time-varying constraints, input saturation, uncertainty, and unknown disturbance. Initially, this control problem is also treated as a tracking control problem of the dual-coupled adsorption system (DCAS). A novel nonlinear time-varying state-dependent function (NTVSDF) is first designed to tackle the full-state constraint problem. Subsequently, a feedforward tracking control method is designed, that uses the command-filtered backstepping technique, to transform the tracking control problem into an equivalent differential game problem (DGP) of closed-loop systems. Then, a zero-sum game strategy is presented, that uses the idea of adaptive dynamic programming (ADP) algorithm, to determine the DGP. The whole control method ensures that the closed-loop signals are uniformly ultimately bounded (UUB). Furthermore, another problem is that the partial system states are not accessible.To overcome this problem, a high-gain observer is utilized to reconstruct the state vector, and an output feedback controller is developed. The feasibility of the proposed control scheme is demonstrated in simulation.

针对金属部件表面损伤点云数据对分割网络局部特征分析能力要求高,局部特征分析能力较弱的传统算法对某些数据集无法达到理想的分割效果问题,选择采用相对损伤体积等特征进行损伤分类,将金属表面损伤分为6类,提出一种包含空间尺度区域信息的三维图注意力特征提取方法.将得到的空间尺度区域特征用于特征更新网络模块的设计,基于特征更新模块构建出了一种特征更新的动态图卷积网络(Feature Adaptive Shifting-Dynamic Graph Convolutional Neural Networks)用于点云语义分割.实验结果表明,该方法有助于更有效地进行点云分割,并提取点云局部特征.在金属表面损伤分割上,该方法的精度优于PointNet++、DGCNN(Dynamic Graph Convolutional Neural Networks)等方法,提高了分割结果的精度与有效性.

Unmanned aerial vehicles(UAVs) technology and target detection technology are increasingly integrated. However, the target detection networks are cumbersome and complex, and it is hard to directly deploy these networks on the UAVs. To address this problem, a Lightweight-YOLOv5s which can be applied to the UAVs is proposed to detect the camouflaged personnel in the woodland. Depthwise separable convolutions are added to the YOLOv5s for reducing the parameters and computational effort. Then a new channel pruning method based on Batch Normalization(BN) layer scaling factor with adaptive sparsity training is proposed for network compression, which can balance the compression ratio and accuracy of the detection. The pruned network is then fine-tuned to detect the camouflaged personnel. In addition, a special dataset is built for the camouflaged personnel detection in the woodland. The quantitative results on this dataset prove that the proposed Lightweight-YOLOv5s greatly improve the YOLOv5s. Parameters from 7.06M to 1.86M, floating point operations from 16.4G to 9.9G and inference time from 8.16ms to 7.06ms. Besides, mAP of the Lightweight-YOLOv5s is 0.83, which shows the excellent detection performance of the proposed method.