Computed Tomography (CT) is currently the mainstream method for reconstructing bronchial environments. Although this technique can reveal the internal structure of the airways, it provides relatively limited information, leading to the neglect of many useful details. This paper proposes a novel bronchial environment semantic reconstruction algorithm that integrates geometric constraints with region growing. The aim is to address the limitations of traditional reconstruction methods, which often fail to capture sufficient airway detail. Clinically, the diameter of surgically relevant bronchial structures typically exceeds 2 mm, yet conventional CT-based reconstructions struggle to preserve fine features at this scale. By combining region growing with geometric constraint-based segmentation of the airway, our method achieves high-precision semantic bronchial modeling. Experimental results using ex vivo porcine lungs demonstrate that the preoperative semantic bronchial point cloud models reconstructed by our method provide surgeons with richer intraoperative information and achieve a reconstruction accuracy of up to 1 millimeter.
To address issues in underwater images, such as blurriness, low contrast, and color distortion, we propose a Generative Adversarial Network (GAN) based on a hybrid attention mechanism and Transformer, named HAT-GAN. HAT-GAN is designed to efficiently recover image details and improve visual quality. The proposed framework incorporates a Multi-Domain Attention Transformer, which processes query information in the frequency domain while handling key-value information in the spatial domain to generate attention maps with high information density. Additionally, in the up-sampling and down-sampling modules, different attention mechanisms are applied at different sampling stages to acquire richer information. Finally, in the discriminator, a Dual Attention mechanism is employed, enabling the discriminator to focus on critical regions of the image. Through experimental comparisons and ablation studies, we show that our approach successfully mitigates issues such as image detail blur and color distortion in underwater images, producing visually pleasing results.
Robotic welding is a critical component of industrial automation, with weld seam detection playing a vital role in ensuring both accuracy and efficiency in welding operations. While precise laser sensors and point cloud cameras are commonly employed to improve welding detection accuracy, current methods often fall short of meeting the speed requirements essential for industrial applications. This paper proposes an efficient multi-weld seams detection method that leverages CAD-guided image segmentation to achieve fast and accurate detection. Our approach employs computationally efficient image segmentation to detect weld seams, in contrast to conventional point cloud processing methods, which involve heavy computational overhead. The detected weld seams in image space are then mapped to the 3D point cloud, enabling the final robot-executable welding paths. Furthermore, to enhance detection accuracy, we introduce a background removal method guided by the CAD model of the workpiece, providing clean data for the image segmentation process. Extensive experiments conducted on the real-world robotic welding system validate the proposed method, demonstrating significant improvements in both accuracy and efficiency compared to existing state-of-the-art techniques.
In this study, we focused on the thoracic joint of a click beetle and proposed a two-link type snap mechanism. This mechanism mimics the click beetle's principle for storing and releasing elastic energy, and is composed of links, springs, motors, viscoelastic peg, etc. This mechanism normally functions as a general active joint, and can generate snap rotational motion at any time. The developed snap mechanism was likened to a click beetle and the relationships between the mechanism's parameter and jumping characteristics were validated. As a result, we confirmed that the size of the mechanism, weight distribution, and hardness of the ground affect the jumping performance.
Semantic segmentation of aerial images is vitally important to many aspects of autonomous driving. However, when applied to the segmentation of large-scale aerial imagery acquired from diverse geographic regions, pre-trained deep learning models often fail to produce consistently accurate predictions. To address this challenge, this paper introduces a novel unsupervised domain adaptation (UDA) method leveraging entropy-driven attention for the semantic segmentation of aerial images. The entropy-driven attention strategy contains a two-stage adversarial learning process, which utilizes the entropy map to explicitly measure the data distribution distance between two domains and then guides the model to focus on the poorly aligned features. We conducted extensive experiments on the LoveDA dataset to validate the effectiveness of our proposed UDA method. The quantitative results indicate that our approach surpasses six other state-of-the-art UDA methods, achieving superior performance.
This study comparatively evaluated two advanced object detection models—Detectron2 and YOLOv11—for strawberry segmentation tasks. We utilized a custom dataset consisting of 551 images of 30 strawberries of various sizes captured through wide-angle, close-up, and individual shots. Both models were trained under similar computational conditions, using an NVIDIA RTX 4090 GPU, except that Detectron2 was trained for 3450 iterations and YOLOv11 for 50 epochs. The models were assessed using multiple metrics: mAP@50, mAP@50:95, accuracy, recall, F1 score, and inference time. Our results showed that YOLOv11 outperformed Detectron2, achieving superior segmentation accuracy, recall, F1 score, and significantly faster inference time. Specifically, YOLOv11 achieved 96% mAP@50:95, 99.4% mAP@50, and 98.4% F1 score, which were better than those of Detectron2. Furthermore, YOLOv11 realized a shorter inference time (9.1 ms) than Detectron2 (5.52 s). An ablation study on varying dataset sizes (25%, 50%, 75%, and 100%) was also conducted to evaluate each model’s robustness under limited data scenarios—a critical factor in agricultural applications. This analysis provides new insights into data efficiency and highlights YOLOv11’s superior adaptability. These results highlight that YOLOv11 is the preferred model for real-time segmentation tasks in complex environments, offering higher segmentation precision and efficiency than Detectron2.
This research proposes an improved RFNN-EKF (Recurrent Fuzzy Neural Network-Extended Kalman Filter) algorithm and integrates it with Robot Operating System 2 (ROS2) to achieve autonomous navigation for physical mobile robots in obstacle-filled environments. While previous studies have validated the feasibility of the RFNN-EKF algorithm through simulations, physical testing introduces additional challenges such as motor dynamics, sensor noise, and environmental disturbances. To address these issues, this study modifies the RFNN-EKF algorithm to better suit real-world applications. The system employs a distributed architecture: A PC handles the computationally intensive RFNN-EKF algorithm and processes data from LiDAR and IMU, while an ESP32-based microcontroller manages motor control and speed monitoring. Communication between the PC and microcontroller is facilitated through Micro-ROS, ensuring real-time data transmission. Experimental results demonstrate that the proposed system achieves effective path planning and obstacle avoidance in dynamic environments, enhancing the autonomy and adaptability of mobile robots.
Depth estimation serves as the foundation for many computer vision applications. As an emerging perception sensor inspired by biological retinas, event cameras have garnered significant attention in recent years due to their high dynamic range, ultra-high temporal resolution, low latency, and other distinctive features.This paper provides a systematic review of depth estimation methods based on event cameras, including related datasets, evaluation metrics, and future development prospects. In terms of methodological classification, the approaches are divided into frame-based methods and learning-based methods (the latter further subdivided into Spiking Neural Networks (SNN) and Artificial Neural Network (ANN)) based on sensor type. Additionally, the supervision modes (supervised/unsupervised) and open-source availability of these methods are examined in detail, summarizing the algorithmic principles, performance, and contributions of various approaches. Subsequently, commonly used datasets and measurement techniques in event camera-based depth estimation are systematically organized. Finally, the development prospects of this field are discussed. We hope this paper can offer a comprehensive survey and serve as an introductory reference for researchers new to this field.
Self-balancing lower limb exoskeletons (SBLLE) have significant potential in mobility assistance and rehabilitation applications. Providing exoskeletons with stable turning capabilities is critical for their practical application. In this paper, we propose an effective control framework designed to enable stable turning motions in SBLLE without reliance on external assistive devices. By modeling the exoskeleton as a Cart-Table system, we develop trajectory generation methods and impedance controller design strategies specifically tailored for turning maneuvers. To evaluate the stability and dynamic performance of the system under practical operating conditions, we conducted comparative experiments involving two scenarios: only executing the gait planning trajectories and incorporating impedance-controller corrections. Additionally, we performed stabilized turning trials with a male subject. The results demonstrate notable advances in the achievement of autonomous and balanced turning movements, highlighting the effectiveness of the proposed framework.
Convolutional Neural Networks (CNNs) have revolutionized deep learning, particularly for image-related tasks. Despite their success, training CNNs remains challenging, particularly in ensuring convergence in dynamic, real-time environments like robotics and autonomous systems. Traditional training methods, such as backpropagation, often lack theoretical convergence guarantees, making stable learning difficult in practical applications. This paper proposes a novel collaborative learning framework for CNNs that enables end-to-end training while ensuring theoretical convergence. By integrating intermediate representations dynamically, the proposed approach enhances feature refinement, leading to improved generalization and accuracy. The collaborative learning method is validated on benchmark datasets, demonstrating improvements in classification accuracy and incremental learning stability compared to forward progressive learning method.
Autonomous robot-assisted minimally invasive surgery is a key development direction to achieve precision medicine and reduce operator fatigue, while the core functions of existing surgical robotic systems are still confined to the master-slave remote-control mode, which makes it difficult to realize task autonomy for fine operations such as suturing and knotting. Especially in surgical suturing scenarios, real-time tracking and 3D reconstruction of suture lines face multiple challenges. To this end, we propose a suture line pose estimation method based on a diffusion probabilistic model. Considering the multi-source noise interference in the surgical scene, we first obtain the segmentation mask of the suture line and complement it as the feasible domain for the pose estimation, and add the start and end constraints to ensure the accuracy of the pose estimation. Subsequently, we model the task of resolving the geometric structure of the suture line as a path-planning problem under conditional constraints. The acquired feasible domains are encoded as global observations, and the suture trajectories are iteratively optimized by a denoising diffusion probabilistic model (DDPM), which ultimately leads to the pose estimation of the sutures. We conduct extensive experimental evaluations on a multi-scene dataset with sutures of different topologies, and the proposed method maintains the accuracy of the results in both non-critical occlusion and self-crossing cases. Based on this, further experiments on 3D reconstruction of sutures are performed and results are presented to validate the possibility of the present method to provide crucial a priori information for robotic surgical autonomy.
Recently, few-shot semantic segmentation (FSS) has gained prominence due to the high costs of data annotation. While meta-learning frameworks are commonly used for feature fusion in FSS, they often yield insufficient accuracy due to scene differences. To address this problem, we propose two innovative approaches. First, we introduce a base learner with background-guided features. This additional learner recognizes targets in the base class individually and uses these predictions as background-guided features for the final target. We also propose a GramAttention module to mitigate scene differences between support and query images during background prediction. Second, we enhance the meta-learner with a separate target category selected enhancement branch. This branch utilizes known information from the support set to enhance target features, reducing background interference and improving the model’s generalization ability. We evaluated our method on the Cityscapes-3i dataset, a few-shot outdoor dataset derived from the Cityscapes database, which emphasizes semantic understanding of urban street scenes. Our experiments verified the effectiveness of the proposed approaches in improving accuracy and generalization of FSS in urban environments.
Master-slave robotic systems hold significant potential to enhance safety and precision in vascular interventional surgery, which have extremely high demands for operational accuracy force feedback sensitivity. However, existing systems face challenges, including signal distortion caused by mechanical vibrations, the limited adaptability of traditional threshold-based methods to complex vascular environments, and fixed mapping strategies fail to meet dynamic operational requirements. To address these issues, this paper proposes an occlusion detection method for robotic vascular intervention. The method employs hybrid cascaded filtering to suppress signal noise of the interventional force and leverages delivery force-insertion depth differentiation analysis to detect guidewire blockages in complex interventional environments. Experimental results validate the feasibility of the proposed framework in improving surgical safety. The method effectively mitigates noise interference in force feedback signals, enhances force feedback precision,The RMSE for axial force was 414.15 mN, and the radial torque error measured 1.01 N·m. And achieves reliable detection of guidewire blockages in intricate anatomical scenarios, the detection success rate is 90% thereby elevating both procedural safety and operational precision.
The traditional five-axis machining path is typically planned in the workpiece coordinate system (WCS). However, when this path is mapped to the machine coordinate system (MCS), challenges such as singularities and abrupt movements of the rotary axis may arise, significantly impacting machining quality and efficiency. To tackle these issues, we propose a path planning method based on Physics-Informed Neural Networks (PINN) that generates an efficient tool path with iso-scallop height and free from singularities. In this approach, the surface is discretized into a triangular mesh, using its vertices as the input for the PINN, while the output are four scalar fields. To train the proposed PINN, we formulate four loss functions aimed at optimizing the iso-scallop height, ensuring path smoothness, addressing the kinematics of the machine tool, and guaranteeing the feasibility of the machining process. The tool path generated by our PINN demonstrates iso-scallop height and singularity-free performance, exhibiting significantly superior smoothness in both the tool path and the rotary motion of the machine tool compared to the existing benchmark.
To address the challenging issue of online fault discrimination in unmanned aerial vehicle (UAV) navigation systems, this paper introduces an innovative fault detection method that leverages spatial motion coordinates. This approach performs online synchronization, linear interpolation and smoothing on the spatial motion coordinate position data. By calculating the position, velocity and acceleration of the given data and selecting the fault discrimination thresholds adaptively through a grid search algorithm, the fault discrimination are fast realized. Then, the simulation date randomly generated is employed to validate the procedure. It turns out that the proposed method discriminate effectively anomalies for the UAV navigation systems in real time based on a single spatial motion coordinate data, such as position, velocity, or acceleration information, which improves the accuracy and stability of the system of the interest. The innovative method for fault discrimination demonstrates high sensitivity and robustness, providing theoretical and technical support for the safety of UAV flight
In order to realize the balance and path tracking control of unmanned motorcycle without self-balancing auxiliary device. In this paper, a coupling control scheme is proposed, which uses auxiliary function and reference model to realize balance and path tracking coupling control only by controlling handlebar steering. Firstly, LADRC is designed to realize the balance control of unmanned motorcycle by controlling the angular speed of handlebar steering. Secondly, the expected roll angle of the motorcycle is generated based on the reference model, and the nonlinear relationship between the yaw error and the lateral position error is combined to design the nonlinear auxiliary function, which realizes the path tracking control of the motorcycle by controlling the steering angle of the handlebar. Finally, the convergence of the linear extended state observer (LESO) and the nonlinear auxiliary function is proved theoretically, and the proposed scheme is simulated. The simulation results show that the proposed control scheme has a good tracking effect on both the classical path and the true path, which further verifies the effectiveness and robustness of the proposed scheme.
Developing a high-precision respiratory motion prediction model is crucial to optimize radiotherapy accuracy. This paper proposes a respiratory motion prediction method based on a Long Short-term Memory (LSTM) network with a weighted attention mechanism. Using the temporal modeling capability of LSTM and the adaptive weight allocation strategy of the attention mechanism, the proposed model effectively captures the nonlinear dynamic characteristics of respiratory signals. Experimental results demonstrate that, compared to conventional prediction methods, the proposed approach achieves lower prediction errors across multiple datasets and exhibits strong generalization capability, providing a novel technical solution for high-precision real-time tracking in radiotherapy.
Cyclic stability is always highlighted in bipedal dynamic walking. Therefore, convergence to limit-cycle gaits is often expected. From another perspective, the most fundamental requirement of achieving stable biped walking is, however, do not fall down, which desires a large basin of attraction (BOA) to tolerate disturbances. Chaotic oscillators tend to possess a larger BOA compared to the limit-cycle. With this consideration, the chaotic gait of the bipedal walker in this study is generated via entrained to a typical chaotic oscillator, e.g., the Rössler attractor. We explore a large stable walking domain of the chaotic bipedal gait by varying its control parameters. The numerical simulation results show that the chaotic bipedal gait owes a larger BOA, compared to limit-cycle walking. Our study not only shows the positive side of chaos in dynamic walking with an extremely simple and efficient control method, but also promises to provide new ideas and solutions for the motion control of robots.
The wide application of robotics technology has been observed in this era, which made the lives of individuals using it convenient. However, the rapid evolution of cyber threats, especially in phishing attempts, has influenced cybercriminals to make it an attack target. To solve this problem, we proposed real-time machine learning (ML)-based techniques to effectively detect phishing activities in a robotic environment. Stochastic Gradient Descent (SGD) and Passive-Aggressive algorithms are used to enable an accurate real-time phishing detection. Even though the dataset is static, our proposed methodology has attempted to simulate prediction in real-time through processing the test samples in sequence, treating each instance received as if it is arriving in real-time. The performance of our ML-based real-time phishing detection in a robotics environment has shown promising results, achieving over 99.8% for both SGD and passive-aggressive classifiers.
Extensive workspace traversal by multiple robots may compromise planning efficiency due to over-exploration and redundant calculations in inefficient regions, while trying to ensure the optimality of motion planning in multi-robot systems (MRS). This paper proposes a Voronoi-inspired random tree (VRT*), to improve planning efficiency in MRS by mitigating inefficient workspace exploration and reducing redundant computations. In particular, generalized Voronoi graph (GVG) is used to transform workspace into searchable objects. Subsequently, a voronoi guide tree is constructed to capture the connectivity and optimization potential of workspace, enabling the identification of critical workspace and facilitating the planning of heuristic paths. Through integrating the Expansion direction selection strategy into a discrete rapidly-exploring random tree, optimal solutions can be achieved by VRT* while minimizing redundant computations. Robust theoretical proofs and extensive experimental validations show that VRT* efficiently focuses computational resources on critical workspace in MRS planning, rapidly capturing high-quality initial solutions and reliably converging to optimal path.
