IET Cybersystems and Robotics最新文献

The perception of Bird's Eye View (BEV) has become a widely adopted approach in 3D object detection due to its spatial and dimensional consistency. However, the increasing complexity of neural network architectures has resulted in higher training memory, thereby limiting the scalability of model training. To address these challenges, we propose a novel model, RevFB-BEV, which is based on the Reversible Swin Transformer (RevSwin) with Forward-Backward View Transformation (FBVT) and LiDAR Guided Back Projection (LGBP). This approach includes the RevSwin backbone network, which employs a reversible architecture to minimise training memory by recomputing intermediate parameters. Moreover, we introduce the FBVT module that refines BEV features extracted from forward projection, yielding denser and more precise camera BEV representations. The LGBP module further utilises LiDAR BEV guidance for back projection to achieve more accurate camera BEV features. Extensive experiments on the nuScenes dataset demonstrate notable performance improvements, with our model achieving over a $��4�\times �$ reduction in training memory and a more than $��12�\times �$ decrease in single-backbone training memory. These efficiency gains become even more pronounced with deeper network architectures. Additionally, RevFB-BEV achieves 68.1 mAP (mean Average Precision) on the validation set and 68.9 mAP on the test set, which is nearly on par with the baseline BEVFusion, underscoring its effectiveness in resource-constrained scenarios.

Accurate state estimation is a fundamental problem for autonomous robots. To achieve locally accurate and globally drift-free state estimation, multiple sensors with complementary properties are usually fused together. Local sensors (camera, IMU (inertial measurement unit), LiDAR, etc.) provide precise poses within a small region, whereas global sensors (GPS (global positioning system), magnetometer, barometer, etc.) supply noisy but globally drift-free localisation in a large-scale environment. In this paper, we propose a sensor fusion framework to fuse local states with global sensors, which achieves locally accurate and globally drift-free pose estimation. Local estimations, produced by existing visual odometry/visual-inertial odometry (VO/VIO) approaches, are fused with global sensors in a pose graph optimisation. Within the graph optimisation, local estimations are aligned into a global coordinate. Meanwhile, the accumulated drifts are eliminated. We evaluated the performance of our system on public datasets and with real-world experiments. The results are compared with those of other state-of-the-art algorithms. We highlight that our system is a general framework which can easily fuse various global sensors in a unified pose graph optimisation.

This paper addresses a crucial challenge in the domain of smart factories and intelligent warehouse logistics, focusing on conflict-free planning and the smooth operation of large-scale nonlinear mobile robots. To tackle the challenges associated with scheduling large-scale mobile robots, an improved space–time multi-robot planning algorithm is proposed. The cloud servers are adopted in this algorithm for computation, which enables faster response to the planning requirements of large-scale mobile robots. Furthermore, enhancements to a model-free adaptive predictive control method are proposed to enhance the networked control effectiveness of the nonlinear robots. The algorithm's capability to accommodate conflict-free path planning for large-scale mobile robots is demonstrated through simulation results. Experimental findings further validate the effectiveness of the cloud-based large-scale mobile robot planning and control system in achieving both conflict-free path planning and accurate path tracking. This research holds substantial implications for enhancing logistics transportation efficiency and driving advancements in the field of smart factories and intelligent warehouse logistics.

This paper proposes a trajectory planning approach based on the configuration space (C-space) generated from large-scale digital twinning. Leveraging GPU-based parallelism, the C-space of a multi-degree-of-freedom (multi-DoF) manipulator in a complex task space with obstacles can be mapped out through extensive simulation of motion and collision of multiple virtual robot arms known as digital twins. An optimal search algorithm is incorporated with artificial potential field generated in the C-space to allow the prioritising of safety in accordance with the varying risks associated with the obstacles by means of variable repulsive potential. To extend the high-degree path to smooth and continuous joint trajectories, a spline operation is applied. Finally, a 7-DOF physical manipulator is deployed for the execution of the planned trajectory in a task space filled with obstacles. Results demonstrated a 16.3% improvement in success rate achieved by utilising the safety-prioritisable search algorithm. With this unified formulation of the control and planning problem in the C-space, the kinematics complexity of a large DOF manipulator in obstacle-present task space could be truly relieved from the joint control loop. This simplification, in turn, opens up prospective work in dynamic reconstruction of the C-space.

Real-time detection of micro-defects in pumped storage generator stators and rotors remains challenging due to small-target obscurity and edge deployment constraints. This paper proposes EdgeFault-detection transformer (DETR), a lightweight transformer model that integrates three innovations: (1) dynamic geometric-photometric augmentation for robustness, (2) a FasterNet backbone with Partial Convolution (PConv) that reduces the number of parameters by 40% (from 20 × 10⁶ to 12 × 10⁶), and (3) cross-scale small-object head enhancing defect localisation. Experiments on 8763 industrial images demonstrate 75.38% [email protected] (+17.25% over RT-DETR) and 49.3% mAP_small (+42.9% from baseline). The model achieves 22 FPS on NVIDIA RTX A4000 GPUs (640 × 640 resolution), validating real-time industrial applicability. Strategic computation allocation increases GFLOPs (giga floating-point operations) by 16.4% (from 58.6 to 68.2) to prioritise safety–critical precision, justifying the trade-off for detecting high-risk anomalies (e.g., insulation cracks).

Acupoint therapy plays a crucial role in the prevention and treatment of various diseases. Accurate and efficient intelligent acupoint recognition methods are essential for enhancing the operational capabilities of embodied intelligent robots in acupoint massage and related applications. This paper proposes a lightweight hand acupoint recognition (LHAR) method based on middle finger cun measurement. First, to obtain a lightweight model for rapid positioning of the hand area, on the basis of the design of the partially convolutional gated regularisation unit and the efficient shared convolutional detection head, an improved YOLO11 algorithm based on a lightweight efficient shared convolutional detection head (YOLO11-SH) was proposed. Second, according to the theory of traditional Chinese medicine, a method of positional relationship determination between acupoints based on middle finger cun measurement is established. The MediaPipe algorithm is subsequently used to obtain 21 keypoints of the hand and serves as a reference point for obtaining features of middle finger cun via positional relationship determination. Then, the offset-based localisation approach is adopted to achieve accurate recognition of acupoints by using the obtained feature of middle finger cun. Comparative experiments with five representative lightweight models demonstrate that YOLO11-SH achieves an [email protected] of 97.3%, with 1.59 × 10⁶ parameters, 3.9 × 10⁹ FLOPs, a model weight of 3.4 MB and an inference speed of 325.8 FPS, outperforming the comparison methods in terms of both recognition accuracy and model efficiency. The experimental results of acupoint recognition indicate that the overall recognition accuracy of LHAR has reached 94.49%. The average normalised displacement error for different acupoints ranges from 0.036 to 0.105, all within the error threshold of ≤ 0.15. Finally, LHAR is integrated into the robotic platform, and a robotic massage experiment is conducted to verify the effectiveness of LHAR.

For the issue of low positioning accuracy in dynamic environments with traditional simultaneous localisation and mapping (SLAM), a dynamic point removal strategy combining object detection and optical flow tracking has been proposed. To fully utilise the semantic information, an ellipsoid model of the detected semantic objects was first constructed based on the plane and point cloud constraints, which assists in loop closure detection. Bilateral semantic map matching was achieved through the Kuhn–Munkres (KM) algorithm maximum weight assignment, and the pose transformation between local and global maps was determined by the random sample consensus (RANSAC) algorithm. Finally, a stable semantic SLAM system suitable for dynamic environments was constructed. The effectiveness of achieving the system's positioning accuracy under dynamic interference and large visual-inertial loop closure was verified by the experiment.

Trajectory planning of a robotic arm has a significant impact on its operational efficiency and success rate. However, due to the complexity of the environment and the vastness of the search space, it often ends up falling into local optima. In this paper, we propose a novel algorithm that combines particle swarm optimisation (PSO) with differential evolution (DE), namely the PSO-DE algorithm, to alleviate the problem. Firstly, the initial path of the robotic arm is represented by spline curves in the joint space. Then, the trajectory optimisation problem of the robotic arm is established, including constraints such as obstacle cost, acceleration cost, torque cost etc. Finally, the PSO-DE algorithm is proposed for optimisation, from which the PSO ensures the search space range through individual collaboration, whereas the DE generates new solutions through individual differentiation with local search. The combination of the two algorithms can fully leverage their respective advantages, ensuring the global optima within a large search space. Experiments are conducted in a simulation environment using the Python Robotics Toolbox and the PyBullet simulation platform. The results demonstrate that the proposed algorithm can effectively plan the trajectory of the robotic arm with significant improvements in success rates compared to the PSO algorithm.

Exploring unknown environments by multiple robots is promising but challenging. The challenges are posed not only by the coordination among multiple robots to improve exploration efficiency, but also by dynamic obstacles that suddenly appear on planned paths. To address those two challenges, this paper proposes a two-layer architecture where the high-level layer generates target locations for each robot to explore the unknown environment, while the low-level layer plans paths in the dynamic environment for each robot. Specifically, in the high-level design, a novel auction algorithm is proposed, which considers both the distance of robots to target locations and the number of frontiers within the clustering domain of target locations. This approach enables robots to explore different target locations while reducing redundant exploration compared to traditional exploration algorithms. In the low-level design, a neural network-based Q-learning algorithm is employed for path planning to achieve dynamic obstacle avoidance. Robots can dynamically adjust their actions through interaction with the external environment, thus avoid obstacles and reach the target position. To validate our methods, a series of simulation experiments are conducted. The experimental results demonstrate that robots can not only efficiently accomplish exploration tasks in unknown environments, but also achieve effective obstacle avoidance when faced with suddenly appearing dynamic obstacles.

Obstructive sleep apnoea (OSA) is a prevalent condition that can lead to various cardiovascular and cerebrovascular diseases, such as coronary heart disease, hypertension, and stroke, posing significant health risks. Polysomnography (PSG) is widely regarded as the most reliable method for detecting sleep apnoea (SA), but it is limited by a complex acquisition process and high costs. To address these issues, some studies have explored the use of single-lead signals, although they often result in lower accuracy due to noise-related information loss. Time context information has been applied to mitigate this issue, but it can lead to overfitting and category confusion. This paper introduces a novel approach utilising time sequence contrastive learning with single-lead electrocardiogram (ECG) signals to detect SA events and assess OSA severity. The proposed method features a Transformer encoder fusion module and a contrastive classification module. First, a multi-branch architecture is employed to extract features from different time scales of the ECG signal, which aids in SA detection. To further enhance the network's focus on the most relevant extracted features, a channel attention mechanism is incorporated to fuse features from different branches. Finally, contrastive learning is utilised to constrain the features, resulting in improved detection performance. A series of experiments were conducted on a public dataset to validate the effectiveness of the proposed method. The method achieved an accuracy of 91.50%, a precision of 92.06%, a sensitivity of 94.37%, a specificity of 86.89%, and an F1 score of 93.20%, outperforming state-of-the-art detection techniques.