Low weight and good toughness thin plate parts are widely used in modern industry, but its flexibility seriously impacts the machinability. Plenty of studies focus on the influence of machine tool and cutting tool on the machining errors. However, few researches focus on compensating machining errors through the fixture. In order to improve the machining accuracy of thin plate-shape part in face milling, this paper presents a novel method for compensating the surface errors by prebending the workpiece during the milling process. First, a machining error prediction model using finite element method is formulated, which simplifies the contacts between the workpiece and fixture with spring constraints. Milling forces calculated by the micro-unit cutting force model are loaded on the error prediction model to predict the machining error. The error prediction results are substituted into the given formulas to obtain the prebending clamping forces and clamping positions. Consequently, the workpiece is prebent in terms of the calculated clamping forces and positions during the face milling operation to reduce the machining error. Finally, simulation and experimental tests are carried out to validate the correctness and efficiency of the proposed error compensation method. The experimental measured flatness results show that the flatness improves by approximately 30 percent through this error compensation method. The proposed method not only predicts the machining errors in face milling thin plate-shape parts but also reduces the machining errors by taking full advantage of the workpiece prebending caused by fixture, meanwhile, it provides a novel idea and theoretical basis for reducing milling errors and improving the milling accuracy.

To address the issues of intense image noise interference and computational intensity faced by traditional vision-based weld tracking systems, we propose WeldLight, a lightweight and noise-resistant convolutional neural network for precise classification and positioning of welding seam feature points using single-line structured light vision. Our approach includes (1) an online data augmentation method to enhance training samples and improve noise adaptability; (2) a one-stage lightweight network for simultaneous positioning and classification; and (3) an attention module to filter features corrupted by intense noise, thereby improving stability. Experiments show that WeldLight achieves an F1-score of 0.9668 for seam classification on an adjusted test set, with mean absolute positioning errors of 1.639 pixels and 1.736 pixels on low-noise and high-noise test sets, respectively. With an inference time of 29.32 ms on a CPU platform, it meets real-time seam tracking requirements.

In this paper, a real-time assistive torque controller based on sliding-mode control is proposed for a Series Elastic Actuator (SEA)-driven lower limb assistive exoskeleton. To address the problem of the lack of buffering properties and the uneven torque output in traditional exoskeletons, a novel SEA is designed for the hip joint lower-limb exoskeleton. This structure features excellent cushioning properties and smooth torque output. On this basis, to enhance the torque tracking performance of the hip joint exoskeleton, in this study, a robust composite control strategy is proposed, which can maintain accuracy in the presence of unknown external disturbances and model parameter inaccuracies. The strategy consists of an adaptive phase oscillator for outputting the phase of the gait, a single-peak curve to provide a reference assistive torque, and a low-level controller to track the torque. The low-level controller employs Continuous Sliding-Mode Control (CSMC) to obtain a continuous control law and utilizes an Extended State Observer (ESO) to estimate the lumped disturbance. It ensures that the tracking error is asymptotically convergent with minimized chatter. The closed-loop stability of the system is theoretically proven by the Lyapunov method. The validity of the proposed algorithm is validated on a designed exoskeleton.

Logistics activities in mixed-model assembly lines involve the material flow that depends heavily on the arrival sequence of products to be assembled in these lines. The main effort of this paper is to analyse how the arrival sequence of mixed models to a specific assembly line affects the logistics activities of the line, where a smoothly- and evenly- distributed material flow is desired. However, the uncertainty caused by the stochastic arrival sequence of customer orders in a make-to-order environment increases the complexity of logistics activities. Confronted with this uncertainty and complexity, we put the logistics activities into a more vivid and intuitive scenario by simulating the behaviour of materials. To begin with, the logistics activities occurring in a specific automobile assembly line is investigated. Secondly, a logistics simulation model of the assembly lines is built. Using the genetic algorithm toolbox embedded in Matlab7.0 package, we obtain subsequently an optimised model sequence that enables an evenly- distributed material flow. Finally, simulation results generated from Matlab7.0 Package are discussed.

Line-laser stripes in robot vision-guided adhesive dispensing exhibit narrow local structures, while keypoint localization is highly sensitive to spatial information and online processing imposes stringent requirements on computational efficiency. To address these challenges, this paper proposes a lightweight three-keypoint recognition method for line-laser-based adhesive dispensing. First, a two-stage top-down vision pipeline is adopted: an object detector locates the region of interest (ROI), after which the left, center, and right points are predicted within the local ROI. The keypoint network uses RepViT-M0.6 as a lightweight feature extraction backbone. To compensate for the low spatial resolution of its final-stage features, LiteFlattenSimCCHead is designed to transform deep features into position-sensitive representations better suited to keypoint localization through low-dimensional feature transformation, progressive spatial reconstruction, lightweight local feature fusion, and position-preserving feature flattening. SimCC coordinate classification and DARK distribution-aware decoding are then combined to obtain the coordinates of the three keypoints. Experimental results show that the proposed method achieves an AP of 0.9532, a mean localization error of 3.6385 px, an RMSE of 4.8272 px, and a PCK@5 of 0.7842. Compared with the Lite-HRNet-18 baseline, the mean error and RMSE are reduced by 16.5% and 20.1%, respectively, while AP increases by 3.80 percentage points. For single-ROI inference in the second stage, the final model reaches 116.34 FPS on the GPU and 49.32 FPS on the CPU. These results demonstrate that the proposed method improves localization accuracy while maintaining a lightweight model and high inference efficiency. Therefore, the reported results demonstrate the performance of the vision and keypoint-recognition method, rather than the positioning accuracy of the complete robotic adhesive-dispensing system.

Reliable extraction of local trajectory geometry from line-laser images is essential for robotic adhesive-dispensing perception, particularly in multi-target scenes affected by reflection, stripe discontinuity, and weak responses. This study presents a two-stage framework that separates full-image target localization from ROI-level keypoint estimation. YOLO26n first detects local trajectory regions, after which an orientation-aware lightweight laser keypoint network (OLKNet) predicts one center control point and two directional reference points within each ROI. OLKNet combines a FastViT-T8 backbone, a lightweight top-down feature pyramid, center-guided rotated anisotropic heatmaps, and channel-dependent DARK decoding. A dataset comprising 4460 images and 5656 target instances was constructed. On the test subset, YOLO26n achieved a precision of 0.9878, a recall of 1.0000, and an mAP_50:95 of 0.8737. On 567 test instances matched to predicted ROIs, OLKNet achieved an AP of 0.9120, a center-point mean Euclidean error of 3.406 pixels, a center-point root mean square error of 4.316 pixels, and an angular error of 2.819^∘. Under a five-frame scheduling strategy, the GPU vision-software pipeline achieved 73.93 FPS for single-target images and 47.17 FPS for multi-target images. Here, real-time performance refers to the vision-software pipeline rather than robotic closed-loop dispensing.