Robotics and Computer-integrated Manufacturing最新文献

{"title":"Industrial application of a human-robot collaborative parallel two-sided destructive disassembly line balancing problem in multi-product, multi-line layouts","authors":"Lei Guo ,&nbsp;Zeqiang Zhang ,&nbsp;Haolin Song ,&nbsp;Yan Li","doi":"10.1016/j.rcim.2025.103209","DOIUrl":"10.1016/j.rcim.2025.103209","url":null,"abstract":"<div><div>Human-robot collaborative maximizes the respective strengths of humans and robots, driving profound transformations in green intelligent manufacturing and supporting efficient completion of diverse disassembly tasks in remanufacturing. However, existing studies mainly focus on single End-of-Life (EOL) product scenarios. With the increasing variety and volume of EOL products, traditional single-line layouts and disassembly modes struggle to meet the demands of large-scale, multi-type product disassembly. To address this, this paper proposes a human-robot collaborative parallel two-sided destructive disassembly line balancing problem (HRC-PTDDLBP) for multi-product, multi-line scenarios. Firstly, a mixed-integer linear programming model is established for HRC-PTDDLBP to minimize weighted workstation count, smoothness index, and safety risk. To effectively derive the Pareto-optimal solutions, an improved Augmented ε-Constraint method (AUGMECON-2) is developed, which introduces slack variables and adaptive ε-step parameters to enhance convergence stability and solution diversity while avoiding weakly Pareto-optimal points. Secondly, an improved multi-objective discrete water wave optimization algorithm is developed for efficient model solving. The algorithm constructs the initial population based on task priorities and component non-disassemblability, incorporates a decoding strategy considering direction and task attribute conflicts, and enhances search performance through refined crossover, local search, and restart strategies. The model and algorithm correctness are validated within the GUROBI commercial solver’s scope. Benchmarking against seven state-of-the-art multi-objective algorithms under two-sided, human-robot non-destructive, and destructive disassembly modes, the proposed approach demonstrates superior performance. Finally, application to disassembly cases of discarded printers and televisions further validates the method. Compared with the second-best algorithm, the smoothness index is reduced by 87.0%, and safety risk is improved by 20.22%, alongside significant gains in line length reduction and idle time minimization. These results illustrate the comprehensive advantages of the proposed method in multi-product, multi-line human-robot collaborative disassembly line balancing, offering a practical and adaptable solution for real-world disassembly systems.</div></div>","PeriodicalId":21452,"journal":{"name":"Robotics and Computer-integrated Manufacturing","volume":"99 ","pages":"Article 103209"},"PeriodicalIF":11.4,"publicationDate":"2025-12-20","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145796201","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Graph-driven Single-Robot Multi-Cognitive Agent System architecture for human–robot collaborative disassembly","authors":"Jianhao Lv,&nbsp;Jiahui Si,&nbsp;Wenchao Li,&nbsp;Ding Gao,&nbsp;Jinsong Bao","doi":"10.1016/j.rcim.2025.103207","DOIUrl":"10.1016/j.rcim.2025.103207","url":null,"abstract":"<div><div>The inherent limitations of single-agent systems in tackling complex tasks, combined with the inefficiencies of traditional multi-agent paradigms—where task decomposition requires distribution among multiple robots, resulting in resource redundancy and escalated costs. To address this critical constraint, a graph-driven Single-Robot Multi-Cognitive Agent System architecture is proposed. Firstly, scene graphs are constructed to transform unstructured visual data from the environment into graph-based triplets. By aligning these triplets with pre-constructed knowledge graphs, historical memories are activated through graph matching to inform system decision-making with precedented insights. Then, an attention-driven collaboration mechanism dynamically designates leader and supporter roles among the different agents, ensuring adaptive role assignment based on contextual demands. Complementing this, a global optimization framework facilitates the collective evolution of the Single-Robot Multi-Cognitive Agent System, enhancing both individual agent performance and inter-agent collaboration. Finally, the Model Context Protocol orchestrates robotic execution by harmonizing external resource utilization with computational processes, ensuring seamless translation of decision outputs into physical actions. Experimental results demonstrate that the method exhibits strong robustness and generalizability in dynamic disassembly queries.</div></div>","PeriodicalId":21452,"journal":{"name":"Robotics and Computer-integrated Manufacturing","volume":"99 ","pages":"Article 103207"},"PeriodicalIF":11.4,"publicationDate":"2025-12-18","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145785002","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"iLSPR: A Learning-based Scene Point-cloud Registration method for robotic spatial awareness in intelligent manufacturing","authors":"Yusen Wan,&nbsp;Xu Chen","doi":"10.1016/j.rcim.2025.103204","DOIUrl":"10.1016/j.rcim.2025.103204","url":null,"abstract":"<div><div>A critical capability for intelligent manufacturing is the ability for robotic systems to understand the spatial operation environment – the ability for robots to precisely recognize and estimate the spatial positions and orientations of objects in industrial scenes. Existing scene reconstruction methods are designed for general settings with low precision needs of objects and abundant data. However, manufacturing hinges on high object precision and operates with limited data. Addressing such challenges and limitations, we propose a novel <strong>L</strong>earning-based <strong>S</strong>cene <strong>P</strong>oint-cloud <strong>R</strong>egistration framework for automatic <strong>i</strong>ndustrial scene reconstruction (iLSPR). The proposed iLSPR framework leverages point cloud representation and integrates three key innovations: (i) a <strong>M</strong>ulti-<strong>F</strong>eature <strong>R</strong>obust <strong>P</strong>oint <strong>M</strong>atching <strong>N</strong>etwork (MF-RPMN) that learns from both raw data and deep features of the objects to accurately align point clouds, (ii) a <strong>G</strong>eometric-<strong>P</strong>rimitive-based <strong>D</strong>ata <strong>G</strong>eneration (GPDG) method for efficient synthetic data generation, and (iii) a digital model library of industrial target objects. During operation, vision sensors capture point clouds in the scenes, and the iLSPR method registers high-fidelity object models in the scenes using MF-RPMN, pre-trained with GPDG-generated data. We introduce an <strong>I</strong>ndustrial <strong>S</strong>cene <strong>O</strong>bject <strong>P</strong>oint-cloud <strong>R</strong>egistration (ISOPR) dataset in IsaacSim to benchmark performance. Experimental results demonstrate that iLSPR significantly outperforms existing methods in accuracy and robustness. We further validate the approach on a real-world robotic manufacturing system, demonstrating reliable digital reconstruction of industrial scenes.</div></div>","PeriodicalId":21452,"journal":{"name":"Robotics and Computer-integrated Manufacturing","volume":"99 ","pages":"Article 103204"},"PeriodicalIF":11.4,"publicationDate":"2025-12-16","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145785000","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"A temporal spatial human digital twin approach for modeling human behavior with uncertainty","authors":"Hongquan Gui ,&nbsp;Ming Li","doi":"10.1016/j.rcim.2025.103203","DOIUrl":"10.1016/j.rcim.2025.103203","url":null,"abstract":"<div><div>Modeling human behavior is paramount in the process of human-robot interaction (HRI). Human motion during HRI is uncertain. Even when the same individual performs the same action, it is impossible to ensure that the motion trajectory will be identical every time. These factors make modeling human behavior extremely challenging. Beyond human uncertainty, there are dynamic temporal spatial dependencies in HRI. Effectively capturing uncertainty while fully integrating temporal spatial features presents a significant challenge. Moreover, representing human behavior solely through human skeleton is insufficient. Recently, human digital twins have been developed to represent human geometry. However, current human digital twins are not well-suited for dynamic HRI scenarios, as they struggle to accurately depict high-dimensional human parameters, leading to issues such as nonlinear mapping and joint drift. In summary, existing methods find it difficult to address the human uncertainty, the temporal spatial dependencies, and the high-dimensional human parameters. To address the above challenges, this study proposes a temporal spatial human digital twin (TSHDT) for modeling human behavior in HRI. The TSHDT is based on predicted human skeletons and integrates forward and inverse kinematics along with diffusion prior distribution to represent high-dimensional human parameters, thus preventing joint drift and nonlinear mapping between joints. In developing the TSHDT, we introduce the human robot temporal spatial (HRTS) diffusion model to mitigate the uncertainty in human motion. The unique diffusion and denoising processes of the HRTS diffusion model can effectively submerge uncertainty in noise and accurately predict human motion during subsequent denoising steps. To ensure that the denoising process favors accuracy over diversity, we propose the temporal spatial fusion graph convolutional network (TSFGCN) to capture temporal spatial features between humans and robots, embedding them into the HRTS diffusion model. Finally, the effectiveness of the TSHDT was validated via predictive collision detection in human-robot fabric cutting experiments. Results demonstrate that the proposed method accurately models human behavior in collision detection experiments, achieving outstanding F1 scores.</div></div>","PeriodicalId":21452,"journal":{"name":"Robotics and Computer-integrated Manufacturing","volume":"99 ","pages":"Article 103203"},"PeriodicalIF":11.4,"publicationDate":"2025-12-13","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145731184","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Auction-based privacy-preserving cloud-edge collaborative scheduling considering flexible service ability for multi-source manufacturing tasks","authors":"Weimin Jing ,&nbsp;Yong Yan ,&nbsp;Yonghui Zhang ,&nbsp;Xiang Ji ,&nbsp;Wen Huang ,&nbsp;Youling Chen ,&nbsp;Huan Zhang","doi":"10.1016/j.rcim.2025.103206","DOIUrl":"10.1016/j.rcim.2025.103206","url":null,"abstract":"<div><div>In the context of cloud manufacturing service, scheduling manufacturing tasks is a crucial area of research because it directly influences service quality and efficiency. As intelligent manufacturing technologies advance, edge manufacturing service providers have gained increasingly flexible ability, enabling them to adjust local production schedules to adapt cloud manufacturing tasks. However, because local manufacturing information is private, traditional centralized cloud manufacturing scheduling methods cannot fully leverage edge flexibility to collaboratively schedule multi-source manufacturing tasks (including both cloud manufacturing tasks and local tasks of edge service providers) without risking the disclosure of sensitive information, thereby limiting improvements in both performance and efficiency of cloud manufacturing service. Therefore, we propose a privacy-preserving cloud-edge collaborative decision-making approach based on auction theory to schedule multi-source manufacturing tasks. First, a mathematical model that accounts for the objectives of both service providers and demanders is established to characterize the collaborative scheduling of multi-sourced tasks. Subsequently, a cloud-edge collaborative scheduling decision framework is introduced. Building upon this, a multi-stage scheduling method based on combinatorial iterative auctions is proposed, featuring novel bidding with a flexible execution timeline and distributed winner determination process incorporating bid consolidation mechanisms to enhance the efficiency of cloud-edge collaborative decision. Finally, to validate the superiority of the proposed method, computational experiments are conducted, comparing it with traditional centralized manufacturing task scheduling methods. The results present that the proposed method not only completes cloud manufacturing tasks within a relatively shorter makespan but also provides higher-value manufacturing services to demanders. Moreover, as the cloud manufacturing task load increases, this advantage becomes even more pronounced.</div></div>","PeriodicalId":21452,"journal":{"name":"Robotics and Computer-integrated Manufacturing","volume":"99 ","pages":"Article 103206"},"PeriodicalIF":11.4,"publicationDate":"2025-12-13","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145753457","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"A hierarchical human behavior modeling framework for safe and efficient human-robot collaborative assembly","authors":"Guoyi Xia ,&nbsp;Zied Ghrairi ,&nbsp;Aaron Heuermann ,&nbsp;Klaus-Dieter Thoben","doi":"10.1016/j.rcim.2025.103202","DOIUrl":"10.1016/j.rcim.2025.103202","url":null,"abstract":"<div><div>Human-robot collaboration (HRC) offers promising solutions in industrial assembly by combining human dexterity and robot efficiency. However, the proximity of the human and robot raises safety concerns that can limit efficiency. Previous studies have typically addressed safety or efficiency separately, relying on incomplete models of human behavior, which has led to robot control strategies with limited adaptability. To bridge this gap, this paper proposes a hierarchical human behavior modeling framework that integrates human motion prediction (HMP) and human action segmentation. The proposed model captures both the fine-grained motion dynamics and the higher-level task structure, enabling a more complete and context-aware understanding of human behavior. The model can enhance robot decision-making for both proactive safety mechanisms and dynamic task allocation. HMP compares three predictive models (Convolutional Neural Networks - Long Short-Term Memory (CNN-LSTM), Spatial-Temporal Graph Convolutional Networks (ST-GCN), Transformer). CNN-LSTM and ST-GCN outperformed Transformer, demonstrating better short-term predictive accuracy. Human action segmentation includes feature extraction, dimensionality reduction, clustering, and two-stage temporal segmentation. The HMP (CNN-LSTM) based features achieve the highest clustering performance. Two-stage segmentation demonstrates high accuracy, achieving normalized edit distances (NED) of 0.029 and 0.07 for task-level and sub-task-level segmentation, respectively. Evaluation results show that proactive collision avoidance using predicted motions increased safety distance (from 0.4633 m to 0.4717 m), while dynamic task allocation based on action segmentation improves robot efficiency (from 84.95 % to 98.56 %). These results validate the effectiveness of the proposed hierarchical human behavior modeling framework in simultaneously enhancing safety and efficiency in HRC assembly.</div></div>","PeriodicalId":21452,"journal":{"name":"Robotics and Computer-integrated Manufacturing","volume":"99 ","pages":"Article 103202"},"PeriodicalIF":11.4,"publicationDate":"2025-12-10","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145731795","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Target-oriented collision-free robot grasping using task-attendance teachers-student knowledge distillation for various dense-clutter scenarios","authors":"Shaodong Li ,&nbsp;Cheng Xiang ,&nbsp;Wei Du ,&nbsp;Xi Liu ,&nbsp;Huajian Song ,&nbsp;Feng Shuang","doi":"10.1016/j.rcim.2025.103201","DOIUrl":"10.1016/j.rcim.2025.103201","url":null,"abstract":"<div><div>The target-oriented collision-free robot grasping in dense-clutter scenarios has made significant progress. However, the prior studies primarily focus on the complex problem brought about by an increase in object quantity. Actually, the various scenarios will also result in the challenge. That means skill models will inevitably start to obtain bloat, as more subtasks appear. Our previous study presents a multi-teacher distillation strategy, thereby effectively avoiding bloat. Unfortunately, it still has trouble in execution capability when the scenarios have more subtasks and highly similar subtasks. The root cause is attributed to the inadequate capability of distilled student in subtask identification. It is naive to develop the subtask classifier based on the traditional data-driven way when human experience fails to work. Therefore, we propose a dataset-free classifier by migrating the classification capability of teachers to the student. Then, we further propose a task-attendance teachers-student knowledge distillation strategy to afford the various scenarios involving more and highly similar subtasks, thus significantly enhancing the performance of grasping. And we also leverage the language instruction to ensure the mask map of target object through LLM, improving the intuitiveness of human-robot cooperation. Extensive comparative experiment verifies the advantage of our framework. We measure the capability of teachers-student distillation, and value of dataset-free classifier in our framework. Importantly, the performance of segmentation strategy is tested under ambiguous instruction, visual occlusion, and color conflict. Furthermore, the excellent generalization and robustness are exhibited according to the real-world experiments involving unseen object, unseen task modality, and disturbance existing.</div></div>","PeriodicalId":21452,"journal":{"name":"Robotics and Computer-integrated Manufacturing","volume":"99 ","pages":"Article 103201"},"PeriodicalIF":11.4,"publicationDate":"2025-12-10","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145731770","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Design of support head with joint-synchronized structure and tuned mass damper for robotic mirror milling","authors":"Kun Chen ,&nbsp;Chenghao Huang ,&nbsp;Haonan Ma ,&nbsp;Peng Xu ,&nbsp;Bing Li","doi":"10.1016/j.rcim.2025.103200","DOIUrl":"10.1016/j.rcim.2025.103200","url":null,"abstract":"<div><div>Machined thin-walled parts are critical to lightweight aerospace structures; however, their low stiffness poses significant challenges in controlling deformation and vibrations during their machining. To counteract these force-induced errors, robotic mirror milling has emerged as a promising technique, because a support robot can actively neutralize the cutting forces. However, the support robot requires an appropriate end-effector to protect the workpiece surface, while providing stable support. In this study, a support head with plastic rollers and a multi-branched tuned mass damper (TMD) is developed. The support head, synchronized with the Joint 6 of the support robot, contacts the thin-walled part via plastic rollers. A follow-up motion plan algorithm is proposed for the support head to synchronize its motion with the milling tool. The multi-branched TMD is designed to suppress the vibrations of the thin-walled parts and support head in different directions, thus ensuring an effective contact between them. A frequency-response function model for the TMD is proposed for parameter design. Simulations and experiments confirm that the designed motion-planning algorithm reduces the velocity of Joint 6 by 78.6 %, while the TMD decreases the machined surface roughness by 32.8 % and 21.4 % at axial depths of cut of 1.0 and 1.5 mm, respectively. The results demonstrate that the joint-synchronized support head, combined with the multi-branched TMD, effectively ensures stable support to the workpiece, while preventing damage to its surface.</div></div>","PeriodicalId":21452,"journal":{"name":"Robotics and Computer-integrated Manufacturing","volume":"99 ","pages":"Article 103200"},"PeriodicalIF":11.4,"publicationDate":"2025-12-08","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145697314","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Time–torque coordinated optimization for trajectory planning of industrial robots","authors":"Zeyun Xiao,&nbsp;Danfeng Sun,&nbsp;Donglai Zhu,&nbsp;Yong Wang,&nbsp;Yi Yan,&nbsp;Huifeng Wu","doi":"10.1016/j.rcim.2025.103199","DOIUrl":"10.1016/j.rcim.2025.103199","url":null,"abstract":"<div><div>Trajectory optimization is vital for improving the operational efficiency and reliability of industrial robotic arms. However, increasing task execution speed frequently induces sharp fluctuations in joint torque, which elevates energy consumption, places excessive stress on actuators, and excites structural vibrations. To tackle these issues, this study introduces a unified time–torque optimization framework that combines predictive modeling with heuristic search. The framework adopts a neural network incorporating frequency-domain correlation and a delay-aware mechanism to model dynamic torque variations. Based on the predicted torque profiles, the robot’s motion trajectory is parameterized by quintic polynomials, and a multi-objective loss function is constructed to jointly minimize execution time and torque variation rate. Particle Swarm Optimization (PSO) is employed to perform a global search for intermediate joint velocities and accelerations, improving convergence toward near-optimal solutions. Experiments on a six-axis industrial robotic platform demonstrate that the proposed method effectively reduces execution time and smooths torque transitions, confirming its practicality for industrial applications.</div></div>","PeriodicalId":21452,"journal":{"name":"Robotics and Computer-integrated Manufacturing","volume":"99 ","pages":"Article 103199"},"PeriodicalIF":11.4,"publicationDate":"2025-12-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145651051","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"A lightweight detection network integrating multi-scale semantic refinement for steel strip defects","authors":"Baicheng Bian ,&nbsp;Long Chen ,&nbsp;Zongwang Han ,&nbsp;Shiqing Wu ,&nbsp;WeiDong Li ,&nbsp;Hongguang Chen","doi":"10.1016/j.rcim.2025.103188","DOIUrl":"10.1016/j.rcim.2025.103188","url":null,"abstract":"<div><div>Real-time surface defect detection in steel strips requires high accuracy under computational constraints. Existing methods often struggle to detect low-contrast defects in complex backgrounds while maintaining computational efficiency. This paper proposes MSR-Det, a lightweight network built on a multi-scale semantic refinement framework to address this challenge. The core innovations include: a Hierarchical Texture Enhancement Module (HTEM) that strengthens discriminative textural patterns, a Specific Cross-Level Interaction Module (SCLIM) that facilitates granular feature fusion across network depths, and a Feature Enhancement Module (FEM) that refines defect representations through deep-to-shallow semantic guidance. The architecture further incorporates two key modules: a Lightweight Large-Kernel Bottleneck Module (LLKBM) for efficient receptive field expansion, and an Adaptive Gating Spatial Pyramid Pooling (AGSPP) for dynamic multi-scale context integration. Extensive experiments on NEU-DET, GC10-DET, and our industrial SS-DET dataset demonstrate that MSR-Det achieves state-of-the-art performance of 86.6% Mean Average Precision at 50% Intersection over Union (<span><math><msub><mrow><mtext>mAP</mtext></mrow><mrow><mn>50</mn></mrow></msub></math></span>) on SS-DET with only 5.0M parameters, also attaining a real-time speed of 180 FPS. This work provides a robust and practical solution for automated visual inspection in industrial settings, effectively balancing high precision with operational efficiency.</div></div>","PeriodicalId":21452,"journal":{"name":"Robotics and Computer-integrated Manufacturing","volume":"99 ","pages":"Article 103188"},"PeriodicalIF":11.4,"publicationDate":"2025-11-29","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145614038","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}