The advancement of Industry 5.0 has positioned human–robot collaboration (HRC) as a critical component of smart manufacturing, requiring robots to work seamlessly alongside humans in dynamic, unstructured environments. However, several challenges remain, including the heavy reliance on object-specific data for training dynamic perception models, the reliance on manually prepared CAD models for 6D pose estimation, and the lack of interaction semantics in grasp planning. To this end, a one-shot human-centric robotic grasping framework that does not require object-specific training or manually prepared CAD models, namely OH-Grasp, is proposed for HRC scenarios. This framework uses a single masked reference RGB-D observation to support instance segmentation, 6D pose estimation, and human-centric grasp planning for query RGB-D observations. First, the one-shot learning-based instance segmentation (OSeg) module is proposed without object-specific training, which utilizes the semantic consistency and geometric constraints of pre-trained visual encoders to guide a promptable segmentation foundation model, achieving robust one-shot instance segmentation. Second, the one-shot learning-based 6D pose estimation (OPose) module is proposed without manually prepared CAD models, combining single-view 3D reconstruction with differentiable rendering optimization to address the challenges of scale and 6D pose estimation for new objects. Finally, the human-centric robotic grasping (HGrasp) module is proposed without object-specific training, which leverages Vision-Language Models (VLMs) to parse object functional regions and integrates them with 3D geometric constraints and a geometry-center prior to generate grasping strategies that align with human interaction habits. Real-robot experiments on 15 tools, 5 parts, and three multi-object scenes show that OH-Grasp achieves grasping quality values of 95.89%, 89.92%, and 95.16%, with an overall end-to-end system success rate of 91.11% over 180 trials. Under conservative safety settings, the scene-level execution times are 25.91 s, 29.15 s, and 32.24 s. A human-centric handover evaluation with 12 participants and 28,800 pairwise Likert ratings further shows that OH-Grasp is preferred in 20 of 24 Directness and Security comparisons. These results demonstrate the effectiveness of OH-Grasp for task-level one-shot human-centric grasping in dynamic and partially unstructured HRC scenarios.

Resilient manufacturing plays a key role in Industry 5.0, focusing on personalised demands and emphasising the combination of production flexibility and smart technologies. It enables workshops to adjust production more flexible to meet dynamic production needs, while optimising manufacturing processes. Autonomous Mobile Robots (AMRs) have become crucial in enhancing resilient manufacturing by dynamically adapting to real-time material handling tasks. Based on multi-agent communication, this paper proposes a distributed collaborative scheduling scheme comprising the autonomous perception, collaborative interaction, and scheduling decision aimed at enhancing workshop material handling resilience. First, based on the 'perceive-reason-decide' process, an AI-Agent model for AMR is constructed, providing them with sensing, analyzing, and processing capabilities. Second, using the principles of single and bidirectional feedback modes, a collaborative interaction model is established to achieve their communication capabilities. Third, an AMR scheduling model is built based on Multi-agent Deep Deterministic Policy Gradient algorithm (MADDPG), which evaluates the scheduling result of transportation tasks by incorporating attention mechanisms and reward values. Finally, the collaborative scheduling scheme is validated through case studies. It can quickly respond to changes in material handling and make optimal allocation decisions, thus significantly improving the resilience of the workshop.

Human-robot collaborative assembly (HRCA) for mass customization frequently exhibits multiple uncertainties in assembly sequences, operator behaviors, and robot states. Existing research often addresses these uncertainties in isolation and lacks a knowledge-aware framework, thereby limiting context-aware adaptation and causing a decoupling between planning and execution. This study proposes a knowledge-augmented probabilistic decision model to synergistically handle these multiple uncertainties in HRCA. Firstly, a unified knowledge representation method integrating human-robot-process-component semantics is proposed using a pre-trained language model. Secondly, a knowledge-augmented partially observable Markov decision process model is established that embeds the structured knowledge into a probabilistic decision framework. The model quantifies human-robot uncertainties through online action prediction probabilities and target part pose scores. Based on the quantification, it enables online policy generation. Thirdly, an execution-level grasp adjustment strategy combining 2D detection and pose scores is proposed to guide end-effector corrections. Experimental results on reducer assembly tasks show that the proposed model achieves an efficiency improvement of 14.40% compared to the manually constrained partially observable Markov decision process baseline. The proposed model offers a robust solution for collaborative decisions in HRCA under multiple uncertainties.

Corner smoothing for linear tool paths is critical to improving trajectory tracking accuracy in complex manufacturing tasks for industrial robots. However, existing C3 continuous smoothing methods for 6R industrial robots fail to preserve the path shape, and shape-preserving smoothing methods for five-axis machines cannot adapt to the coupled kinematics of 6R robots. To address this problem, this paper proposes an analytical C3 continuous local corner smoothing method with shape preservation for industrial robots. First, the tool tip position is smoothed directly in the workpiece coordinate system (WCS) by constructing micro B-spline curves and adjusting control points to ensure that the tool path interpolates the discrete tool path points precisely. The constructed transition curves are optimized to reduce curvature while satisfying the deviation error tolerance. Then, the tool orientation is smoothed after being converted into three rotary Euler angles. Finally, parameter synchronization of the tool orientation with respect to the tool tip position is achieved by adjusting transition lengths and using shared curve parameters. The proposed local corner smoothing method not only ensures C3 continuity and shape preservation but also significantly improves tool path precision and smoothness. The effectiveness of the proposed method is verified through experimental results on an industrial robot with 6 rotational (6R) joints.

Objective Accurate intention recognition is required for human-robot collaborative assembly when an operator requests parts through natural interaction. In noisy and dim industrial environments, gesture-based recognition may suffer from low-quality visual inputs, whereas speech-based recognition is sensitive to acoustic interference. To improve recognition reliability under such disturbances, a multimodal method integrating gesture and speech was developed for a part-request task. The task involved eight components of a worm-gear reducer: an input shaft, an output shaft, a worm gear, a self-aligning ball bearing, the inner and outer rings of a tapered roller bearing, a perforated end cover, and a non-perforated end cover. A unified label space from 0 to 7 was adopted to align gesture commands, speech commands, part categories, and robot action commands.Results and Discussions Gesture recognition performance decreased as illumination weakened, but the improved YOLOv5s model consistently outperformed the comparison models under all illumination levels. At 200, 400, and 600 lx, the improved model achieved Precision values of 84.2%, 87.9%, and 93.1%, Recall values of 76.9%, 85.5%, and 92.7%, F1 values of 80.4%, 86.7%, and 92.9%, mAP@0.5 values of 86.6%, 90.7%, and 93.9%, and mAP@0.5:0.95 values of 57.3%, 62.1%, and 66.7%, respectively. With 4.6 M parameters and 8.5 G computational cost, the model achieved a favorable balance between recognition accuracy and lightweight deployment. Speech recognition performance declined as noise increased. The improved Temporal Convolutional Network with Selective Kernel attention (TCN-SK) achieved better performance than Long Short-Term Memory, Recurrent Neural Network, and the original Temporal Convolutional Network under all tested signal-to-noise ratios. At 0, 5, 10, and 15 dB, its Accuracy values were 68.5%, 78.7%, 89.7%, and 96.7%, and its F1 values were 68.6%, 78.8%, 89.8%, and 96.8%, respectively. The model had 1.76 M parameters, 0.44 G computational cost, and an inference time of 5.3 ms. Ablation results showed that wavelet scattering improved feature stability, while Selective Kernel attention enhanced adaptive multi-scale feature selection. Multimodal fusion effectively compensated for single-modal degradation. Under 600 lx + 15 dB, the fusion method achieved 98.4% Accuracy and 98.3% F1. When the signal-to-noise ratio decreased to 5 dB at 600 lx, fusion Accuracy remained 95.6%. When illumination decreased to 200 lx at 15 dB, fusion Accuracy remained 96.8%. Under 200 lx + 5 dB, the fusion method still achieved 91.7% Accuracy and 91.5% F1, exceeding the best single modality by 11.3 percentage points.Conclusions A gesture-speech multimodal intention recognition method was established for human-robot collaborative assembly under low illumination and noise interference. The improved YOLOv5s model reduced model complexity while maintaining static gesture recognition performance, and the improved TCN-SK model enhanced short digital speech recognition through wavelet scattering and adaptive temporal feature selection. The Dempster-Shafer decision-level fusion strategy used complementary modal information to improve recognition stability when one modality was disturbed. The method can support online part-delivery tasks in collaborative assembly. Future work should address synchronized video-speech stream recognition, dynamic gestures, long-sentence commands, few-shot generalization, and open-set rejection.

Convolutional neural networks (CNNs) have been widely employed for photovoltaic (PV) cell defect detection in electroluminescence (EL) images, yet they still suffer from high hardware dependency and the persistent challenge of balancing detection accuracy with inference speed. To overcome these challenges, this paper introduces FRNet, a lightweight network specifically designed for detecting PV cell defects in EL images. The core contribution is the Feature Reconstruction Convolution (FRConv) module, which effectively addresses two inherent drawbacks of conventional convolutions: rigid geometric structures that limit feature adaptability and inefficient utilization of channel information. The proposed FRConv module enhances feature extraction using a dual mechanism. It combines dynamically deformable convolution with channel importance weighting to adaptively adjust sampling positions. Additionally, gated cross-reconstruction operations are employed to fuse odd- and even-channel features, thereby improving the information flow. The network architecture uses depthwise separable convolutions with feature-bridged connections. This forms an efficient hierarchical structure capable of extracting defect features at multiple scales. To mitigate class imbalance in PV cell defect datasets, we introduce an adaptive focal loss that integrates learnable class weights with dynamic scaling factors, promoting balanced optimization across classes. Extensive experiments demonstrate that FRNet achieves a detection accuracy of 98.58 % on public EL datasets, while maintaining a lightweight design with only 1.30 M parameters and 0.78 G FLOPs, thereby significantly outperforming existing state-of-the-art methods. On an industrial-scale dataset, the model achieves 93.84 % and 89.71 % accuracy for 2-class and 14-class classification tasks, respectively. Moreover, FRNet delivers real-time inference at 29 FPS on resource-constrained hardware, highlighting its strong generalization capability and practical applicability in real-world scenarios.

Industry 5.0 represents a paradigm shift from efficiency-oriented automation to human-centric, resilient, and sustainable manufacturing, where human–robot collaboration (HRC) plays a crucial role by combining human flexibility with robotic precision. However, current HRC systems remain reactive and fragmented, lacking the alignment across perception, cognition, and execution required for seamless collaboration and robust generalization. While generative large models (GLMs) are emerging as a promising solution to these challenges, their integration into HRC exhibits a notable temporal lag compared to robotic domains, necessitating a systematic cross-domain synergy. This paper presents a review of GLM-enhanced HRC and proposes a prospective blueprint of multimodal perception, mutual cognition, and embodied execution for HRC in Industry 5.0. This blueprint outlines potential pathways toward human-centric smart manufacturing by synergizing generative artificial intelligence and embodied intelligence.

In the context of escalating geopolitical tensions and increasing supply uncertainty, global supply chains are increasingly shifting from the efficiency-oriented just-in-time (JIT) paradigm to a more redundancy-based just-in-case (JIC) strategy. The primary objective of this study is to examine the topological and functional implications of this strategic shift. To achieve this, this study first develops a directed and weighted five-tier supply chain network evolution model that incorporates dynamic compensation mechanisms, including multi-source substitution and internal inventory reserves. The model shows that the JIC strategy alleviates the “rich get richer” effect, leading to more balanced resource allocation. Based on this model, we investigate underload cascading failures triggered by supply shortages and evaluate network robustness using the order fulfillment rate. The results show that the evolved network has obvious dual characteristics of robustness and vulnerability. It can relatively resist small random disturbances, but it is particularly vulnerable to targeted attacks on key enterprises, and the system will fail rapidly due to bidirectional load propagation. In addition, although the JIC strategy improves the resilience to severe disturbances, excessive hoarding costs increase the threshold for the network’s survival and operation, and may also lead to early underload cascading failures under slight shocks. Managerially, this implies that focal enterprises must dynamically optimize, rather than blindly maximize, their strategic reserves to balance resilience with capital lock-up costs. These findings offer a theoretical basis for understanding the trade-off between resilience and cost in strategic supply chain design and provide managerial insights for focal enterprises seeking to build more resilient supply networks under uncertainty. A limitation of this study is its reliance on numerical simulations and future research should empirically validate these mechanisms using real-world industry data.

Industry 5.0 emphasizes a human-centric concept, aiming to construct highly intelligent, sustainable, and resilient manufacturing systems. While a large body of literature has explored its concepts, architectures, enabling technologies, and practical applications, literature specifically focused on production planning and control solutions in industry 5.0 shops are scarce. Recent literature indicates that the well-being and skills of human workers significantly impact shop performance due to their highly variable activities and behaviors. Workload control has been recognized as a simple yet effective solution to mitigate the effects of high variability - both human and machine - through a three-layer filter for high-variety make-to-order shops, offering potential for Industry 5.0. However, the existing workload control concept has two significant limitations. First, it primarily focuses on the workload of machines while ignoring the potential impacts of humans, and; Second, this concept relied on the fixed processing times and lack flexibility to cope with changes in human subjective behaviors. In response, this study first presents a human-centric order release method based on workload control, enhancing its adaptability by considering uncertain human processing times. Furthermore, we introduce five shop floor priority dispatching rules to further investigate the potential impacts of additional factors on our proposed method. Simulation results show that the human-centric method outperforms the traditional machine-centric method, particularly in pure job shops. Meanwhile, when combining the human-centric order release method with the shop floor dispatching rules, the load-oriented dispatching rules significantly improve the shop's performance in terms of throughput time, while the time-oriented dispatching rules increase order delivery performance. Counterintuitively, integrating human-centric concept into the shop floor dispatching stage is noteworthy, i.e. human-centric shop floor dispatching rule. It does not enhance shop performance compared to the original dispatching rules, but rather deteriorates the performance of order release on most measures. The findings of this study have important implications for both research and practice in Industry 5.0.

Social manufacturing (SM) is a task-oriented production paradigm that enables coordinated scheduling across distributed manufacturing resources. However, its operational efficiency is often limited by reliance on centralized management systems. In this context, blockchain technology shows strong potential for decentralized management in SM. However; the corresponding models for early warnings of production progress, especially for task fulfillment and delay risks, remain largely unexplored. To address the limitations, this article first proposes a blockchain-based decentralized production progress early warning system in SM. Furthermore, a corresponding early warning workflow is developed to define the overall interaction logic among participants. Then, a hybrid state machine is implemented within the smart contract to construct a task fulfillment status early warning model. Afterward, a Bayesian network-based model is designed to assess the evolving delay risks of interdependent tasks. Finally, a case study on air-conditioner production is conducted to validate the proposed models. Results show that the prototype system maintains scalable performance, achieving up to 17.56 transactions/s under write-intensive workloads and 52.45 transactions/s under read-intensive workloads, demonstrating more stable scalability than traditional centralized and directed acyclic graph (DAG)-based systems.

In the industry 5.0 era, the rise of new information technologies has fostered collaborative manufacturing networks (CMN) to meet with large-scale individualized manufacturing modes. Currently, the complex coupling of business processes in manufacturing systems presents management challenges for CMN, which might lead to cascading network failures. However, designing high-resilience CMN can mitigate losses when cascading failures happen in manufacturing systems. Therefore, to explore the resilience change mechanism and improvement method when cascading failure happens in CMN, this paper first proposes a new method for evaluating node importance based on global and local characteristics of the network. Further, a normally distributed capacity model is proposed considering the underload and overload of nodes, which is more in line with the real CMN, and the corresponding network load redistribution strategy is designed to improve the network resilience. Finally, taking company G's CMN as an example, when the cascading failure happened in CMN, the network resilience influenced by different load distribution strategies is studied, which shows that the influence of affordable load fluctuation range parameter beta on network resilience is greater than capacity factor parameter sigma, and our external cooperation strategy has a better effect than the traditional strategy in network resilience optimization.

Milling deformation is a prevalent quality issue in thin-walled parts due to their low stiffness and high material removal rates. However, conventional milling uses fixed cutting parameters that cannot respond to dynamic disturbances, leading to low efficiency and uncontrolled deformation. To address the above issues, this paper proposes a novel multi-access edge computing (MEC) enabled edge-cloud collaborative framework for the online optimisation and control of milling deformation for thin-walled parts, which significantly improves efficiency and reduces deformation in thin-walled parts. The framework predefines a set of theoretical milling parameters with an improved NSGA-II algorithm at the cloud level. Then, during the milling process, the MEC level dynamically optimises the theoretical milling parameters according to the changes in milling deformation caused by milling disturbances. Finally, the feasibility and effectiveness of the proposed approach are demonstrated through the milling process of a typical thin-walled part.

Nowadays, Industry 5.0 starts to gain attention, which advocates that intelligent manufacturing should adequately consider the roles and needs of humans. In this context, how to enhance human capabilities or even liberate humans from the processes of perception, learning, decision-making, and execution has been one of the key issues to be addressed in intelligent manufacturing. Large language models (LLMs), as the breakthrough in new-generation artificial intelligence, could provide human-like interaction, reasoning, and replies suitable for various application scenarios, thus demonstrating significant potential to address the above issues by providing aid or becoming partners for humans in perception, learning, decision-making, and execution in intelligent manufacturing. The combination of LLMs and intelligent manufacturing has inherent advantages and is expected to become the next research hotspot. Hence, this paper primarily conducts a systematic literature review on the application of LLMs in intelligent manufacturing to identify the promising research topics with high potential for further investigations. Firstly, this paper reveals the concept, connotation, and foundational architecture of LLMs. Then, several typical and trending interdisciplinary LLM applications, such as healthcare, drug discovery, social & economic, education, and software development, are summarized, on which an LLM-enabled intelligent manufacturing architecture is designed to provide a reference for applying LLMs in intelligent manufacturing. Thirdly, the specific pathways for applying LLMs in intelligent manufacturing are explored from the perspectives of design, production, and service. Finally, this paper identifies the limitations, barriers, and challenges that will be encountered during the research and application of LLMs in intelligent manufacturing, while providing potential research directions to address these limitations, barriers, and challenges.

Based on Automation ML (AML), Intelligent Production Lines (IPLs) for Industry 4.0 can effectively organize multi-dimensional data and models. However, this process requires interdisciplinary and multi-team contributions, which often involve the dual pressures of private data encryption and public data sharing. As a transparent decentralized network, blockchain’s compatibility with the challenges of AML collaboration processes, data security, and privacy is not ideal. This paper proposes a new method to enhance the collaborative evolution of IPLs. Its innovations are, firstly, developing a comprehensive two-layer management model, combining blockchain with the Interplanetary File System (IPFS) to build an integrated solution for private and public hybrid containers based on a collaborative model; secondly, designing a version co-evolution management method by combining smart contract workflows and AML multi-dimensional modeling processes; meanwhile, introducing a specially designed conflict resolution mechanism based on the graph model to maintain consistency in version multi-batch management and; finally, using the test cases established in the lab’s I5Blocks for verification.

In the context of Industry 4.0, knowledge recommendation serves as the basis for intelligent process planning. However, the limited interpretability of knowledge recommendation systems make it challenging for users to understand and trust the recommendation process. Consequently, this paper defines an interpretable knowledge recommendation process (iKRP) task that transforms the knowledge recommendation process into a sequential decision-making task through deep reinforcement learning (DRL). It then generates relational paths to the answers based on the topic entities within the knowledge graph. To improve the interpretability of the recommended process knowledge, the following research approaches are proposed: (1) a framework for recommending sequences of process decision knowledge; (2) a TransEx knowledge graph embedding model that integrates attention mechanisms and complex-valued embeddings, with the accuracy improvements of 5.56 % over baseline method; (3) a process knowledge recommendation network based on DRL through the asynchronous superior actor-critic algorithm to achieve interpretability; (4) enhanced interpretability of the recommended process knowledge via the presentation of clear decision paths. Finally, the validity and reliability of the proposed method are demonstrated through application cases, which achieve a final accuracy rate of 0.8148.

Currently, the manufacturing production mode has gradually evolved from mass production to mass customisation production. This poses the new challenges for intelligent workshop to implement the rapid response, flexible adjustment and collaborative production, and thus to meet the diversified and changing customer demands. However, in current workshop, the collaborative interaction ability between machining and assembly processes is lacking. The machining parts are solely simply integrated with the assembled products, which ignores the dual-shop production information. It leads to the low production efficiency and high inventory cost. To bridge the gap, this paper proposes and defines a machining-assembly dual-shop production scheduling problem (DSPSP). On it, the novel collaboration scheduling model is built to optimise the overall production efficiency and ensure the high-efficiency delivery. Then, an improved graph neural network-based deep reinforcement learning (GNN-DRL) algorithm is designed. Finally, the case study is given. Through 4 different case tests and 1 practical case validation, the results show that the proposed method could achieve a better performance than the traditional methods in scheduling domain. In addition, our work could effectively reduce waiting time for assembly processes, which provides a meaning reference for mass customisation production.

As a pivotal step in translating design into production, process planning significantly influences product quality, cost, production efficiency, and market competitiveness. The process knowledge base, a fundamental element of process planning, determines the intelligence level of product manufacturing. Methods that construct process knowledge bases using Knowledge Graphs (KGs) have increasingly become critical technologies for supporting intelligent process planning. However, traditional deep learning-based named entity recognition methods for constructing KGs require extensive manual effort in domain-specific data annotation, resulting in inefficiencies, prolonged construction cycles, and high costs. To address these challenges, this paper introduces a Large Language Model-enabled method for constructing Machining Process KGs (LLM-MPKG). Initially, Large Language Models (LLMs) are employed to pre-annotate machining process text datasets. A verifier is then developed to assess and filter the pre-annotated datasets, with domain experts re-annotating deficient data to create a highquality annotated machining process dataset. Subsequently, using this dataset and a fine-tuned LLM, a machining process knowledge extraction model, MPKE-GPT, is constructed. MPKE-GPT is then applied to extract knowledge from process planning case data for 50 parts within an enterprise, leading to the creation of the MPKG. A prototype system was also developed to support intelligent process planning. Compared to traditional deep learning methods, the proposed method reduces construction time by 48.58%, lowers costs by 46.44%, and enhances performance by 1.96%.

The balancing of U-shaped human-robot collaborative assembly (HRCA) lines presents significant challenges due to tight task-precedence constraints and the complexity of multi-mode operations. Existing optimization methods often struggle with maintaining solution diversity and avoiding premature convergence. This paper introduces a novel LLM-guided improved non-dominated sorting whale optimization algorithm (LLM-INSWOA) to address these issues. The proposed method effectively integrates the semantic reasoning and task planning capabilities of large language models (LLMs) with the global search properties of swarm intelligence. A hybrid encoding-decoding mechanism ensures the satisfaction of task precedence constraints, while the incorporation of Lévy flight and cosine-adaptive strategies enhances population diversity and accelerates convergence. Experimental evaluation on an industrial electric bus battery assembly line demonstrates that the LLM-INSWOA successfully minimizes cycle time, total cost, and workload imbalance, outperforming established metaheuristic algorithms such as the standard whale optimization algorithm, particle swarm optimization, and differential evolution. The results confirm the potential of combining cognitive models with evolutionary algorithms for complex manufacturing system optimization.

Models that predict tool life based on wear mechanism knowledge are typically inaccurate, as the use of simplified model parameters can have a significant effect on this prediction. While a tool life prediction model based on sample cutting data is limited to specific working conditions, which makes tool life prediction difficult to generalize, and needs a large amount of historical data as support. In this paper, the empirical formula of tool life based on wear mechanism knowledge was combined with a neural network, which can significantly improve prediction accuracy. Firstly, a concept of tool life grade is proposed, and its classification standard is outlined. Secondly, a prediction model based on the empirical life formula and experimental data was established. Thirdly, a tool wear prediction model based on a convolutional neural network (CNN) was established through the real-time tool condition data, and the corresponding life compensation strategy can be determined by comparing this with the historical data. Finally, the empirical life grade was adjusted to obtain the real-time tool life grade. A case example shows that the data-driven knowledge-guided prediction model can significantly improve the recognition accuracy of tool life grade.