Accurate electrode instance identification from X-ray images is essential for reliable overhang analysis in lithium-ion battery electrode inspection. However, industrial X-ray radiographs often exhibit blurred edges, contrast attenuation, and structural occlusions caused by layer overlap, leading to discontinuous or distorted segmentation masks requiring post refinement. Existing mask refinement methods struggle in such low-contrast and densely packed patterns because they mostly rely on data-driven training without explicit physical knowledge. This paper presents a Bayesian refinement method that corrects coarse electrode masks through probabilistic inference with context-aware geometric priors. The proposed model integrates curvature, lattice, and global assembly priors with the image-based likelihood to derive a maximum a posteriori formulation, while an additional language-conditioned hyperprior estimated by an LLM adaptively modulates the strength of each prior according to morphological context. The core contribution is the interpretable geometric-prior Bayesian framework, which remains the sole optimization engine and accounts for most of the accuracy gain; the LLM is an enhancement module that only supplies per-image hyperparameter guidance and neither drives the optimization nor generates any mask. This design enables training-free, interpretable, and physically consistent restoration of electrode boundaries across diverse imaging conditions. Experiments on real X-ray datasets of cylindrical cells demonstrate that the proposed approach significantly improves the mask quality especially boundary accuracy and continuity compared with existing refinement methods in practical battery inspection. Beyond segmentation fidelity, the refined masks improve measurement-grade overhang analysis, reducing the standard deviation of electrode height and spacing by approximately 18% and 22%, respectively.

In human-robot collaboration (HRC), explicit and implicit interactions typically follow distinct workflows: explicit interactions are actively initiated by the operator to convey task objectives, while implicit interactions are continuously perceived and responded to by the system to adapt robot behavior to the current collaboration context. Existing HRC systems often handle these interaction types separately, lacking a unified cognitive and decision-making mechanism, which can lead to interaction conflicts, execution interruptions, or delayed safety responses in complex scenarios. Although large language models (LLMs) demonstrate tremendous potential in reasoning, they have yet to effectively integrate explicit and implicit interactions into robotic decision-making. To address this, this paper proposes an LLM-Agent-based framework for explicit-implicit interaction fusion, which models both types of input within a unified semantic space and incorporates a two-stage fusion decision mechanism in LLM reasoning. This framework helps robots maintain consistent responses to task objectives and mitigate potential collaborative safety risks. Preliminary small-scale experimental results on representative HRC tasks reveal favorable trends in task completion ratio, collaborative safety performance and decision consistency when compared with the baseline scheme without explicit-implicit fusion. The proposed method provides an interpretable and scalable approach for deploying LLM-based HRC systems in real-world settings.

Fatigue prognostic studies in composites often model the delamination evolution over time in one or two dimensions, for example, by depicting one-dimensional (1D) length or two-dimensional (2D) shape contour. However, these representations generally provide limited information about the spatial variation in damage severity within the delaminated region. In this context, this work proposes a prognostic framework for severity-graded fatigue delamination maps in composite laminates. Firstly, to facilitate downstream prognostic modeling, ultrasonic C-scan images are processed through pixel-intensity-based discretization, where pixel intensity values are transformed into discrete damage severity levels and replaced with integer class labels, resulting in severity-graded feature maps that simultaneously characterize the spatial extent and severity distribution of delamination-related damage. Then, a U-Net–transformer model is trained with prior-information-aided loss functions and employed to perform long-horizon prognostics of future severity-graded delamination maps. The framework is demonstrated through fatigue experiments on composite laminates monitored using ultrasonic C-scan inspections. The codes and data are publicly available at https://github.com/JingYuan93/3D-delamination-prognostics .

Robotic learning from demonstration (LfD) can significantly minimize robotic programming difficulties due to its intuitiveness and efficiency. Gaussian mixture model (GMM)/Gaussian mixture regression (GMR) is an effective method to implement LfD. However, there are limitations in GMM/GMR, such as the requirement of creating multiple demonstrations, sensitivity to demonstration defects, and underfitting/overfitting generalization in industrial tasks. To tackle the issues, in this study, an improved GMM/GMR approach with three improvements is proposed. Firstly, a Gaussian noise scattering strategy is designed to minimize the required number of multiple demonstrations and eliminate unsmooth turning areas/jitters defects in the demonstration. Secondly, based on a new evaluation criterion of incorporating the Gaussian cluster number, Bayesian information criterion, and dynamic time warping, the particle swarm optimization algorithm is introduced to achieve an optimal balance between feature preservation and smoothness for a regression solution, thereby effectively mitigating the overfitting/underfitting issues. Thirdly, the B-spline curve-based reconstruction of the regression solution is devised to enhance the adaptability of GMM/GMR to the changes in target positions. Experiments on retired product recycling are conducted to validate the effectiveness of the approach presented in this study, providing a reliable solution to facilitate LfD in industrial environments.

Personalized production is widely adopted in discrete manufacturing due to its ability to meet diverse customer requirements through flexible combinations of components. However, it also poses significant challenges to production synchronization, which requires all necessary operators, tools, materials, and industrial robots to be available within the prescribed time and space. In complex multi-stage manufacturing processes, execution deviations can accumulate into operation-level uncertainty, namely the risk that an operation cannot start, proceed, or finish as scheduled because the required resources and execution conditions are no longer aligned. Such uncertainty undermines resource utilization and may trigger cascading delays across the entire process. To address this challenge, this paper proposes a spatial–temporal event compiler (STEC) framework comprising digital, knowledge, and reasoning engines for compiling and analyzing manufacturing events. Specifically, the state information of digital models is compiled into a spatial–temporal event graph, which is continuously updated through a multi-clock alignment scheme that aligns distinct planning, scheduling, and execution time scales. Within the reasoning engine, a look-around reasoning approach is developed to capture historical consistency, the current execution context, and near-future evolution trends by integrating look-backward, look-present, and look-forward perspectives, thereby enabling uncertainty assessment. Furthermore, an LLM-based synchronization mechanism is incorporated into the framework to provide status reports through multi-turn interactions, allowing managers to query the current execution status and assess potential downstream impacts for timely intervention. Finally, a case study demonstrates that STEC improves the accuracy and stability of uncertainty identification.

Highlights: What are the main findings? A dual-layer collaborative framework is developed to synergize task allocation and path planning, ensuring global feasibility under heterogeneous constraints. The multi-altitude layering strategy effectively balances vertical takeoff and landing costs against horizontal detour efficiency, significantly improving 3D airspace utilization. What are the implications of the main findings? The proposed closed-loop optimization mechanism resolves the fragmentation issue in traditional sequential planning, offering a robust theoretical model for urban logistics. By quantifying obstacle distributions and flight dynamics across altitude layers, this study provides actionable technical references for engineering large-scale multi-UAV delivery systems. Efficient UAV logistics in complex urban airspaces requires a synergistic approach to task allocation and path planning. However, traditional methods often decouple these two phases, leading to physically infeasible or sub-optimal delivery schedules. This paper proposes a Dual-Layer Optimization Framework (D-LOF) to address the Multi-UAV delivery problem in 3D urban environments. The upper layer utilizes an improved Genetic Algorithm (GA) with a specialized constraint repair operator to optimize task sequences for a heterogeneous UAV fleet. The lower layer employs an altitude-aware A* algorithm that dynamically balances vertical energy costs and horizontal cruise efficiency across multiple altitude layers. Unlike conventional models, our framework iteratively feeds precise 3D flight costs from the lower layer back to the upper layer to guide evolutionary search. Simulation results demonstrate that the D-LOF consistently achieves global convergence within 20 generations. Compared to single-altitude planning and rule-based strategies, the proposed method can reduce total operational costs and maintains zero time-window violations in high-density obstacle scenarios. This study provides a robust decision-making tool for “last-mile” urban logistics by navigating the trade-offs between 3D spatial constraints and delivery punctuality.

Path planning for cargo unmanned aerial vehicles (UAVs) in continuous indoor–outdoor heterogeneous environments poses a critical challenge: promoting payload stability under sharp turns and abrupt altitude variations while maintaining navigational efficiency. To address this issue, this paper proposes a hybrid A*–APF path planning framework that embeds trajectory smoothness optimization directly into the planning process rather than treating it as a post-processing step. An improved A* algorithm is developed by incorporating a trajectory smoothness term into its cost function to penalize sharp turns during global path generation. The resulting path is further refined using an enhanced artificial potential field (APF) method with virtual target points and multi-field force synthesis to mitigate local minima. In addition, the Ramer–Douglas–Peucker algorithm is employed to remove redundant waypoints, and a trajectory generation module based on B-spline interpolation and minimum snap optimization is introduced to produce smooth and dynamically feasible trajectories. Numerical simulation results demonstrate that, in indoor warehouse environments, the proposed method reduces the average turning angle by 88.4% (to 23.1°) compared with the standard A* algorithm while maintaining a comparable path length of 135.11 m. In large-scale outdoor urban scenarios, it achieves a path smoothness of 0.0124 with an average turning angle of 40.0°, substantially outperforming the Genetic Algorithm (104.6°) and Particle Swarm Optimization (83.5°) on turning angle while delivering competitive computation times of 0.52–1.51 s. An ablation study confirms that the improved A* and enhanced APF components each contribute independently to turning angle reduction and local minima avoidance, respectively, and that their integration yields the optimal balance across all metrics. These results indicate the proposed framework’s effectiveness for UAV-based last-mile delivery in scenarios requiring seamless indoor–outdoor transitions under payload stability constraints.

Modern railway maintenance increasingly requires diagnostic systems that not only detect anomalies but also align decisions with established engineering standards and remain consistent across inspection cycles; however, many existing solutions rely either on operational alarm thresholds or on purely data-driven detectors, with limited integration between numerical inference and domain semantics. This study proposes a Multi-Agent Cognitive Digital Twin (MA-CDT) architecture designed to distribute cognitive responsibilities across specialised agents linked through shared semantic representations to support traceable information flow between physical inspection processes and digital reasoning components. To provide an initial implementation of the proposed architecture, its diagnostic component was instantiated and empirically evaluated using real operational track-geometry inspection data. The implemented module integrates neural anomaly detection, engineering intervention thresholds and ontology-guided semantic validation within a single decision structure. Comparative experiments against an operational alarm baseline and a neural-only configuration show that the integrated design reduces unsupported alerts while preserving threshold-defined intervention detections under noisy and imbalanced sensing conditions. Ablation analysis isolates the contributions of neural inference, rule-based constraints and semantic validation, demonstrating how their structured interaction stabilises decision behaviour. The diagnostic case study illustrates how one cognitive agent can operate within the proposed semantic framework and exchange information through the architectural knowledge structures. The results provide an initial empirical validation of the diagnostic component as a proof of concept within the proposed MA-CDT architecture and suggest that embedding domain constraints into neurosymbolic diagnostic reasoning could enhance decision stability and traceability in threshold-driven maintenance systems.

Low-visibility environments induced by sea fog severely constrain the navigational efficiency and safety in narrow waterways, where traditional radar and Automatic Identification Systems (AIS) frequently encounter challenges such as perception blind spots and information lag. To address this critical issue, this study proposes a UAV-based perception and decision-making methodology for main navigational routes in fog, integrating physical priors with unmanned aerial vehicle (UAV) vision. Firstly, a joint physical dehazing and fog-domain adaptive detection network is constructed. This network addresses the overcomes the interference of non-uniform fog through feature-level enhancement, generating a spatio-temporally continuous visibility field and ship probability grids under a bird’s-eye view (BEV). Subsequently, a quantified “Sailability Score” model is established, providing a scientific basis for the dynamic diversion, speed limitation, and safe distance maintenance of main navigational routes. Simulation-based verifications using real-world fog navigation scenarios in the Qiongzhou Strait, coupled with a joint analysis of Vessel Traffic Service (VTS) and AIS data, suggest that at the critical visibility threshold (≤500 m), the proposed method improves the recall rate of long-distance small target detection by approximately 16.2% and reduces the visibility estimation error by 19.3%. Furthermore, the consistency between the proposed Sailability Score and the actual VTS navigation restriction windows reaches 82.1%, exhibiting a conservative preference for safety (i.e., risk preference ratio (Formula presented.)). Additionally, by introducing a temporal anti-jitter mechanism (parameterized by a smoothing window (Formula presented.)), the proposed method extends the navigable time window of the main routes by approximately 12.4% while ensuring navigational safety. The simulation results indicate the framework’s potential perception capabilities and engineering applicability, providing reliable technical support for smart shipping and intelligent VTS systems.

Embodied intelligence (EI) is the latest paradigm in the evolution of artificial intelligence (AI), which becomes a key enabler for advancing intelligent manufacturing in industrial scenarios. In recent years, the explosive growth of large language models (LLMs) has provided a critical pathway for realizing EI, rapidly spawning numerous groundbreaking applications. The widely application of industrial robots is the core characteristics of intelligent manufacturing, where the emergence of EI makes industrial robots becoming smarter and smarter with the capability of executing ever more complex tasks today. Consequently, research on LLMs for industrial EI has rapidly attracted attention from researchers worldwide in the past few years. To better understand the current situation in this field, this paper reviews the progress of LLMs towards industrial EI and discusses the corresponding research status and development trend from the perspectives of the capability breakthrough of single-modal LLMs, the fusion enhancement of multimodal large language models (MLLMs), and the embodiment realization of LLMs. This review aims to provide useful references for the further advancement of LLMs to support the development of industrial EI.