Keywords: Machine learning, Autonomous Agents, Planning, Scheduling and Coordination
Abstract: Training robust policies in Multi-Agent System (MAS) remains challenging due to high sample complexity, unstable policy updates, and insufficient coordination among agents. Multi-Agent Proximal Policy Optimization (MAPPO) has recently demonstrated strong performance in cooperative multi-agent tasks; however, it does not explicitly leverage curriculum-aware optimization and lacks mechanisms that encourage consistent team-level coordination. In this paper, we propose an enhanced MAPPO framework that improves learning efficiency and cooperative behavior in complex environments. Specifically, we introduce Curriculum-Aware Advantage Weighting, a policy optimization strategy that dynamically scales policy gradients based on opponent difficulty during training. This mechanism allows agents to progressively emphasize more informative experiences from challenging scenarios while maintaining stability during early learning stages. In addition, we propose a multi-agent coordination loss that regularizes individual agent advantages toward a team-level objective, encouraging coordinated behaviors and reducing conflicting strategies among agents. This paper considers a typical multi-agent system scenario using the Google Research Football (GRF) environment, where a full 11-vs-11 match is adopted as the testbed for evaluating the proposed approach. Simulation results demonstrate that the proposed approach improves convergence speed, training stability, and final episodic returns compared with baseline MAPPO variants. These results indicate that integrating curriculum-aware policy optimization with explicit coordination regularization provides an effective framework for improving cooperative behavior and performance in large-scale multi-agent reinforcement learning tasks.

Keywords: Machine learning, Big data Analytics for Large-scale Energy Systems
Abstract: Accurate prediction of nitrogen oxide (NOx) emissions in coal-fired boilers is challenged by multi-variable coupling, dynamic time delays, and noise. To address this, this paper proposes a novel framework integrating causal time-series alignment and non-linear dimensionality reduction. Following rigorous data preprocessing, the Peter-Clark Momentary Conditional Independence algorithm is employed to identify causal time-delay relationships for multivariate alignment. Subsequently, the Hilbert-Schmidt Independence Criterion Lasso method extracts intrinsic non-linear features, effectively mitigating redundant interference. ly, a Long Short-Term Memory neural network is constructed for precise forecasting. Experimental results demonstrate that this method reduces the overall prediction error by approximately 45%, exhibiting exceptional tracking capabilities for both concentration trends and local fluctuations.

Keywords: Machine learning, Big-Data and Data Mining, Industrial and Service Robotics
Abstract: Understanding the technological structure of robotics innovation is increasingly important as robotic systems expand across manufacturing, healthcare, logistics, and service applications. Patent documents provide a comprehensive source for analysing technological development, yet the rapidly growing volume of robotics patents makes large-scale technology mapping challenging using traditional approaches. This study presents a semantic analysis of 19,823 robotics patents granted by the United States Patent and Trademark Office between 2006 and 2025. Transformer-based document embeddings derived from SciBERT are used to represent patent abstracts in a high-dimensional semantic space, followed by unsupervised clustering to identify dominant technological themes. The analysis reveals five major domains of robotics innovation: mechanical actuation, articulated manipulation, sensing and autonomous control, navigation and environment interaction, and intelligent field robotics. These clusters collectively delineate a layered technological architecture spanning hardware design, perception, control, and system-level deployment. The results demonstrate the effectiveness of transformer-based semantic analysis for large-scale technology mapping and provide insights into the evolving structure of robotics innovation.

Keywords: Machine learning, Cloud Computing For Automation, AI and Machine Learning in Healthcare
Abstract: Federated Learning (FL) is an emerging computing paradigm to collaboratively train Machine Learning (ML) models across multi-source data while preserving privacy. The major challenge of “meaningful” implementation of FL for any ML model is how to guarantee that the federated ML model can achieve comparable performance compared to the global model trained using the combined data. Moreover, there are very limited studies on FL of functional regression models that analyze functional data, a commonly encountered type of data in many fields. This study develops the first-of-its-kind federated Gradient Boosting algorithm with the Least Squares Approximation (fed-GB-LSA) for efficient, privacy-preserving federated learning of the function-on-function regression with several distinct merits: (1) The GB-based algorithm allows the sparse selection of multivariate functional and non-functional features in the function-on-function regression prediction, which is not straightforward in the functional regression; (2) The parameter estimation by the GB algorithm results in separate sub-optimization problems with explicitly analytical solutions for each of the features, providing an efficient estimation algorithm for the function-on-function regression; (3) The LSA-enabled fed-GB provides a “one-shot” approach for FL that is communicationally and statistically efficient, providing theoretical guarantees to the federated model’s performance without data sharing across local servers. The proposed fed-GB-LSA is tested in extensive simulation studies by considering real-world challenges such as device heterogeneity and applied in a real-world dataset for privacy-preserving telemonitoring of Obstructive Sleep Apnea (OSA).

Keywords: Machine learning, Foundations of Automation, AI-Based Methods
Abstract: Deep drawing tool try-outs are highly iterative and still manual processes, driven by the heterogeneous knowledge of individual toolmakers. This knowledge is implicit and experience-based, making it difficult to standardize, and reproduce systematically. Data-driven methods offer the potential to capture patterns from past try-outs in a more reproducible way, thereby decreasing the dependence on individual expertise. To reduce this manual effort, there is a growing interest in applying data-driven methods, such as neural networks, to automate the try-out process. However, neural networks require formalized knowledge and large, well-structured datasets, which are difficult to obtain in tool try-outs. To bridge this gap, this paper proposes an ontology-based framework for knowledge-infused neural networks. The ontology serves as a single source of truth, incorporating scientific, world, and expert knowledge in a unified, machine-interpretable structure. The framework maps typical knowledge representations, such as simulation results, equations, and logic rules, to integration points in the machine learning pipeline, and provides guidance on suitable method choices. A representative use case formalizes the initial tool try-out step by learning a mapping from spotting patterns, process parameters, and blank parameters to material removal on the active tool surfaces. Limited real try-out data can be addressed by using finite element simulation of contact pressure distributions and a proxy model that aligns them with real spotting patterns.

Keywords: Machine learning, Learning and Adaptive Systems, Deep Learning in Robotics and Automation
Abstract: Learning-based models are increasingly used to represent nonlinear dynamical systems in prediction, estimation, and control pipelines. In these settings, assessing the sensitivity of the learned model to perturbations in inputs and measurements is important, since prediction errors may propagate through the control loop and degrade performance.

This paper studies the sensitivity of neural network models for nonlinear dynamical systems to optimized additive perturbations. Adversarial optimization is employed as a computational tool to generate perturbations that increase the model prediction error. As a case study, we consider a popular autoencoder-based architecture for nonlinear system identification.

Two perturbation settings induced by the architecture are analyzed: an online setting, in which perturbations act on the transition-dynamics input, and an offline setting, in which corruption affects the measurement history provided to the encoder.

Numerical simulations on benchmark systems show that even small perturbations can significantly degrade prediction accuracy, highlighting the need for robustness assessment and certification when learning-based models are deployed in control applications.

Keywords: Robot Networks, Deep Learning in Robotics and Automation, Task Planning
Abstract: Vision-Language-Action (VLA) models have found increasing applications across various domains. However, their application in underwater environments remains underexplored due to degraded visual perception, complex hydrodynamics, and limited onboard computational resources. To bridge these gaps, we propose RoboFish-VLA, an embedded vision-language-action framework designed for intelligent swimming control of biomimetic robotic fish. We construct a real-world multimodal dataset containing approximately 11K egocentric vision-language-action samples collected from robotic fish deployments. RoboFish-VLA predicts short-horizon action sequences conditioned on visual observations, language task prompts, and historical motion states, allowing the model to compensate for slow actuation dynamics and oscillatory motion typical in biomimetic swimming. The system is deployed on an embedded robotic fish platform and evaluated through real-world experiments including narrow-gate navigation and language-guided target pursuit. The results suggest that language-conditioned task specification and historical motion feedback can improve robust decision-making for controlled underwater robotic-fish navigation.

Keywords: Robotics and Automation in Life Sciences
Abstract: The deployment of autonomous robotic systems in chemistry laboratories is accelerating experimental workflows and providing the foundational data for AI-driven scientific discovery. However, despite the success of data-driven methods in acquiring dexterous skills, safety remains a primary barrier to their deployment in high-risk domains, such as early-stage materials chemistry experiments. Specifically, learning-based policies frequently struggle to distinguish between safe and unsafe actions, leading to overconfident extrapolation and potentially catastrophic failures. To mitigate these safety risks, we propose textbf{SAFE-CHEM}, an uncertainty-aware framework designed for robust, learning-based robotic chemists. Our approach leverages an ensemble of recurrent neural network-based imitation learning policies to quantify epistemic uncertainty online through the variance of action predictions. By characterising the success-conditioned density of this variance using kernel density estimation, we introduce a hybrid control architecture that autonomously switches from the learned policy to a deterministic, rule-based backup controller when uncertainty exceeds a calibrated safety threshold. We evaluate textbf{SAFE-CHEM} across three fundamental laboratory manipulation tasks, where our empirical results demonstrate that this hybrid strategy improves overall task success rates and reduces critical safety violations compared to traditional single-policy baselines. Finally, we demonstrate the practical viability of the framework through zero-shot sim-to-real transfer onto a physical Franka Production 3 robot manipulator.

Keywords: Deep Learning in Robotics and Automation, Model Learning for Control, Motion Control
Abstract: State-of-the-art quadrupedal locomotion approaches integrate Model Predictive Control (MPC) with Reinforcement Learning (RL), enabling complex motion capabilities and terrain-adaptive behaviors through planning. However, they often face compounding errors over long horizons and have limited interpretability due to the absence of physical inductive biases. We address these issues by integrating Lagrangian Neural Networks (LNNs) into an RL–MPC framework, enabling physically consistent dynamics learning. At deployment, our inverse-dynamics infinite-horizon MPC scheme avoids costly matrix inversions, improving computational efficiency by up to 4× with minimal loss of task performance. Ablations show improved sample efficiency and reduced long-horizon error vs. unstructured neural dynamics. Lastly, we test our framework in the real world to demonstrate viability.

Keywords: Sensor Fusion, Failure Detection and Recovery, AI-Based Methods
Abstract: Time-of-flight (ToF) depth cameras lose up to 78% of their pixels on reflective surfaces such as polished floors and glass doors, making per-pixel depth fusion unreliable precisely where indoor mobile robots need it most. We propose a self-calibrating fusion system that turns this failure into a calibration signal: the sensor's surviving valid pixels anchor learned monocular depth to metric scale at runtime, without offline data or external ground truth. The resulting policy organizes sensing into a deliberate hierarchy (LiDAR as geometric anchor, hardware depth where valid, learned depth where ToF fails), evaluated on a navigation costmap. Offline replay on real corridor and pedestrian data shows 55--110% more costmap obstacle coverage than LiDAR alone. A distilled student (EfficientViT-B1, 5.31,M parameters) runs at 218,FPS on a Jetson Orin Nano at 2.7,GB and reaches 9/10 navigation success with zero collisions in closed-loop Gazebo, matching the foundation baseline at a fraction of the cost. An on-robot trial in a glass-walled corridor shows that the V9-fusion safety gate stops at the glass surface that LiDAR and raw ToF fail to detect.

Keywords: Sensor Fusion, Industrial and Service Robotics, Calibration and Identification
Abstract: Ultra-Wideband (UWB) ranging has become a viable option for industrial Autonomous Mobile Robot (AMR) localization due to improved accuracy and low cost. However, real-world deployments remain limited by two recurring challenges: calibrating static anchors can be time-consuming and error-prone, and integrating UWB with existing onboard sensors requires careful design to ensure robust and consistent pose estimation. Addressing these challenges, this paper presents an end-to-end pipeline that combines automatic anchor calibration with a generic multi-sensor estimator tailored to surface-bound vehicle motion. It targets existing AMR stacks in scenarios where robot pose priors are available for initialization. The calibration stage estimates anchor positions and range biases, while the localization stage fuses UWB with proprioceptive sensing in a bias-aware Extended Kalman Filter to improve consistency without extensive parameter tuning. Experiments on a commercial logistics AMR in a warehouse setting demonstrate accurate positioning indoors and across outdoor transitions, with improved consistency compared to an earlier estimator formulation. Evaluation on an independent forklift dataset further indicates transferability to other platforms. The method remains effective in test cases with limited line-of-sight and sparse anchor coverage. These results show that UWB localization can be deployed with substantially reduced manual effort while preserving the accuracy required for industrial AMRs. The collected warehouse dataset is made publicly available.

Keywords: Sensor Networks, Probability and Statistical Methods, Cyber-physical Production Systems and Industry 4.0
Abstract: Accurate temperature-field information is essential for structural safety, condition monitoring, and predictive maintenance in mechanical systems. However, dense sensor deployment and extensive simulations are often costly, making full-field thermal characterization difficult in practical industrial scenarios. To address this challenge, this paper proposes a jointly optimized variational probabilistic framework for reconstructing full-field temperature distributions from sparse sensor observations under varying operating conditions (OCs) and randomly available sensor configurations. The proposed method combines latent Gaussian process modeling with structured OC integration to capture the intrinsic low-dimensional structure of temperature fields and explicitly model OC-induced variations. In addition, observation-pattern similarity is incorporated into latent regularization to improve robustness under random sensor availability. The framework enables accurate full-field reconstruction and uncertainty quantification from limited measurements within a unified probabilistic learning process. Experimental results demonstrate improved accuracy and robustness under varying OCs and sparse sensor configurations, showing its potential for cost-effective industrial monitoring and digital-twin applications.

Keywords: Grasping, RGB-D Perception, Perception for Grasping and Manipulation
Abstract: To address the challenge of robust 6-DoF grasp prediction in cluttered,

unstructured environments, this paper proposes a geometry-aware two-stage grasping framework. First, a plug-and-play Structural Edge-aware Keypoint Extractor is introduced to selectively retain salient edge geometry while mitigating point loss during downsampling. Second, a graspable feature extractor of GraspLAKAN-Net is developed to

extract graspable features from stacked point clouds by integrating local aggregation with nonlinear mapping via an improved KAN-based architecture. To further enhance cross-object generalization, a coarse-to-fine Geometry-Aware Grasp Pose Estimation strategy is designed, enabling accurate and explainable 6-DoF pose prediction across both surface-dominant and edge-dominant objects. The proposed method is evaluated on a newly constructed MSDRG dataset comprising both simulated and real-world multi-object scenes. Comprehensive experimental results indicate the effectiveness of the proposed framework, which significantly outperforms state-of-the-art baselines in both recognition precision and operational efficiency.

Keywords: Haptics and Haptic Interfaces, Telerobotics and Teleoperation, Multi-Modal Perception for HRI
Abstract: Recent advancements in virtual reality and robotic teleoperation have greatly increased the variety of haptic information that must be conveyed to users. While existing haptic devices typically provide unimodal feedback to enhance situational awareness, a gap remains in their ability to deliver rich, multimodal sensory feedback encompassing force, pressure, and thermal sensations. To address this limitation, we present the Multimodal Feedback Exoskeleton (MFE), a hand exoskeleton designed to deliver hybrid haptic feedback. The MFE features 20 degrees of freedom for capturing hand pose. For force feedback, it employs an active mechanism capable of generating 3.5-8.1 N of pushing and pulling forces at the fingers' resting pose, enabling realistic interaction with deformable objects. The fingertips are equipped with flat actuators based on the electro-osmotic principle, providing pressure and vibration stimuli and achieving up to 2.47 kPa of contact pressure to render tactile sensations. For thermal feedback, the MFE integrates thermoelectric heat pumps capable of rendering temperatures from 10 to 55 degrees Celsius. We validated the MFE by integrating it into a robotic teleoperation system using the X-Arm 6 and Inspire Hand manipulator. In user studies, participants successfully recognized and manipulated deformable objects and differentiated remote objects with varying temperatures. These results demonstrate that the MFE enhances situational awareness, as well as the usability and transparency of robotic teleoperation systems.

Keywords: Human-Centered Automation
Abstract: Object placement is a fundamental component of aerial manipulation tasks, yet existing systems typically require the desired placement position to be specified explicitly in metric coordinates. Such interfaces are not intuitive and require users to reason about coordinate frames and scene geometry, making them difficult to use in practical deployments. In contrast, humans often communicate spatial goals through a combination of language and pointing gestures. Inspired by this observation, we present AERMANI-PLACE, a framework for language-guided object placement with aerial manipulators. Given a scene image and a natural language instruction, an image editing model generates a modified version of the scene containing a visual marker that indicates where the object should be placed. This marker is then grounded into the physical environment using depth observations to recover a metric place point, after which a placement trajectory is generated and executed by the aerial manipulator. We evaluate the proposed approach on a test set of 100 language-guided placement tasks and demonstrate successful execution on a real aerial manipulation platform. Experimental results show that the proposed method reliably infers placement locations from language instructions with an average success rate of 87% on the test-set and transfers effectively to real-world aerial manipulation with an average success rate of 72%.

Keywords: Industrial and Service Robotics, Automation in Construction, Sensor-based Control
Abstract: Bridges play a critical role in transportation infrastructure, and their structural safety is therefore essential. However, the stay cables of cable-stayed bridges are susceptible to environmental degradation. Conventional inspection methods require inspection equipment to be manually transported to specific locations along the cable, which is time-consuming, labor-intensive, and often conducted at night, thereby increasing operational risks. This paper presents the development of a cable-climbing robot designed for automated inspection of stay cables. The system adopts a modular architecture consisting of a control and sensing module, a drive wheel module, an idler wheel module, and a frame. A single-board computer (Raspberry Pi) serves as the primary controller and is remotely operated from a ground laptop via Wi-Fi. The robot employs an accelerometer to measure cable resonance frequencies and onboard cameras to visually inspect cable surfaces. The developed robot can inspect cables with diameters of 200-216 mm, traverse obstacles up to 5 mm in height, and incorporates a slow-descent safety mechanism to improve operational safety. Field experiments demonstrate the feasibility and effectiveness of the proposed system for bridge cable inspection.

Keywords: Industrial and Service Robotics, Product Design, Development and Prototyping, Motion Control
Abstract: Complex natural disasters and severe industrial accidents create challenging environments for search and rescue (SAR) operations. Navigating these environments often exceed the capabilities of single-modal robots, resulting in a critical ``capability-environment" mismatch. To bridge this gap, this paper proposes a compact triphibious platform designed for multi-domain inspection and search missions. Unlike complex hybrids, our design integrates three propulsion modes to enhance adaptability with minimal structural overhead. We establish a unified modeling framework utilizing a 12-dimensional state representation to derive domain-specific kinematics, complemented by stability analysis to ensure longitudinal and lateral terrestrial equilibrium. To maximize traversability, we define a global operational space and implement a Finite State Machine to orchestrate robust cross-domain transitions through complex barriers. Extensive experiments, from fundamental verification to full-mission outdoor trials, demonstrate the system's efficacy. Results confirm that the robot successfully negotiates environments where traditional bi-modal systems fail, offering a holistic solution for multi-domain inspection and search.

Keywords: Intelligent and Flexible Manufacturing, Product Design, Development and Prototyping, Motion and Path Planning
Abstract: There is a growing interest in utilizing computational design techniques to optimize the structural and geometric parameters of robots along with their joint values. Physics simulators play a crucial role in this process by enabling the training and refinement of computational models. In this paper, we investigate the use of GPU-accelerated differentiable simulation to find good robot designs for reaching a sequence of target positions. This allows for parallel optimization of independent problem instances which we utilize to determine a set of non-dominated solutions for the underlying multi-objective optimization problem. We evaluate the method of identifying an optimal robot manipulator design for specific paths, focusing on both performance and design evaluation. We compare this gradient-based approach against a random baseline and 3 multi-objective evolutionary algorithms.

Keywords: AI-Based Methods, Machine learning
Abstract: Abstract— Thermo-mechanical topology optimization is im- portant for the lightweight design of aerospace structures operating under coupled mechanical and thermal loads. However, conventional density-based optimization methods usually require repeated finite element analysis and iterative design updates, which can be computationally expensive when multiple loading scenarios are considered. This work-in-progress paper explores a data-driven generative framework for accelerating thermo-mechanical topology optimization using a conditional latent diffusion model. The proposed framework first constructs optimized topology samples using a conventional SIMP-based thermo-mechanical topology optimization solver. A variational autoencoder is then used to encode optimized structures into a compact latent space, where a conditional diffusion model learns the distribution of optimized layouts. To represent the relative influence of thermal and mechanical effects, a Thermal Compliance Ratio is introduced as part of the conditional information, together with boundary conditions and volume constraints. Preliminary results indicate that the proposed framework can generate topology layouts with similar structural patterns to reference optimization results. This study provides an initial investigation into diffusion-based generative design for thermo-mechanical topology optimization, and future work will focus on quantitative validation, physical consistency evaluation, and generalization under unseen loading conditions.

Keywords: Robust/Adaptive Control, Inventory Management, Manufacturing, Maintenance and Supply Chains
Abstract: Blockchain transaction packaging in account-based blockchains requires selecting and ordering transactions under block-capacity, conflict, and account-state constraints. This paper studies a multiblock transaction packaging problem in which feasibility requires nonnegative account balances at every execution prefix. We formulate a compact integer programming model and propose a Dantzig–Wolfe column-generation framework, where each column represents a feasible block-level execution path. A Fix-and-Optimize heuristic is further developed to improve integer solutions. Computational results show that the proposed method obtains high-quality feasible packaging solutions while preserving dynamic balance feasibility.

Keywords: Semiconductor Manufacturing, Intelligent and Flexible Manufacturing, Machine learning
Abstract: In semiconductor wafer fabrication, scheduling decisions at the ion implantation stage often face uncertainty in wafer arrival times due to variations in upstream processing and equipment conditions. Existing two-stage methods (predict-then-optimize) treat parameter prediction and decision optimization as independent stages and train prediction models based only on prediction errors. However, for discrete scheduling problems, prediction error does not necessarily reflect decision error, which may lead to frequent schedule adjustments, increased decision costs, and reduced production stability. To address this issue, this paper proposes a joint prediction-optimization method. A hybrid loss function is designed to simultaneously account for prediction accuracy and scheduling performance. To handle the non-differentiability of the decision error, a surrogate gradient mechanism is introduced based on an Actor-Critic structure, enabling closed-loop feedback between prediction and optimization. Experimental results show that the proposed method effectively reduces decision errors and decreases the number of decision failures compared with the two-stage method. The research provides a new perspective for scheduling optimization in the ion implantation stage and offers a practical method for integrating machine learning with combinatorial optimization in manufacturing systems.

Keywords: Simulation and Animation, Software, Middleware and Programming Environments, Logistics
Abstract: Without having a robotic bin packing setup, the prediction whether a packing plan results in a stable pallet is based on an informed guess based on prior experience. Since such a setup is costly, for many algorithms it is unclear if they are suitable for production. To close this gap, we provide our simulation framework PET-Physics that enables researchers to assess a solver's quality. Based on the pose recording of all bodies, it is possible to check not only for static stability, but also dynamic stability of a packing plan. A novel feature of our tool is the provision of data that supports the detection of boxes that might be overloaded. We demonstrated the realism of our framework by comparing simulation runs with real-world palletization. Our packing evaluation tool considering physics is publicly available at https://github.com/released/after-acceptance.

Keywords: Smart Grids, Big data Analytics for Large-scale Energy Systems, Environment Monitoring and Management
Abstract: Real-time power flow analysis and thermal-limit violation detection under N − k contingencies pose significant computational challenges in high-renewable power systems. This paper proposes the Physics-Inspired Topology-Gated Network (PITG-Net), a graph-based deep learning framework for rapid power flow computation and operational risk assessment under N−k contingencies. The framework embeds Alternating Current power flow physics into structured message passing with hard topology gating to adapt to discrete network changes, and employs a boundary-adaptive correction strategy that selectively applies Newton–Raphson refinement near thermal limits to minimize false positives and false negatives at acceptable computational cost. Experiments on the IEEE 118-bus system demonstrate that PITG-Net significantly outperforms conventional approximation methods in both regression accuracy and thermal-limit violation detection reliability.

Keywords: Sustainability and Green Automation, Intelligent and Flexible Manufacturing
Abstract: Accurate long-term energy baseline forecasting is critical for smart microgrid optimization in green intelligent manufacturing; however, existing methods are limited by high load variability and insufficient real-time adaptability. To overcome these limitations, this work proposes Advanced-AVMIII (third-generation AVM system). The framework incorporates three key innovations. First, it applies nonlinear feature engineering to reduce autocorrelation-induced time-lag. Second, it adopts a Transformer + LSTM-AT dual-model architecture to capture both global and local temporal dependencies. Third, it implements a Sequential Global Similarity Index (SGSI)-triggered dual-phase mechanism that monitors anomalies in real-time (Phase I) and automatically fine-tunes the model based on prediction error (Phase II). Validated on semiconductor facility chiller and air compressor data, Advanced-AVMIII achieves a 58.4% reduction in MAE over the baseline LSTM in Original AVMIII [1], which reaches a Phase I MAE of 1.20 kWh, and has an overall 36.2–37.5% system-level improvement over Original AVMIII [1].

Keywords: Deep Learning in Robotics and Automation, Intelligent and Flexible Manufacturing, Semiconductor Manufacturing
Abstract: In advanced semiconductor packaging, products often exhibit high unit cost, complex structures, and stringent precision requirements. Inspection systems such as Scanning Acoustic Microscopy (SAM) are employed to detect internal defects, particularly within the Epoxy Molding Compound (EMC) layer. However, ultrasound images are susceptible to various noise sources that degrade image quality and hinder defect identification. This study proposes a self-supervised image denoising framework called Multi-Branch Task U-Net (MBTUNet), which is a U-Net backbone with a multi-branch decoder. The model enables multi-task learning without requiring paired clean-noisy data or prior knowledge of noise characteristics. An empirical study of semiconductor packaging is conducted to validate the proposed MBT-UNet by comparing it with several benchmark methods (e.g., paired supervised learning approaches such as SC-UNet and RIDNet). The results show that MBT-UNet achieves competitive performance, as evaluated by the feature similarity (FSIM) and learned perceptual image patch similarity (LPIPS) metrics, and is comparable to supervised models despite being trained without paired data.

Keywords: Deep Learning in Robotics and Automation, Manipulation Planning, AI-Based Methods
Abstract: Open-vocabulary, task-oriented grasping of specific functional parts, especially with dual arms, remains a significant challenge. This is because existing approaches, even when leveraging powerful VLMs for task understanding, still struggle to generate precise grasps under defined constraints or to effectively coordinate dual-arm actions. We propose UniDiffGrasp, a unified framework integrating VLM reasoning with guided part diffusion to address these limitations. UniDiffGrasp leverages a VLM to interpret user input and identify semantic targets, which are then grounded via open-vocabulary segmentation. Critically, the identified parts directly provide geometric constraints for a Constrained Grasp Diffusion Field using its Part-Guided Diffusion, enabling efficient, high-quality 6-DoF grasps without retraining. For dual-arm tasks, UniDiffGrasp defines distinct target regions, applies part-guided diffusion per arm, and selects stable cooperative grasps. In completely zero-shot settings, UniDiffGrasp achieves grasp success rates of 0.876 in single-arm and 0.767 in dual-arm scenarios, significantly surpassing existing state-of-the-art methods, demonstrating its capability to enable precise and coordinated open-vocabulary grasping in complex real-world scenarios.

Keywords: Deep Learning in Robotics and Automation, Reinforcement
Abstract: Recent advances in bipedal robot locomotion control have demonstrated remarkable progress, with reinforcement learning methods exhibiting excellent performance in complex environments. However, most current research focuses on the design of privileged information and reward functions, a process that is not only time-consuming and cumbersome but also relies on prior knowledge of the problem itself. Moreover, training solely based on reward functions bypasses critical physical constraints, potentially leading to dangerous actions by the robot. To address these issues, this study proposes a novel locomotion control method for bipedal robots that integrates non-privileged information with constrained reinforcement learning. Without relying on prior knowledge, the method utilizes only real sensor-acquired self-perception data and achieves locomotion control for bipedal robots through minimal rewards and constraints, thereby enabling traversal of complex terrains. Simulations and physical experiments were conducted to validate the effectiveness of the proposed method. The results demonstrate that the proposed control method outperforms existing approaches and can effectively advance the development and application of bipedal robots in complex environments.

Keywords: Deep Learning in Robotics and Automation, Reinforcement, Motion Control
Abstract: The manipulation of deformable materials such as cloth remains a challenging task in robotic automation due to high-dimensional, nonlinear dynamics and material-dependent variability. In conventional reinforcement learning (RL), rewards are difficult to define for such tasks and may exhibit limited transferability under changes in dynamics. We investigate Generative Adversarial Imitation Learning (GAIL) with domain augmentation as a reward-free imitation framework for cloth manipulation. Given expert demonstrations of a cloth-laying task for a single cloth configuration, we train a discriminator using GAIL. We show that this discriminator is agnostic to dynamic parameter variations and can thus be used as a reward signal in reinforcement learning to train policies under domain augmentations in cloth size and mass. The resulting policies are more robust to dynamics changes, achieving better performance and lower variability when generalizing to unseen cloth configurations.

Keywords: Deep Learning in Robotics and Automation, Robot Networks, AI-Based Methods
Abstract: 视觉-语言-行动（VLA）模型显示出强大 可能实现机器人操作，但政策预先训练 不要总是公开精确所需的信息 行动生成。特别是，有效的控制可以 受益于（i）反映近期时间背景 行动塑造了当前的场景，（ii）感知 根据所暗示的控制角色组织线索 任务。我们介绍RoVLA，一种基于GR00T-N1的增强技术 该模块通过互补模块解决这两个问题： （i） 一种动作条件编码器（ACE），其使用 作为时间先验的历史行动 状态转移上的动作条件接地， （ii） 一个用于三元的控制注入（CI）模块 将感知特征结构化到预训练策略中 通过零初始化路径以&#

Keywords: Deep Learning in Robotics and Automation, Robot Networks, AI-Based Methods
Abstract: Vision-Language-Action (VLA) models have shown strong promise for general-purpose robotic manipulation, but existing approaches often condition control on entangled multimodal features that do not explicitly preserve task-relevant structure during action generation. This limitation is especially pronounced in long-horizon or relationally ambiguous tasks, where reliable control requires persistent grounding of object roles, spatial targets, and interaction intent. We present Structured-VLA, a framework that introduces a fixed-format structural interface between multimodal understanding and continuous control. A structure-aware Vision-Language Model predicts predefined fields encoding semantic entities, geometric task anchors, agent-centric relations, and latent action codes. Their field-aligned hidden states are then reused as compact conditioning signals for a DiT-based Flow Matching policy, while dense visual features are preserved separately for local geometric refinement. Across LIBERO, Simpler-WidowX, and real-world SO101 evaluation, Structured-VLA consistently improves over a reproduced GR00T-N1.5 baseline, raising average success from 94.7% to 96.5% on LIBERO, from 39.8% to 45.1% on Simpler-WidowX, and from 73.1% to 77.3% on SO101. The gains are particularly notable on long-horizon, relation-aware, and multi-step manipulation tasks, while controlled ablations suggest that both the structural interface and the policy-side conditioning design contribute to the observed improvement. Overall, these results suggest that explicit structural conditioning is a promising way to improve downstream control performance and cross-domain robustness in pretrained VLA systems for robotic manipulation.

Keywords: Rehabilitation, Medical Robots and Systems, Motion Control
Abstract: Cervical traction is an important approach for the rehabilitation of neck dysfunctions. However, the traditional manual and mechanical traction methods are difficult to standardize, and lack the flexibility required for individualized treatment. Most existing robotic devices for cervical rehabilitation are limited by insufficient end-effector degrees of freedom and restricted workspace, which constrain their capability to perform multi-dimensional dynamic cervical traction training. This study presents a multi-DOF robot based on a serial-parallel mechanism for dynamic cervical traction. It provides four degrees of freedom for cervical motion, including vertical traction, flexion--extension, lateral bending, and axial rotation. A mechanical release mechanism is designed and installed between the moving platform and the head connecting ring, allowing automatic detachment when the interaction force exceeds a predefined threshold, thereby enhancing safety during cervical traction training. The kinematic model and workspace analysis of the system are presented in detail. Then, position control was implemented, and experiments were conducted to evaluate the position-tracking accuracy. The results show that the proposed robot provides a workspace sufficient to cover the range of motion of a healthy human neck. The normalized root mean square error of position tracking is below 10% in most movement directions. This study demonstrates the preliminary feasibility of the robot for cervical rehabilitation. Future work will investigate its potential application in therapeutic training for patients with degenerative cervical disc disease.

Keywords: Reinforcement, Deep Learning in Robotics and Automation
Abstract: This paper investigates whether sequential context improves goal-conditioned Reinforcement Learning in sparse-reward manipulation tasks. While Hindsight Experience Replay (HER) addresses reward sparsity through goal relabeling, its operation on isolated transitions limits its ability to capture temporal dependencies inherent in joint-space control. We hypothesize that incorporating motion history can enhance policy learning and introduce CONTHER, which integrates a Transformer-based architecture with a modified HER replay buffer. The Transformer encodes sequences of prior states and goals to provide temporal awareness, while the buffer populates experience with artificially successful trajectories. Two architectural variants are analyzed to examine how contextual information should be integrated. In simulated point-reaching tasks with a UR3 manipulator, CONTHER achieves a 38.46% higher average success rate compared to baselines and outperforms the strongest baseline by 28.21%, with faster convergence and more stable learning. The framework is further evaluated on three dynamic tasks requiring complex trajectory following and obstacle avoidance, where temporal context is critical. By operating directly on joint velocities, the approach provides a foundation for transfer to physical systems. The primary contribution is a systematic investigation into fusing sequential context with goal relabeling, offering insights into how temporal awareness benefits policy learning.

Keywords: Reinforcement, Learning and Adaptive Systems
Abstract: Existing quadrupedal locomotion learning paradigms usually rely on extensive domain randomization to alleviate the sim2real gap and enhance robustness. It trains policies by introducing various disturbances into simulated environments to improve their adaptability under uncertainty. However, since optimal performance without disturbances often conflicts with the need to handle worst-case scenarios, this paradigm has a trade-off between optimality and robustness. This trade-off forces the learned policy to prioritize stability in diverse and challenging environments over efficiency and accuracy in no-disturbed ones, leading to overly conservative behaviors that sacrifice peak performance. Inspired by disturbance rejection control, we propose a two-stage framework that integrates policy learning with motion reference to mitigate this phenomenon. This framework enhances the conventional reinforcement learning (RL) approach by incorporating imagined transitions as reference inputs. The imagined transitions are derived from an optimal policy and a dynamics model operating within an idealized setting without disturbances. By providing this reference, we found that the policy could generate actions that are better aligned with the desired motion, which led to accelerated training and reduced tracking errors under external disturbances.

Keywords: Remanufacturing, Assembly, Diagnosis and Prognostics
Abstract: Error detection is crucial for achieving smart and intelligent automation in assembly and disassembly processes. Various errors can arise during the disassembly of fasteners, potentially leading to operational uncertainties. In this study, we systematically collected torque-rotation angle data during the unfastening process across three distinct setups, aiming to develop a robust machine learning approach for error case detection. We initially benchmarked a Multilayer Perceptron (MLP), a one-dimensional Convolutional Neural Network (1DCNN), and a Fully Convolutional Network (FCN), as well as Long Short-Term Memory (LSTM) and extended LSTM (xLSTM) models, evaluating their application in identifying errors. The results demonstrate an impressive accuracy of up to 99%. To address explainability, explainable AI (XAI) is integrated via the LIMESegment method to identify the most relevant torque–rotation-angle segments contributing to model decisions, which supports domain experts (e.g., screw joint engineers) in interpreting the models. Additionally, we designed and developed a user-friendly graphical interface to facilitate practical applications, enabling monitoring of the unfastening process along with integrated machine learning error detection capabilities. This approach not only enhances the reliability and explainability of the robotic unfastening process but also paves the way for future advancements in smart disassembly.

Keywords: Remanufacturing, Intelligent and Flexible Manufacturing, Assembly
Abstract: Remanufacturing involves contact-rich disassembly and assembly operations that remain challenging for robotic automation. While disassembly trajectories can be obtained reliably, collecting assembly demonstrations is often costly and impractical. This study proposes a framework that enables assembly skills to be learned from disassembly experience. A task-oriented remanufacturing model is introduced to integrate perception, planning, and execution in an autonomous workflow. Disassembly trajectories are used to construct an axis-based progress representation combining pose and force–torque feedback, which provides guidance for reinforcement learning without manually designed reward functions. The method is validated in simulation and on a physical robotic platform using three representative skills: direct removal, lateral extraction, and twist–pull. Results show improved training efficiency and higher assembly success rates compared with conventional reward design and inverse reinforcement learning methods. Physical experiments further demonstrate robustness under typical remanufacturing disturbances such as grasping deviation and axial misalignment, indicating that disassembly knowledge can effectively guide assembly learning.

Keywords: Remanufacturing, Sustainability and Green Automation, Intelligent and Flexible Manufacturing
Abstract: End-of-life aerostructures are commonly processed by bulk scraping because removing millions of fasteners individually is prohibitively time-intensive, limiting circular value retention. This paper proposes a contact-centric destructive disassembly approach in which a straight-cutting tool (line saw/wire) performs continuous sweep cuts along part contacts to separate major aerostructure elements, similar to filleting a fish. We present a CAD-driven planning pipeline that estimates inter-part contact, constructs a contact-interface-following sweep by ordering locally separating cut lines, and generates coordinated dual-robot trajectories for executing the cut with a tensioned wire tool. In simulation, the method yields collision-free trajectories that separate the leading edge from the upper skin of an aircraft winglet sub-assembly. We further validate feasibility on a small-scale dual-robot hot-wire setup on styrofoam, where the planned cut sequence removes multiple attached subparts and achieves the target geometry.

Keywords: Logistics, Cyber-physical Production Systems and Industry 4.0, Intelligent Transportation Systems
Abstract: Cyber-Physical Internet (CPI) is a transformative global logistics paradigm that aims to replicate the efficiency and interconnectivity of the digital internet for the physical movement of goods. However, traditional logistics systems suffer from inefficient data handling, non-standardized addressing, and unreliable communication in multi-tier environments. This paper highlights an advanced IP-centric framework for CPI. The framework integrates a hierarchical IPv6-based addressing scheme, a memory-efficient Left-Child Right-Sibling (LCRS) algorithm, and a reliable Read-and-Reply communication mechanism. The proposed addressing scheme utilizes a Wide Area Network (WAN), Catchment Area Network (CAN), and Local Area Network (LAN) hierarchy to provide scalable, globally unique identification for logistics nodes and Protocol Shipment Units (PSUs). To support resource-constrained IoT devices, the LCRS algorithm is introduced to restructure the routing table, achieving a 53.85% reduction in memory footprint. This optimization enables the routing table to fit entirely within the on-chip SRAM of edge devices, effectively overcoming the memory wall challenge. Furthermore, the Read-and-Reply mechanism—inspired by the TCP three-way handshake—ensures reliable task coordination and goods transfer, maintaining a 97.25% communication reliability rate even under a 20% packet loss rate. Experimental results demonstrate that this framework significantly reduces routing table overhead and enhances communication reliability compared to traditional flat addressing and push-based protocols.

Keywords: Intelligent and Flexible Manufacturing, Additive Manufacturing, Planning, Scheduling and Coordination
Abstract: Additive Manufacturing (AM) is widely used for small-volume and highly customized production with multiple parts processed on the same machine. However, batch scheduling of heterogeneous parts in AM remains challenging due to two-dimensional packing constraints and batch-dependent processing characteristics. This paper studies a geometry-constrained batch scheduling problem on a single AM machine with the objective of minimizing makespan. We propose SeqPPO, a deep reinforcement learning method that combines a sequence-to-sequence architecture with Proximal Policy Optimization (PPO). It employs a state encoder and separate actor--critic decoders to directly generate part-selection and batch-termination actions for sequential batch formation. The state representation integrates static part attributes and dynamic batch features, while a hierarchical reward is designed to align local packing efficiency with global makespan minimization. Experimental results show that SeqPPO achieves near-optimal performance within 2.1% of mixed-integer linear programming (MILP) solutions while reducing solution time from up to 3600 s to less than one second. In addition, SeqPPO improves makespan by 25.3% on average over three heuristic methods. These results demonstrate the effectiveness and scalability of SeqPPO for real-time batch scheduling in AM.

Keywords: Assembly, Collaborative Robots in Manufacturing, Intelligent and Flexible Manufacturing
Abstract: Human-centric manufacturing has gained increasing attention in the context of Industry 5.0, and augmented reality (AR) has been widely adopted in human–robot collaboration (HRC) to enhance human operators' performance. However, despite the growing maturity of AR-assisted manufacturing technologies, excessive information presentation may impose additional burdens on operators. To address this issue, this paper designs a human-factor-driven AR-assisted human–robot collaborative assembly system and investigates how different types of AR information affect human state and system performance during collaborative assembly tasks. The effects of AR assistance are evaluated from both subjective and objective perspectives, including situational awareness, cognitive load, and task performance. The results provide empirical insights into AR information design in HRC and establish a basis for multi-dimensional optimization of collaborative assembly systems, including information presentation, interaction design, and task configuration, thereby supporting the development of more human-centric collaborative assembly systems in smart manufacturing.

Keywords: Logistics, Planning, Scheduling and Coordination, Optimization and Optimal Control
Abstract: A hybrid robotic order picking system (HROPS) featuring two distinct types of collaborative robots has been widely adopted in modern fulfillment centers. In such a system, picking robots work within storage aisles, retrieving bins between storage layers and buffer tiers (located at the bottom of racks), while moving robots transport bins between buffer areas and picking workstations to fulfill orders. By enabling two types of robots to operate in parallel, the system significantly enhances efficiency. To estimate performance of HROPS, we compare two scheduling policies (sequential vs. parallel) and three class-based storage policies (vertical, horizontal, and diagonal). Under the sequential policy, a picking robot first retrieves the required bin to the buffer tier before a moving robot is assigned, thereby avoiding idle waiting but resulting in longer cycle times; in contrast, the parallel policy assigns both robots simultaneously at the beginning, reducing cycle time at the risk of potential idle waiting. The vertical and horizontal storage policies store fast-moving bins near the buffer tier (to minimize vertical travel for picking robots) or near workstations (to reduce horizontal travel for moving robots) respectively, while the diagonal policy seeks a balanced trade-off. We model the system as semi-open queuing networks (SOQNs) and solve them using matrix-geometric method (MGM). Simulation experiments validate the accuracy of the analytical models. Numerical experiments assess policy performance in terms of throughput capacity, throughput time, and robot utilization. Finally, we formulate a cost-minimization problem subject to a service-level constraint to determine the optimal number of moving robots and rack configuration.

Keywords: Deep Learning in Robotics and Automation, Sensor Fusion, Force and Tactile Sensing
Abstract: We study vision-to-tactile prediction for robot grasping, where a short RGB sequence is used to infer bilateral tactile maps before a physical tactile sensor is available at deployment. The goal is to predict not only whether contact occurs, but also where it appears and how it evolves during grasp closure and hold. To support this problem, we build a high-fidelity paired visuo-tactile pipeline for Franka-TG2 and OmniViTac grasp scenes, and study a family of predictors that couple visual contact inference with temporal and structural tactile reasoning. Our aim is to establish vision-to-tactile prediction as an intermediate representation for grasp-state understanding, grasp stability estimation, and future sensor-light manipulation systems

Keywords: Rehabilitation, Health Care Management, Medical Robots and Systems
Abstract: In this study, a robotic system for image-based rehabilitation evaluation is designed by utilizing the robot’s mobility. Using the hand to spine task as an example, 16 volunteers are recruited for testing experiments. The results show that the accuracy of side view evaluation reached 95.8%, significantly higher than 58.3% achieved from front view only. Such a system compensates for the limitation of fixed-angle imaging, thereby improving the accuracy of evaluation and enabling more precise and personalized intervention plan for rehabilitation.

Keywords: Human-Centered Automation, Collaborative Robots in Manufacturing, Task Planning
Abstract: The ability to learn generalized skills from demonstrations and apply the acquired skills in various real-world scenarios is a key challenge for robot manipulation. Different from the typical robot learning tasks that learn the action from multiple demonstrated samples in a single task, zero-shot robot manipulation requires the robot to efficiently leverage multiple learned robot skills to accomplish a new task. In this paper, we propose an action decomposition/composition framework that efficiently transfers key manipulation skills to various new derivative tasks. Specifically, we first decompose one demonstration that encompasses several foundation skills that do not contain the derivative task. Then we adopt an action prediction approach to generate possible manipulation poses and the end pose for each subtask in the derivative task based on the robot's action and the video frames from the robot cameras. Since the generated poses may be impacted by previous misleading actions, we denoise the action by selecting the most possible manipulation poses based on the task to guide the robot manipulation. In the experiment, we show our framework can manipulate both robotic arm and automated Guided Vehicle (AGV) in the derivative task.

Keywords: Motion and Path Planning, Industrial and Service Robotics, Human-Centered Automation
Abstract: This paper addresses the integration of current and predicted human poses into the trajectory optimization problem of a robotic manipulator operating in a shared workspace, to achieve proactive collision avoidance. We propose a two-step approach. In the first step, a Feedforward Neural Network with Discrete Cosine Transform (DCT) preprocessing is trained to predict human poses over a finite prediction horizon, given a sequence of past observed poses. In the second step, we formulate a nonlinear trajectory optimization problem that minimizes control effort and torque variations, subject to dynamic, actuation, and joint-velocity constraints, and includes a collision-avoidance term in the objective function based on both current and predicted human poses. Numerical experiments on a scenario involving a 6-degree-of-freedom UR5 manipulator and a human operator demonstrate that the DCT-based approach limits the pose prediction error and that the proposed collision-avoidance term effectively maintains a safe human–manipulator distance throughout the planned trajectory.

Keywords: Intelligent and Flexible Manufacturing, Human-Centered Automation, Assembly
Abstract: The flexible assembly of high-end equipment is critical for improving production efficiency, reducing human manual workload, and enhancing adaptability to frequent design changes, which also aligns with the human-centric principles advocated in Industry 5.0. In this process, Mobile Parallel Robots (MPRs) play an essential role in executing transportation, alignment, and docking tasks for large-scale components, such as aircraft fuselages or cylindrical sections. However, current MPR systems generally lack sufficient cognitive autonomy to perceive and understand complex manufacturing environments, limiting their operational scalability in modern factories. To tackle this problem, this paper proposes an embodied intelligence framework for MPRs to enable flexible assembly of heavy-load cylinder sections in large, unstructured manufacturing environments. The framework utilizes dynamic 3D scene graphs with a rapid local update mechanism to achieve open-vocabulary vision-language semantic understanding and real-time spatial reasoning in factory environments. Simultaneously, Standard Operating Procedures (SOPs) are embedded into Large Language Models (LLMs) as physical constraints to ensure the generated task sequences comply with industrial safety requirements. During physical operation, the system integrates a hybrid visual relocalization method with A* chassis navigation to guide the MPR in completing collision-free, high-precision micro-insertion. This work is expected to establish a new paradigm of embodied, human-centric intelligent manufacturing in Industry 5.0, while driving forward both academic research and industrial innovation.

Keywords: Sustainability and Green Automation, Sustainable Production and Service Automation
Abstract: Laser welding technology has been broadly employed for manufacturing industry by virtue of high productivity, great processing quality, and excellent flexibility. Nevertheless, the laser welding process suffers from heavy energy consumption, poor energy efficiency, and severe carbon emission resulting in adverse impacts on the environment. The existing studies focused on energy consumption and carbon emission modeling, evaluation, and optimization, and there is a lack of comprehensive quantification on the relationship between carbon intensity and processing productivity in the laser welding process. To this end, this research performs the mathematical modeling between carbon intensity with processing productivity for laser welding considering multi-production stages. Firstly, the boundary of the typical single laser welding system is analyzed. Then, the carbon intensity and processing productivity indicators are proposed in terms of the ecological and productive properties, and the mathematical modeling between them is studied. Finally, the laser butt welding of aluminum alloy 6061 was carried out for verifying the reliability and practicality of the model, and the error analysis shows that the root mean squared error is 0.011 and the R-squared is 0.948. This research can lay the theoretical foundation for processing parameters optimization and carbon emission reduction in laser welding technology for sustainable manufacturing.

Keywords: Human-Centered Automation, Computer Vision for Manufacturing, Motion Control
Abstract: Accurate and efficient human motion prediction is essential for enabling safe and adaptive human–robot collaboration in future intelligent manufacturing systems. Existing methods either predict a single deterministic motion that ignores uncertainty or generate diverse outcomes that often lack realism and contextual consistency. To address these limitations, this paper introduces LoM-Diffusion, a novel diffusion-based model for human motion prediction that balances efficiency, accuracy, and diversity. The model employs a Transformer backbone enhanced with two key innovations: a Low-Rank Compressed Attention mechanism that significantly reduces computational complexity while preserving expressive power, and a Multi-Granularity Excitation module that adaptively recalibrates features across different temporal scales. Extensive experiments on the Human3.6M benchmark dataset demonstrate that the proposed model achieves competitive prediction accuracy with a reduced model size, providing a compact and effective motion prediction framework for human-centric collaborative manufacturing research.

Keywords: Planning, Scheduling and Coordination, Task Planning, Hybrid Logical/Dynamical Planning and Verification
Abstract: Outage scheduling of nuclear power plants (NPPs) is a critical component of energy-based Manufacturing Cycle Industrial Systems (MCIS), presenting a complex optimization problem to seek equilibrium among safety, availability, and economic efficiency. Aligning with the emerging paradigms of smart industries and data-driven intelligent optimization, this study's main contributions are reflected in three aspects. Firstly, it formulates a multi-objective optimization mathematical model that integrates safety, availability, and economic considerations, overcoming the limitations of traditional single-objective approaches in industrial and systems engineering. Secondly, an evolutionary learning algorithm, specifically an improved genetic algorithm closely coupled with the research domain, is designed to perform omni-dimensional organic scheduling. It employs task-priority-based encoding and a serial scheduling decoding mechanism to ensure all solutions satisfy the rule base and critical resource constraints while enabling efficient global search. Finally, the quantitative indicators of Probabilistic Safety Assessment (PSA) are innovatively incorporated, elevating safety objectives from qualitative restraint to quantitative decision-making. The study demonstrates that the proposed framework can effectively generate a set of Pareto-optimal solutions, intuitively revealing the nonlinear trade-off relationship between project duration and cost through industrial automation visual simulations. This provides a scientific, data-driven decision-support tool with significant benefits for the lean management and intelligent automation of large-scale energy MCIS.

Keywords: Collaborative Robots in Manufacturing, Human-Centered Automation, AI-Based Methods
Abstract: Human motion in human-robot collaboration (HRC) is intrinsically stochastic, with repeated task executions often resulting in varied motion trajectories. Predicting such motion thus requires modeling a probability distribution over future states, rather than estimating a single deterministic path. Although diffusion-based methods have recently advanced probabilistic motion generation, they remain constrained by two critical shortcomings in HRC settings: first, they frequently emphasize sample diversity at the expense of kinematic plausibility, and second, they lack explicit mechanisms to capture the structured, multi-scale interdependencies between human and robot movements. To address these limitations, this paper introduces GCFM—a graph-conditioned flow matching model for interactive motion forecasting. By adopting a flow matching objective, GCFM learns a time-continuous vector field that deterministically transports a simple noise distribution toward the data manifold of human-robot collaboration sequences. The model further incorporates a dual-graph prior, combining a fixed topological graph that enforces skeletal and mechanical constraints with a learnable collaboration graph that captures task-driven collaboration patterns. These relational structures are embedded into a flow prediction network, which performs multi-scale feature aggregation through graph convolutions and temporal filtering. To preserve coherence with observed motion history, a mask-based conditioning mechanism anchors known frames throughout the flow integration process. Experimental results demonstrate that GCFM achieves great prediction accuracy while maintaining competitive inference speed.

Keywords: Collaborative Robots in Manufacturing, Assembly, Human-Centered Automation
Abstract: Contact-rich assembly often exhibits clear stage structure, but under visual demonstration settings, stage transitions are not always reflected by obvious appearance changes. In this work, we formulate stage recognition in contact-rich assembly as a constraint-induced stage transition detection problem. We propose a structured framework that constructs Relational State Encoding, predicts frame-level states through Temporal State Modeling, and infers stable stage boundaries by Constraint-Aware Transition Inference under legal transition constraints. The method explicitly models relational states rather than raw appearance features, providing an interpretable approach to stage understanding in contact-rich assembly.

Keywords: Semiconductor Manufacturing, Planning, Scheduling and Coordination, AI-Based Methods
Abstract: Cluster tools are widely used for wafer fabrication processes such as etching and deposition processes. Since a cluster tool usually operates in a vacuum environment, the tool is called vacuum cluster tool (VCT) in this work. In practice, for many semiconductor tools, the VCT is connected with an equipment front-end module (EFEM) by two loadlocks to construct a front-end-module-equipped cluster tool (FEM^CT). In the FEM^CT, the two robots need to work cooperatively to transport wafers, which introduces additional challenges for the scheduling problem of cluster tools. Moreover, concurrently processing of multi-type wafers is also required in modern semiconductor fabs to improve the utilization efficiency of FEM^CTs. Thus, this work aims to schedule a single-armed FEM^CT with the concurrently processing of multi-type wafers. To do so, a novel deep reinforcement learning (DRL) method is developed to optimize the entry sequence of multiple wafer types into the tool for processing, so as to further improve the tool productivity. Finally, experiments are carried out to demonstrate the effectiveness and efficiency of the proposed method.

Keywords: Optimization and Optimal Control, Planning, Scheduling and Coordination, Task Planning
Abstract: The problem of integrated order acceptance and scheduling (OAS) plays a critical role in a make-to-order (MTO) production mode with production capacity constraints. This paper investigates the order acceptance and scheduling in non-permutation flow shops (OAS-NPFS) with the objective of maximizing total net profit, defined as the revenue of accepted orders minus the makespan-related cost. We formulate mixed-integer linear programming (MILP) and constraint programming (CP) models to solve small-scale instances. For large-scale instances, we design a logic-based Benders decomposition (LBBD) method that decomposes the problem into an MILP master problem (MP) for order acceptance and a CP subproblem (SP) for non-permutation scheduling. To enhance computational performance, the lower bound on the makespan is incorporated into the MP. Extensive computational experiments across small, medium, and large-scale instances validate the superiority of the proposed methods. The LBBD method successfully obtains optimal solutions for large-scale instances of up to 100 orders and 8 machines within minutes, demonstrating scalability to large-sized instances and applicability to real-world manufacturing systems.

Keywords: Collaborative Robots in Manufacturing, Planning, Scheduling and Coordination
Abstract: Timely robot assistance is critical in human–robot collaborative assembly under dynamic disturbances. Rework, sequence deviations, and human-state changes can invalidate part of the original schedule and require online updates. Excessive deviation between the updated and original schedules can disrupt human expectations, reduce collaboration fluency, and cause unnecessary rescheduling of tools, parts, and robot actions, leading to resource waste and increased waiting time. To address this issue, this paper presents an adaptive task planning method that improves schedule stability, preserves collaborative schedule continuity, and enables timely robot assistance under dynamic disturbances. The method is developed as a particle swarm optimization (PSO)-based adaptive task planning approach that combines initial collaborative scheduling with event-triggered partial replanning, where a schedule-stability term penalizes unnecessary start-time deviations from the original plan. After a dynamic event, completed and unaffected ongoing tasks are frozen, release conditions are updated, and only the unfrozen task set is rescheduled. Robot-supported atomic actions are modeled as pre-assistance actions associated with human-executed primitive tasks. A BV704 gas valve assembly case and larger scale simulated cases are used for evaluation.

Keywords: Planning, Scheduling and Coordination
Abstract: Scheduling cluster tools with an equipment front-end module (EFEM) requires close coordination among the vacuum module (VM), loadlock module (LLM), and EFEM. Since VM and EFEM operate under different pressure environments, loadlocks (LLs) must perform both wafer transfer and pressure switching, making robot–LL cooperation critical to system performance. This abstract revisits a cyclic scheduling framework for EFEM cluster tools with one-in-one-out LLs, focusing on VTR and ATR activity sequences and their cooperative relationship with LLs. Motivated by practical equipment configuration, an alternative LL configuration is further discussed, in which one LL is configured as a two-in LL for incoming raw wafers and another LL as a two-out LL for outgoing processed wafers. Representative robot activity sequences are outlined to show how this configuration may change robot cooperation and loadlock utilization, providing preliminary insights for future configuration-aware scheduling analysis.

Keywords: Task Planning, Machine learning, Model Learning for Control
Abstract: This paper proposes a modular grasping framework that combines Gaussian Processes (GPs), pose distributions, and Programming by Demonstration for generalizable and adaptive grasping. From user demonstrations, the system learns grasping splines that train GPs to predict grasp feasibility across object geometries, distinguishing robot-only, human–robot collaborative, and infeasible grasp regions. Furthermore, we integrate SpyroPose for SE(3) pose distributions, each bin is queried with the trained GPs to estimate grasp success probabilities. Unlike existing approaches limited to single-grasp evaluation, our method integrates pose uncertainty and reachability to identify optimal grasps across entire objects, enabling robust, geometry-aware grasp planning. The resulting method is integrated and tested on the real system consisting of a Universal Robot UR5e arm.

Keywords: Deep Learning in Robotics and Automation, Collaborative Robots in Manufacturing, Intelligent and Flexible Manufacturing
Abstract: This study presents an adaptive human-robot collaboration (HRC) system for electric vehicle battery disassembly workstations, integrating real-time fatigue detection through minimal physiological monitoring. The proposed approach relies on only two dry EEG electrodes embedded in a standard safety helmet, combined with eye-blink rate analysis and head movement detection, to classify operator vigilance into three states: alert, fatigued, and highly fatigued. Among seven machine learning algorithms evaluated via k-fold cross-validation (k=2 to k=10), a medium gaussian Support Vector Machine (SVM) achieved the highest accuracy of 95.2%. Detected fatigue levels drive dynamic adjustments of collaborative robot safety parameters, including protective zone distances and operational speed, according to human-robot proximity. Experimental validation using RoboDK on a battery disassembly station confirms effective real-time adaptation, reducing collision risks while preserving productivity. Compared to traditional 16-32 electrode configurations, the two-electrode setup ensures minimal intrusive integration into existing industrial personal protective equipment (PPE), aligning with Industry 5.0 principles for safer and more adaptive work environments.

Keywords: Task Planning, Motion and Path Planning, Manipulation Planning
Abstract: Robot manipulation of physical puzzles is an important requirement for automatic assembly and disassembly tasks. However, to enable robots to solve physical puzzles, manipulation skills need to be learned, which requires large training datasets, the generation of which is often time con- suming and tedious. To overcome this problem, we propose a framework that leverages procedural content generation (PCG) for automated generation of synthetic datasets of manipulation puzzles, which we call PhyRoGen. PhyRoGen is a general- purpose puzzle generator, which can generate physical puzzles with interlocking joint dependencies. Based upon PhyRoGen, we define six concrete generators which we use to generate 24 physical puzzles. By using a benchmarking framework, we are able to solve all puzzles in 1 to 300 seconds using sampling-based planning algorithms. Finally, we demonstrate that every generated puzzle is manipulatable by using a KUKA LBR iiwa robot in a physical simulation. This shows that our framework is able to procedurally generate unique, solvable robot manipulation puzzles, which is a crucial ingredient to benchmark manipulation algorithms and to develop robust foundation models.

Keywords: Model Learning for Control, Deep Learning in Robotics and Automation, Reinforcement
Abstract: Perceptive locomotion for legged robots requires adapting to dynamic environments. Model Predictive Control (MPC) provides interpretable planning with constraints but struggles with high-dimensional visual inputs and dynamic terrains. In contrast, model-free Reinforcement Learning (RL) adapts to visual scenarios but lacks planning. We propose VIP-Loco, which integrates vision-based scene understanding with RL and planning. During training, an internal model encodes proprioception and depth into kinodynamic features for the RL policy, which, at deployment, drives an infinite-horizon MPC. We validate VIP-Loco in simulation on diverse terrains across quadruped, biped, and wheeled-biped robots. Ablations show VIP-Loco unifies perception and planning for robust, interpretable locomotion.

Keywords: Robust/Adaptive Control, Industrial and Service Robotics
Abstract: We addresses the challenge of simultaneously compensating for state-dependent uncertainties and enforcing time-varying state constraints in Euler–Lagrange systems—common requirements in robotics yet underserved by existing control designs. A novel adaptive control framework is developed that combines an artificial time delay based uncertainty estimation (aka time delay estimation, TDE) with a barrier Lyapunov function to enforce constraint-aware control design. A state-dependent upper bound on the TDE approximation error is formulated, and an adaptive law is constructed to estimate its parameters online, enabling real-time state-dependent uncertainty compensation without relying on prior model knowledge. To ensure constraint compliance, the barrier Lyapunov function-based controller enforces time-varying bounds on both position and velocity. The resulting architecture is provably stable via Lyapunov analysis. Experimental results on a five degrees-of-freedom robotic manipulator validate the framework’s capability over the state-of-theart in adhering to safety-critical constraints under dynamic uncertainties.

Keywords: Collision Avoidance, Physically Assistive Devices, Sensor Fusion
Abstract: Autonomous wheelchairs using 2D LiDAR suffer collisions with elevated obstacles that sit above the scan plane and remain undetected. Post-SLAM methods add camera-based 3D awareness to navigation costmaps but never update the occupancy grid map used for localization, creating map-costmap inconsistency that degrades pose estimation. We present PRISMNav, a pre-SLAM fusion method that merges height-filtered depth projections from three cameras with 2D LiDAR via element-wise minimum, embedding elevated obstacles directly into the occupancy grid map before map construction. In threeway controlled trials (20 trials per configuration in each of two environments), PRISM-Nav achieved zero collisions across all 40 runs and reached 33 goals. LiDAR-only suffered 20 collisions across both sites, and the post-SLAM baseline (STVL) exhibited 3 kidnapping events. Ablation over the consensus threshold confirmed that this parameter controls the trade-off between spurious obstacles and missed detections. The system runs in real time (∼10 Hz) on CPU as a standalone ROS 2 node.

Keywords: Automation in Construction, Computer Vision for Manufacturing, Calibration and Identification
Abstract: Robotic practices on the construction site emerge as an attention-attracting manner owing to their capability of tackling complex challenges, especially in the rebar-involved scenarios. Most of existing products and research are mainly focused on the collection of large amounts of data with model training demands. To fulfill this gap, we propose OpenTie, a 3D training-free rebar tying framework utilizing a RGB-to-point-cloud generation and an open-vocabulary rebar detection on the real-world test. We implement the OpenTie via a robotic arm with a binocular camera and guarantee a high accuracy by applying the prompt-based object detection method on the image filtered by our proposed post-processing procedure for the image-to-point-cloud generation framework. Our pipeline requires no training efforts and outperforms the training-based object detection, i.e., YOLO-based method, with the verification on the real-world sequential rebar tying test. The system is flexible for horizontal and vertical rebar tying tasks and holds the potential application to the real construction site with possibility of commercialization.

Keywords: Smart Home and City, Sustainability and Green Automation, Environment Monitoring and Management
Abstract: With a rapidly urbanizing world, cities are increasingly evolving into complex socio-technological systems, illustrating the importance of integrated environment–technology approaches for sustainable smart-city development. Urban riverine zones are at the intersection of such human-nature interfaces where natural hydrological processes interact with the built environment. This study examines the spatial dynamics of two prominent networks, rivers and roads, in Ho Chi Minh (HCM) and Jakarta. First, a city-scale spatial scaling analysis was performed using Hack’s law for river networks and an analogous length–area scaling for road networks. Both cities exhibit strong power-law behavior in river networks (R ≃ 0.96–0.98) and moderate scaling in road networks (R ≃ 0.38–0.48), indicating that arterial roads retain a partially tree-like hierarchical structure shaped by underlying hydrological geometry. Further, a cross-modal Graph Neural Network (GNN) with dual encoders was developed to quantify river–road coupling through road-edge reconstruction. Same-city experiments show weak but non-random predictability (AUC = 0.527 for HCM; 0.577 for Jakarta), suggesting that river morphology provides limited yet detectable information about major road connectivity. Cross-city experiments reveal a strong asymmetry: models trained on Jakarta generalize moderately to HCM (AUC ≃ 0.596), whereas models trained on HCM fail to transfer to Jakarta (AUC ≃ 0.434). This study demonstrates that combining scaling law with cross-modal GNNs provides a powerful framework for understanding natural-engineered network intersections, offering actionable insights for resilient, informed smart-city planning.

Keywords: Telerobotics and Teleoperation, Virtual Reality and Interfaces, Surveillance Systems
Abstract: Underwater environments encompass sensitive ecosystems and critical man-made infrastructures that require careful monitoring to preserve both aquatic biodiversity and structural integrity. However, routine in-water inspection is challenging for both expert and non-expert users. This paper proposes WS Ray: a teleoperated underwater drone that minimizes human intervention in hazardous underwater environments through immersive inspection. WS Ray is a prototype engineered from first principles, designed to deliver an omnidirectional viewing experience through five degrees of freedom (DOF) maneuverability and a servo-actuated, pitch-controllable camera. A whale-shark-inspired hydrodynamic hull was chosen to reduce drag and drone motion is regulated by robust sliding-mode control (SMC). The system centered on a ground station connects a tethered drone and provides internet access. Live video is streamed via WebRTC to an immersive Virtual Reality (VR) app with head-tracked pitch/yaw control, supplemented by a novel auxiliary handheld controller for linear motion. The prototype was tested in a 2 m depth pool environment, remotely controlled over the internet for initial validation. The teleoperation delivered stable control, real-time visual feedback, and an immersive and natural experience, demonstrating WS Ray as a viable platform for remote underwater inspection.

Keywords: Task Planning, Behavior-Based Systems, AI-Based Methods
Abstract: Autonomous aerial object retrieval requires reliable high-level decision-making particularly when perception-driven conditions govern task progression. Conventional Behavior Trees (BTs) are widely used to structure such missions, but their deterministic condition evaluations limit the use of complex semantic perception. Vision–Language Models (VLMs) offer rich semantic reasoning through natural-language queries and therefore appear promising for this purpose; however, their outputs are uncalibrated and inherently uncertain, making direct integration into BTs brittle for safety-critical robotic systems. In this work, we bring these two perspectives together and propose a VLM-Augmented Behavior Tree framework for autonomous aerial manipulation that explicitly accounts for semantic uncertainty. VLM responses are treated as noisy observations of latent task predicates, including target visibility and safe camera transition. A Bayesian belief update mechanism is embedded within phase-specific BT subtrees to enable belief-space decision-making. To model VLM reliability, we introduce both constant and input-conditioned asymmetric confusion models, in which error rates adapt to the image blur measure and the UAV velocity magnitude. Experiments in a simulated aerial object retrieval mission demonstrate that the proposed belief-space reasoning framework improves robustness to semantic uncertainty, increasing success rate, while reducing recovery actions and mission time compared to deterministic and constant-error baselines.

Keywords: Intelligent Transportation Systems, Agent-Based Systems, Machine learning
Abstract: Agent-based evacuation simulations are powerful tools for precise policy evaluation, yet they suffer from high computational costs when exploring the vast design space of shelter locations and evacuation policies during the planning phase of Nuclear Power Plants (NPPs). This paper proposes a rapid AI-based surrogate framework for the automated screening of shelter candidate sites and allocation strategies by approximating high-fidelity simulation outputs at a significantly lower cost. A case study was conducted on the traffic environment around the Precautionary Action Zone (PAZ) of the Kori Nuclear Power Plant, demonstrating that the proposed model achieves up to a Spearman’s ρ of 0.70 and top 50% recall 80%, suggesting that it successfully captures the non-linear complexity of the road network. The surrogate generates predictions in under one second, achieving a 500-fold speedup over conventional simulation runs. These results demonstrate that the proposed framework establishes a critical technical foundation for the preliminary and automated screening of hundreds of thousands of diverse policy candidates, a scale of evaluation that is computationally prohibitive for traditional agent-based simulators.

Keywords: Additive Manufacturing
Abstract: Additive manufacturing (AM) technologies have often utilized for fast fabrication of structures with complex geometries. The first step for process planning and geometric evaluation is the effective segmentation of features of interest from geometric measurements (i.e., point clouds) of as-deposited parts. These features can range from local distortions introduced by process parameter fluctuations to decomposing the whole geometry into a several freeform shapes. However, manual segmentation of large point cloud data is time consuming and often inaccurate. This study proposes an unsupervised framework for automatic segmentation of geometric features caused by freeform designs or excessive material accumulation. The algorithm starts by transforming the point cloud coordinates to a space for which height corresponds to material accumulation. In this space, the quantile function of the transformed height encodes information of the amount of local regions with material accumulation. The quantile function is used to obtain thresholds for contour generation and iterative shape segmentation. A case study involving four Wire Arc Additive Manufacturing (WAAM) printed cylindrical components and one Stereolithography (SLA) printed dental part demonstrates the method’s robustness and interpretability. The results illustrate the effectiveness of the proposed framework in the presence of complex geometries.

Keywords: Machine learning, Optimization and Optimal Control, Semiconductor Manufacturing
Abstract: Achieving optimal parameter settings in epitaxial silicon carbide (Epi SiC) manufacturing is challenging due to the need to simultaneously meet conflicting multi-objective requirements, such as precise thickness control and uniform doping. This paper presents an extended Digital Twin (DT) framework that incorporates the Multi-Objective Optimization with Deep-Feature Gaussian Process (MOODFG) algorithm, explicitly extending our prior DT-in-the-loop framework (MRBORI) in [26] from single-objective tuning to spec-driven multi-objective recipe optimization. The framework enables the effective optimization of high-dimensional and interdependent process parameters, overcoming limitations of traditional methods. The key innovation of the MOODFG algorithm lies in its integration of deep feature extraction and Gaussian Process Regression, which allows dynamic refinement of feature representations and surrogate models through iterative learning. Unlike MRBORI [26] (one objective per run), MOODFG minimizes a weighted distance to a user-specified target vector (e.g., thickness and doping) and provides an uncertainty-aware target-attainment certificate with a practical stopping rule for deployment. This approach ensures robust convergence to optimal parameter values while efficiently balancing multiple objectives. Experimental validation using real-world Epi SiC manufacturing data demonstrates significant improvements in yield, parameter stability, and process adaptability, highlighting the framework’s transformative potential for semiconductor manufacturing.

Keywords: Medical Robots and Systems, Deep Learning in Robotics and Automation, Motion Control
Abstract: Abstract— Transformer-based imitation learning policies, such as Action Chunking with Transformers (ACT), have demonstrated significant potential in fine-grained bimanual surgical manipulation tasks. However, existing action chunking algorithms typically employ an open-loop execution strategy with a fixed step size, neglecting the kinematic consistency between consecutive action chunks. This often leads to severe velocity jumps at the chunk switching boundaries. To mitigate this jitter, traditional methods introduce Temporal Ensembling, which improves smoothness but significantly increases computational cost and introduces phase lag. Such drawbacks make it difficult to meet the high standards of low latency and immediate response required in surgical scenarios. In this paper, we propose DynaChunk, an adaptive dynamic chunking strategy tailored for surgical robots. We formulate the action switching problem as a discrete optimization task within a receding horizon. By constructing a composite objective function that incorporates velocity smoothness, position tracking error, and temporal stability, we dynamically determine the optimal action switching moment. This state-aware method avoids updating the policy during high-speed or highly dynamic periods. Experimental results in a simulated bimanual surgical task demonstrate that DynaChunk maintains a high task success rate (90.0%) while reducing the peak jerk of the manipulator by 57.8% and reducing the total inference time to approximately 1/13 of that required by the Temporal Ensembling method. Our approach not only drastically improves computational efficiency but also successfully eliminates the phase lag associated with traditional methods, achieving smooth and high-frequency real-time control for surgical manipulatio.

Keywords: Motion and Path Planning, AI-Based Methods, Environment Monitoring and Management
Abstract: Dynamic target search using multiple unmanned aerial vehicles (UAVs) is challenging due to stochastic target emergence, partial observability, and strict onboard energy limitations. This paper proposes an energy-aware decentralized search framework for multi-UAV dynamic target search in uncertain environments. A dual-information hexagonal map is introduced to simultaneously represent environmental uncertainty and target existence probability, enabling information-driven search decisions under partial observability. The problem is formulated as a constrained multi-agent partially observable Markov decision process (CMAPOMDP), and a safe multi-agent reinforcement learning (MARL) approach, namely multi-agent proximal policy optimization (MAPPO)-Lagrangian, is developed to learn decentralized policies while enforcing UAV energy constraints. Simulation results demonstrate that the proposed framework achieves higher search efficiency and strictly satisfies energy safety requirements compared with several baseline methods in dynamic target search scenarios.

Keywords: Optimization and Optimal Control, Machine learning, Logistics
Abstract: Stowage planning plays a vital role in steel enterprises, serving as a critical link between production warehousing and maritime transportation. This paper proposes RA-MODE, a surrogate-assisted multi-objective differential evolution algorithm, to optimize stowage plans for steel products. We formulate a multi-objective integer programming model to maximize outbound efficiency, customer concentration, and cargo weight, while respecting maritime loading rules. To enhance search efficiency, we introduce an adaptive MODE featuring a multi-operator mutation pool and a dominance-relation prediction mechanism based on Support Vector Regression (SVR). The proposed method is validated on real-world data from a major Chinese steel enterprise and demonstrates strong potential for practical deployment in maritime logistics operations.

Keywords: Optimization and Optimal Control, Motion Control
Abstract: We present a unified framework that turns a set of 6-DoF inspection viewpoints into a time-optimal, collision-free route for a 9-DoF robotic system. Unlike modular pipelines that fix a single inverse-kinematics (IK) configuration per viewpoint, build an all-pairs travel-time map, and then route, our method jointly optimizes the visiting order and the per-viewpoint configuration in a single global search. The three-dimensional self-motion manifold of each viewpoint is parameterized in closed form so that the pose constraint holds by construction, the rest-to-rest travel time is approximated by a closed-form admissible double-integrator surrogate, and the tour is encoded by random keys. A derivative-free optimizer (CMA-ES) minimizes a cheap penalized objective over order and configuration, after which direct-collocation trajectory optimization is applied only to the edges of the selected route to certify dynamic feasibility and torque limits, and to return exact timings. This reduces the trajectory solves from quadratic to linear in the number of viewpoints and removes the decoupling that prevents modular pipelines from being globally time-optimal. Simulations and real-robot experiments on a KUKA LBR iiwa with a 2-DoF linear stage validate feasibility, smooth execution, and reduced end-to-end inspection time relative to modular and naive distance-based baselines.

Keywords: Planning, Scheduling and Coordination, Machine learning, Intelligent and Flexible Manufacturing
Abstract: In medium plate production, a slab is first rolled into a mother plate and then cut into multiple subplates to satisfy downstream customer orders. This roll-then-cut process creates a strong coupling between rolling and cutting decisions. However, most existing studies optimize only a single process and do not explicitly incorporate rolling strategy into slab cutting. This paper addresses the slab cutting problem with consideration of rolling strategy. We first formulate the problem as a compact integer programming model and then reformulate it as a pattern-based model by Dantzig-Wolfe decomposition. We further propose a branch-and-price algorithm and several acceleration techniques, including initialization based on dynamic programming with neighborhood improvement, pricing acceleration via dynamic programming reuse and monotonic queues, and an arc-flow branching rule. To further improve computational efficiency, we develop a machine learning-based node selection strategy in which each search node is represented as a bipartite graph and evaluated by a GAT-TE model using direct and cascade decision mechanisms. Computational experiments on instances derived from a real steel plant show that the proposed machine learning based branch-and-price algorithm consistently outperforms both CPLEX and the standard branch-and-price algorithm.

Keywords: Additive Manufacturing, Machine learning, Learning and Adaptive Systems
Abstract: Process condition variation, such as increases in substrate temperature due to heat accumulation, is inevitable in metal additive manufacturing, and periodic re-characterization of the process-geometry relationship is required to maintain geometric/quality consistency of deposited beads. Conventional approaches rely on discrete parameter grids and dense sampling, resulting in high experimental cost and limited adaptability when conditions shift.

We introduce a condition-aware response surface adaptation framework for directed energy deposition that models the bead geometry response surface as a structured and continuous deformation along a process-condition axis. By amortizing the evolution of the laser power (LP) – scan speed (SS) mapping across powder rate variations, the method enables data-efficient reconstruction under few-shot and even zero-shot settings within the sampled powder rate range. Across a 5x5 LP–SS grid with multiple leave-one-out evaluations, the framework achieves stable prediction accuracy while reducing measurement requirements by an order of magnitude compared to full-grid characterization.

Keywords: Robust/Adaptive Control, AI-Based Methods, Computer Vision in Automation
Abstract: Efficient deployment of vision-language-action (VLA) policies is difficult because closed-loop robot control repeatedly invokes multimodal inference at every control step. This paper studies Mixture-of-Experts Vision-Language-Action (MoE-VLA), a lightweight Mixture-of-Experts (MoE) routing framework for efficient VLA inference. The key idea is to treat layer-execution paths as computation experts and select a path from action change, visual variation, and temporal context. Instead of using a fixed pruning rule, MoE-VLA adaptively combines cache reuse, visual-token reduction, and layer skipping. To keep the mechanism easy to analyze, we con- struct a controlled Synthetic Mini-VLA benchmark with four tabletop manipulation task groups and evaluate representative efficiency strategies on the same compact backbone. MoE-VLA further uses an on-policy-distillation-style sparse-path objective to preserve dense-policy behavior under routed execution. Experiments show that MoE-VLA achieves a favorable success– efficiency trade-off across Visual Matching and Variant Aggre- gation settings, reaching 83.1% and 61.0% average success with 2.55× and 2.29× effective speed-up, respectively. These results indicate that MoE-style path routing can reduce redundant computation while preserving task-relevant control behavior.nd jointly necessary.

Keywords: Planning, Scheduling and Coordination, Intelligent and Flexible Manufacturing, Additive Manufacturing
Abstract: Mobile Additive Manufacturing (MAM) offers a transformative solution for supplying critical customized parts by enabling production en route. However, relying purely on MAM encounters spatial bottlenecks when serving geographically dispersed customers. To address this, we propose a novel synergistic supply chain paradigm that integrates the high responsiveness of MAM with the broad coverage of Third-Party Logistics (3PL) networks. The MAM unit acts as a roving factory and a dynamic injection point, capable of either delivering parts directly or dropping them off for 3PL distribution. We formulate a mixed-integer linear programming model to jointly optimize batch nesting, machine scheduling, vehicle routing, and delivery-mode selection. Computational experiments demonstrate that the synergistic mode performs superior to centralized AM and pure MAM paradigms. Insights highlight that strategically utilizing 3PL injections for dispersed or less urgent orders help overcome the spatial limitations of single mobile unit.

Keywords: Additive Manufacturing, Planning, Scheduling and Coordination, Optimization and Optimal Control
Abstract: This paper studies the single-machine one-dimensional batch scheduling problem for additive manufacturing with the objective of minimizing total tardiness. To solve this problem exactly, we develop a tailored branch-and-bound (B&B) algorithm that integrates hybrid lower-bounding, adaptive node selection, and dominance-based pruning mechanisms. Computational experiments show that the proposed B&B generally outperforms Gurobi in both efficiency and scalability. The results further reveal opposite computational effects of platform capacity: a smaller platform makes the mixed-integer linear program formulation harder for Gurobi, while improving the efficiency of the proposed B&B algorithm.

Keywords: Additive Manufacturing, Hybrid Strategy of Intelligent Manufacturing, Calibration and Identification
Abstract: 。为解决现有熔融沉积建模（FDM）多材料3D打印中显著的材料浪费、长切换时间和交叉污染问题，本文提出了一种新型欠致动热端切换系统。采用动力再利用机制，系统利用打印机X/Y轴的相对运动驱动机械锁定和释放机构，无需在刀头上额外安装电机。这种方法显著降低了硬件成本和终端执行器的质量，同时实现了高效的热端切换。为确保切换时的高重复性，引入了基于麦克斯韦原理的运动学耦合定位机构。此外，开发了基于触控探针的自动喷嘴偏移校准系统，以补偿装配误差。实验结果显示切换时间仅为2秒。

Keywords: Planning, Scheduling and Coordination, Learning and Adaptive Systems, Factory Automation
Abstract: Disruption recovery in industrial assembly lines requires timely decisions under machine faults, worker absence, and emergency orders. Existing methods either rely on rigid handcrafted recovery logic or learn adaptive policies that do not readily exploit heterogeneous external recovery knowledge at decision time to reduce abnormal recovery time (ART) and preserve on-time delivery (OTD). To address this gap, we propose a phase-aware guidance injection framework that augments a trained recurrent MAPPO (RMAPPO) scheduling policy through logit-level action bias during evaluation. The framework provides a unified decision-time interface for rule-based, replay-based, and online LLM-based guidance, while activating intervention only during abnormal and recovery phases. Experiments on a custom AssemblyLineEnv show that high-quality rule guidance yields the strongest gains, replay-based guidance degrades smoothly under imperfect availability, and online LLM guidance still provides useful intermediate improvements. These results show that decision-time guidance injection can exploit heterogeneous recovery hints without redesigning the actor.

Keywords: AI-Based Methods, Intelligent and Flexible Manufacturing, Hybrid Strategy of Intelligent Manufacturing
Abstract: Addressing challenges of semantic fragmentation and weak engineering constraint modelling in existing process planning knowledge and Large Language Model (LLM) fusion for intelligent manufacturing, a Structure-Constrained Knowledge Reasoning framework is proposed. Firstly, process documents are decomposed into semantically coherent chunks and organised into an explicit topology capturing sequential dependencies, structural associations, and engineering constraints. Secondly, a structure-constrained hybrid retrieval method then combines semantic search with topology-guided expansion to reconstruct complete and coherent process evidence from dispersed information. Thirdly, an engineering consistency-oriented agent response mechanism is introduced to guide the LLM through intent recognition and constraint-aware reasoning, ensuring that generated recommendations satisfy domain-specific feasibility and execution requirements. Experimental results demonstrate that the framework outperforms state-of-the-art methods in knowledge retrieval and engineering consistency response. In particular, it effectively resolves information omission and enhances factual reliability, providing a logically consistent decision knowledge support in reliability-critical manufacturing environments.

Keywords: Collaborative Robots in Manufacturing, Task Planning, Assembly
Abstract: Traditional approaches to two-sided assembly line balancing often ignore the complexities introduced by human-robot collaboration, particularly stochastic operation times and the associated ergonomic risks. This study addresses this problem by developing a novel multi-objective optimization model that integrates these critical, real-world factors. The model explicitly aims to minimize three key objectives: cycle time, total ergonomic fatigue, and the smoothness index. To handle the inherent uncertainty in operation times, we employ chance-constrained programming, while ergonomic risk is dynamically assessed through muscle fatigue-recovery modeling. For optimization, a decomposition-based multi-objective evolutionary algorithm (MOEA/D) with Tchebycheff scalarization is adopted. Through a detailed case study on keyboard assembly and a subsequent sensitivity analysis on operation time variance, the results reveal distinct trade-offs between production efficiency and ergonomic risk. This work provides a robust methodological framework and valuable insights for designing assembly systems that are not only efficient and robust but also human-centric.

Keywords: Intelligent and Flexible Manufacturing, AI-Based Methods, Task Planning
Abstract: This paper addresses the blocking-constrained Distributed Assembly Flowshop Scheduling Problem (DAFSP) with the objective of minimizing the makespan. A mixed-integer linear programming (MILP) model is first established. To solve the problem efficiently, a Heuristic-Based Heterogeneous Graph Neural Policy (H2GNP) is proposed. In H2GNP, the scheduling state is represented by a heterogeneous graph, based on which a heuristic-fused Markov decision process (MDP) is constructed. The policy network consists of a graph-based encoder and an autoregressive decoder. The encoder uses multilayer residual graph convolutions, while the decoder combines a Gated Recurrent Unit (GRU) with scaled dot-product attention. In addition, problem-specific heuristics are further integrated to guide job-to-factory assignment and insertion. The policy is trained via Proximal Policy Optimization (PPO) on instances of varying scales to enhance scalability. Experimental results show that H2GNP achieves the minimum makespan in 20 out of 24 large-scale instances. Moreover, H2GNP exhibits superior efficiency, requiring only 0.005% of the runtime of baselines on the largest instances.

Keywords: Intelligent Transportation Systems, Plug-in Electric Vehicles
Abstract: We study the charging service scheduling problem for workplace electric vehicle charging stations equipped with vehicle-to-grid capability and on-site photovoltaic generation. The problem is challenging due to uncertainties in building load demand and photovoltaic generation, which significantly affect charging scheduling. Existing approaches either ignore uncertainty or rely on overly conservative robust formulations, potentially leading to inefficient decisions. To address this issue, we propose a covariate-driven target-oriented robust satisficing framework that minimizes the worst-case deviation from a predefined operating cost target under distributional ambiguity. The framework leverages observable contextual information, such as weather and calendar variables, to dynamically construct uncertainty bounds for building load and photovoltaic generation. Numerical experiments based on real-world data show that the proposed approach consistently achieves better out-of-sample performance than traditional benchmarks. Further analysis reveals that V2G capability provides increasing economic benefits as time-of-use price spreads widen.

Keywords: AI-Based Methods
Abstract: Knowledge Graph Question Answering (KGQA) enables access to structured industrial knowledge, but Additive Manufacturing (AM) queries often contain dense terminology and multi-conditional constraints that cause semantic confusion and hallucinated reasoning. We propose a two-stage collaborative reasoning framework that combines a fine-tuned M3E retriever with an instruction-tuned LLaMA-3-8B discriminator. The retriever is refined through general LCQMC semantic strengthening followed by AM-specific contrastive learning, while the LLM filters retrieved candidates under KG constraints. Experiments on MechaQA achieve a Question Hit (Hit) of 0.9504 and a Mean Triple Recall (Recall) of 0.9557, with further validation on a 3D printing subset.

Keywords: AI-Based Methods
Abstract: With the recent progress of Large Language Models (LLMs), there has been paradigm-shifting improvement in natural language understanding, significantly outperforming traditional Knowledge Graph Question Answering (KGQA) methods. However, when deployed in specialized vertical domains, domain knowledge deficits of general-purpose LLMs lead to factual hallucinations. Moreover, their black-box generative nature lacks explicit logic chains with poor reasoning interpretability. To address these serious issues, we propose a new Retrieve-and-Discriminate framework that combines the global structure injection technique with local evidence reasoning for better factual accuracy as well as explainability. In particular, to address the lack of domain knowledge, we employ ComplEx embeddings for precise candidate selection and propose a Knowledge Embedding Injection method by soft prompts, fusing continuous graph structure directly into the LLM. We explicitly build multi-hop reasoning path as local evidence to clarify the reasoning logic. Lastly, we adopt a discriminative instruction tuning approach rather than openended generation and essentially restrict hallucination. Our extensive experiments show that our approach reaches 82.2% Hits@1 on the general WebQSP benchmark, and a dominant 90.8% on the industrial Mecha-QA dataset.

Keywords: AI-Based Methods, Plug-in Electric Vehicles, Modelling, Simulation and Validation of Cyber-physical Energy Systems
Abstract: Restless Multi-Armed Bandit (RMAB) is a powerful tool for modeling decision-making problems with weak-coupled components under resource constraints. However, solving RMABs is challenging due to the exponentially large state and action spaces. Though index policies could solve large-scale RMABs effectively by assigning an index value to each arm, designing such policies requires substantial domain expertise. Additionally, existing data-driven methods for learning such policies require specific task properties and often lack interpretability. To address these challenges, we propose using Large Language Models (LLMs) to design and refine interpretable index policies. Our approach eliminates the dependence on difficult-to-verify problem properties and requires no expert knowledge. The generated index policies are presented as executable code in natural language, ensuring interpretability. Experimental results demonstrate the effectiveness of the proposed framework and the scalability of the generated policies. Additionally, this framework can be extended to produce more generalized index policies, like multi-level index policies, showing the potential for practical applications.

Keywords: Modelling, Simulation and Validation of Cyber-physical Energy Systems, Plug-in Electric Vehicles, Simulation and Animation
Abstract: In cyber-physical energy systems (CPES), electric vehicles (EVs) have emerged as a critical coupling interface between transportation and power distribution networks. Their growing penetration intensifies charging demand, while existing studies often overlook traffic dynamics in charging behavior analysis. This paper therefore develops an improved dynamic user equilibrium (IDUE) model for EV charging behavior analysis. The main contributions of this paper are as follows. First, a traffic equilibrium model for mixed vehicle categories is constructed by comprehensively considering road congestion, charging time cost, and electric vehicle state of charge (SOC) constraints during route selection. This achieves a unified description of traffic flow evolution and energy state changes. Second, an EV feasible path screening model is established, incorporating an energy constrained feasible path search and charging station selection mechanism to enhance the model's representation of real-world charging behavior. Finally, a modified flow-swapping algorithm is proposed to solve the equilibrium model. Through collaborative computation involving dynamic network loading (DNL) and path flow updating, the locations and SOC information of EVs under equilibrium conditions are obtained. The model is validated using Nanjing's transportation network as a case study. Simulation results demonstrate that the proposed method provides decision support for coordinated scheduling of traffic-energy coupled systems and optimization of charging infrastructure.

Keywords: AI-Based Methods, Model Learning for Control
Abstract: Under the “dual-carbon” context, the high penetration of photovoltaic (PV) power places increasing demands on both the accuracy and reliability of short-term forecasting. However, strong nonlinearity, pronounced day–night zero inflation, and weather-driven uncertainty significantly limit the effectiveness of conventional point prediction models. This paper proposes a probabilistic PV forecasting framework that integrates multi-scale deep feature extraction with hybrid uncertainty modeling. A CNN–LSTM–Attention architecture is employed to capture local irradiance variations, temporal dependencies, and rapid weather transitions. On this basis, a hybrid output layer combining Gaussian negative log-likelihood and Zero-Masked Quantile Regression (GNLL+ZM-QR) is designed to jointly learn the conditional mean, variance, and quantiles while excluding nighttime zero samples from quantile optimization, thereby mitigating zero-inflation-induced bias and yielding physically consistent prediction intervals. Experiments on 13 months of real-world PV data demonstrate that the proposed method consistently outperforms LSTM, random forest, and CNN-LSTM baselines in both point accuracy and probabilistic reliability. The resulting prediction intervals remain stable under sunny, cloudy, and extreme weather conditions, supporting grid scheduling, energy storage management, and high-penetration renewable integration.

Keywords: Reinforcement, Modelling, Simulation and Validation of Cyber-physical Energy Systems, Power and Energy Systems automation
Abstract: Reinforcement learning (RL) in complex continuous-control tasks is highly sensitive to how observations are represented. Existing representation methods typically compress observations holistically, without explicitly distinguishing action-responsive factors from environment-driven variation. We propose the State Decomposition Model (SDM), which disentangles observations into action-responsive internal states and disturbance-driven external states through a factorized transition prior that conditions external evolution on the updated internal state. This control-aware decomposition integrates with standard RL algorithms (PPO, SAC, TD3) without modifying their objectives. Experiments on MuJoCo benchmarks and an IEEE 118-bus power-grid regulation task show that SDM consistently outperforms VAE-based and forward/inverse-dynamics baselines, suggesting that SDM provides a practical structured inductive bias for RL, especially in control problems with strong coupling and persistent exogenous disturbances.

Keywords: Renewable Energy Sources
Abstract: The large-scale integration of wind power plays a pivotal role in the low-carbon transition of power systems, yet it also introduces significant volatility and uncertainty. Assessing the resilience and reliability of power systems under such conditions requires the consideration of extreme wind power output scenarios. However, due to the rarity of extreme meteorological events, such scenarios are seldom observed in historical datasets. To address this challenge, this paper proposes an improved conditional diffusion model for generating realistic extreme wind power scenarios. The proposed framework explicitly embeds the operational constraints of wind turbines into the diffusion process to ensure the physical feasibility of the generated scenarios. In addition, a hierarchical training strategy is introduced to enhance scenario diversity, enabling the model to capture multiple extreme patterns. Experimental results demonstrate that the proposed model achieves better physical consistency and extreme scenario representation than the baseline models selected in this study.

Keywords: Demand Side Management, Planning, Scheduling and Coordination, Energy and Environment-aware Automation
Abstract: Data centers are under increasing pressure to improve operational efficiency amid growing computing demand and rising energy consumption. Among different workload types, batch workloads constitute an important source of scheduling flexibility because their execution can be adjusted over time while satisfying completion requirements. However, effective utilization of this flexibility requires explicit characterization of both the progressive and resumable nature of batch-task execution and the coupling among batch-task progression, thermal conditions, and cooling demand, which remain insufficiently addressed in existing studies. To bridge this gap, this paper develops an integrated model that unifies batch-task execution, server-side resource provisioning, and zone-level thermal-cooling dynamics. In this model, batch-task progress can be preserved across scheduling periods through suspend--resume execution, while zone thermal states are explicitly incorporated to capture their influence on batch-task progression and cooling operation. The resulting coordinated scheduling problem is formulated to minimize data-center operating cost subject to workload completion and thermal safety constraints. Numerical results show that the proposed method reduces electricity cost, total energy consumption, peak power, and cooling burden while maintaining workload completion feasibility and thermal safety. Moreover, resumable execution and thermal-performance coupling are shown to provide complementary benefits, whose joint exploitation enables more efficient and balanced data-center operation.

Keywords: Deep Learning in Robotics and Automation, Telerobotics and Teleoperation, Human Factors and Human-in-the-Loop
Abstract: We present a Real-Time Operator Takeover (RTOT) paradigm that enables operators to seamlessly take control of a live visuomotor diffusion policy, guiding the system back to desirable states or providing targeted corrective demonstrations. Within this framework, the operator can intervene to correct the robot’s motion, after which control is smoothly returned to the policy until further intervention is needed. We evaluate the takeover framework on three tasks spanning rigid, deformable, and granular objects, and show that incorporating targeted takeover demonstrations significantly improves policy performance compared with training on an equivalent number of initial demonstrations alone. Additionally, we provide an in-depth analysis of the Mahalanobis distance as a signal for automatically identifying undesirable or out-of-distribution states during execution. Supporting materials, including videos of the initial and takeover demonstrations and all experiments, are available on the project website.

Keywords: Differentiable Physics-based Granular Material Manipulation, Motion Control of Manipulators, Optimization and Optimal Control, Robust/Adaptive Control of Robotic Systems
Abstract: Automating the manipulation of granular materials poses significant challenges due to complex contact dynamics, unpredictable material properties, and intricate system states. Existing approaches often fail to achieve efficiency and accuracy in such tasks. To fill the research gap, this paper studies the small-scale and high-precision granular material digging task with unknown physical properties. A new framework, named differentiable digging robot (DDBot), is proposed to manipulate granular materials, including sand and soil. Specifically, we equip DDBot with a differentiable physics-based simulator, tailored for granular material manipulation, powered by GPU-accelerated parallel computing and automatic differentiation. DDBot can perform efficient differentiable system identification and high-precision digging skill optimisation for unknown granular materials, which is enabled by a differentiable skill-to-action mapping, a task-oriented demonstration method, gradient clipping and line search-based gradient descent. Experimental results show that DDBot can efficiently (converge within 5 to 20 minutes) identify unknown granular material dynamics and optimise digging skills, with high-precision results in zero-shot real-world deployments, highlighting its practicality. Benchmark results against state-of-the-art baselines also confirm the robustness and efficiency of DDBot in such digging tasks.

Keywords: Domain-specific Software and Software Engineering, Cloud Computing For Automation
Abstract: Validation of robotic systems critically depends on the operating conditions under which they are assessed. Scenario selection and variation are often manual, experience-driven, and difficult to scale, which harms reproducibility and weakens validation conclusions. We propose a scenario-based methodology that models scenarios compositionally and formalizes how these dimensions are varied, instantiated, executed, and interpreted. Building on this, we introduce RoboVAST, a framework that realizes declarative campaign specifications, plugin-based scenario generation, and scalable containerized execution with integrated result analysis. We demonstrate the approach with a navigation dataset comprising 5480 scenario configurations and over 100000 execution runs across five indoor maps with varied paths, sensor noise, software parameters, and obstacle settings, totaling more than 1800 hours of simulated operation and 1873 km traveled. Twenty repetitions per configuration allow us to distinguish systematic failures from stochastic anomalies.

Keywords: Domain-specific Software and Software Engineering, Software, Middleware and Programming Environments, Big-Data and Data Mining
Abstract: Robotic systems generate large volumes of multimodal sensor data, but converting ROS bag recordings into machine learning datasets is often handled by ad hoc sequential scripts, creating engineering overhead and slow iteration cycles. We model dataset construction as an artifact-based build process over a dependency graph and implement this approach in Bagzel, an open-source Bazel extension for reproducible, incremental dataset generation (including nuScenes-format export). We compare Bagzel and Bagzel-xattr (server-side digest management) against a sequential rosbag2nuscenes baseline. Bagzel reduces runtime in all evaluated execution modes, with the largest gains in iterative workflows (up to 386.26x in warm builds and 7.21x in incremental builds on a 20.4 GB dataset). Across dataset sizes from 5.1 to 20.4 GB, Bagzel variants show markedly better scaling behavior than the baseline, especially in warm and incremental modes. Bagzel-xattr provides additional gains, with a mean runtime reduction of 5.9% compared to Bagzel in the input granularity study. Overall, modeling robotics dataset construction as an artifact-based build process substantially reduces dataset update latency while maintaining a deterministic build design that supports reproducibility.

Keywords: Energy and Environment-aware Automation, Logistics, Industrial and Service Robotics
Abstract: Logistics tasks in sectors such as agriculture and construction often require transporting objects across environments with uneven terrain and obstacles. Although legged robots provide mobility on challenging terrain, their locomotion can be inefficient on flat surfaces where wheeled robots are more suitable. In this work, we present a new platform based on the Dual-Diablo design, a cooperative wheel-legged robotic system designed for flexible and energy-efficient delivery in challenging environments. The system combines two biped-wheeled robots with a rigid tray to transport payloads while enabling multiple locomotion strategies, including wheeled driving, legged walking, and body height adjustment to traverse obstacles. We implement coordinated locomotion gaits that allow the robots to operate together while maintaining payload stability. In addition, we derive a data-driven power consumption model that predicts energy usage as a function of robot velocity, terrain slope, payload mass, and body configuration. This allows the system to utilize the right gait based on the terrain. The proposed design and power model are validated through simulation and laboratory experiments, demonstrating the feasibility of the system for flexible and energy-efficient logistics operations in environments containing traversable obstacles.

Keywords: Factory Automation, Task Planning, Compliant Assembly
Abstract: We present a dynamic robotic packing approach for autonomously and densely packing unknown objects that arrive continually at a packing site, such as those delivered via a conveyor belt. Our system enables simultaneous online object perception, rapid object modeling, dense packing planning, and object picking and packing in the presence of multi-modal sensing uncertainties, approximations in rapid object modeling, and robot motion inaccuracies. The method can accommodate both rigid and certain non-rigid objects, including in scenes where items are piled. Our approach is validated in real experiments with packing tasks involving unknown objects of various shapes and types. Experimental results demonstrate the effectiveness and efficiency of the introduced method in achieving high-density packing for different kinds of objects.

Keywords: Motion and Path Planning, Optimization and Optimal Control, Motion Control
Abstract: Traditional mobile manipulation systems typically decouple base navigation from arm control, which weakens whole-body coordination and limits their ability to operate in complex environments. While recent research has begun exploring integrated control frameworks, many approaches still rely on rigid collision constraints or are tailored to specific scenarios, resulting in limited generalization and poor real-time performance. To address these challenges, we propose a Holistic Model Predictive Control (HMPC) framework that jointly optimizes the mobile base and manipulator to leverage their complementary capabilities. We present a unified optimization formulation that unifies task representations across scenarios and enables flexible task switching through control matrix adjustments. Meanwhile, a micro-spherical decomposition method is introduced to construct collision-aware soft constraints, enabling anticipative, geometrically precise obstacle avoidance with low computation. Furthermore, the unified optimization is embedded into a modular hierarchical architecture to improve adaptability for diverse tasks. Experimental results in both simulation and real-world platforms demonstrate that HMPC outperforms traditional sampling-based planners, achieving a 90.0% task success rate, generating smoother and shorter trajectories, and reducing risky motions. In dynamic tracking tasks, it also achieves better real-time performance with consistently lower end-effector errors. These results highlight HMPC as a unified, efficient, and generalizable control solution for mobile manipulation in complex environments.

Keywords: Collaborative Robots in Manufacturing, Task Planning, Planning, Scheduling and Coordination
Abstract: Modern manufacturing increasingly relies on heterogeneous multi-robot systems, yet coordinating collaboration across both processing and logistics tasks remains challenging. Existing studies typically address these two domains separately and often overlook the fact that a task may admit multiple execution modes, which can lead to inefficient schedules. To address these limitations, this paper proposes a unified framework for the heterogeneous multi-robot collaborative task allocation and scheduling problem. We develop a mixed-integer linear programming (MILP) model that captures diverse collaboration patterns, ranging from tightly synchronized joint assembly to sequential relay transport. By explicitly modeling alternative coalition-based execution modes, the framework enables the optimization of the trade-off between execution efficiency and resource consumption. A stator assembly case study validates the framework in a representative processing-strong/logistics-weak collaborative manufacturing scenario and demonstrates a 31.4% reduction in makespan compared with a non-collaborative baseline. These results show that jointly optimizing collaboration across processing and logistics tasks can significantly improve the utilization of shared heterogeneous robot resources and enhance overall system efficiency.

Keywords: Optimization and Optimal Control, Process Control
Abstract: Timed event graphs (TEGs) are widely used to represent synchronization-driven systems in manufacturing. In such systems, undesirable initial conditions often induce excessive token delays during transient operation, leading to degraded performance. This paper studies a finite-horizon control approach that enforces convergence to a desired 1-cyclic schedule while penalizing excessive token residence times. We formulate the problem as a linear program and investigate how the choice of transient horizon influences delay penalties, and derive bounds that quantify this relationship. Numerical observations indicate that the proposed bounds closely capture achievable performance in practical settings, particularly in robotized manufacturing systems for semiconductor manufacturing.

Keywords: Computer Vision in Automation, Data fusion, Sensor-based Control
Abstract: To address the inconsistency in wafer Automated Optical Inspection (AOI) where cross-machine imaging variations cause the same defect to be classified into different categories, and the practical constraint of single-shot capture (single-modal inference) in actual production, this paper proposes a detection framework based on privileged knowledge learning. First, a coarse-to-fine wafer alignment method is employed to construct a multi-modal paired dataset. Second, a dual-backbone YOLO11 network is built to fuse multi-modal features and generate privileged information. Finally, a two-stage Modality-Dropout Self-Distillation (MDSD) strategy is designed to transfer knowledge from the multi-modal teacher network to the single-modal student network through feature consistency constraints. Experimental results demonstrate that, while maintaining the single-modal inference efficiency required for actual production, the proposed method improves cross-machine detection consistency by 3.35% compared to the baseline model.

Keywords: Semiconductor Manufacturing, Petri Nets for Automation Control
Abstract: Efficient scheduling of cluster tools, the equipment central to semiconductor manufacturing, is critical for productivity. As circuit architecture geometries shrink, chambers require periodic cleaning to remove residues, which forces chambers idle and breaks the optimality of conventional dual-armed schedules. The swap sequence, optimal without cleaning, is particularly affected, motivating partial-loading variants swap(z), swap(a,z), and the m-wafer cyclic swap(A,z). Yet swap(A,z) still leaves a positive-gap band of width 2(m+1)(w+v) on the (p,c)-plane, growing linearly with the chamber count m, and whether this width can be reduced has remained open. In this paper, we first experimentally identify the minimum achievable width as (m+1)(w+v), then analyze the robot sequence structure that attains it: defining robot operations as permutations of the four elementary chamber tasks, classifying them by feasible initial robot state, and identifying the compatible operation set that achieves this minimum width. A computer-assisted symbolic proof verifies the boundary for m = 2, ..., 8 and all w,v > 0.

Keywords: Planning, Scheduling and Coordination, Optimization and Optimal Control, Logistics
Abstract: Automated Pick-and-Place (PnP) robotic systems in dynamic sorting environments require advanced production planning. Traditional scheduling focuses on throughput maximization, neglecting the heterogeneous economic values of objects and complex multi-depot routing. Addressing this gap, we propose a value-aware scheduling framework formulating the dynamic multi-robot sorting process as a Time-Dependent Team Orienteering Problem with Time Windows (TDTOPTW).

We first develop a Mixed-Integer Linear Programming (MILP) model within a rolling-horizon framework to capture continuous object inflows, time-dependent kinematics, and multi-depot assignments. To overcome the computational intractability of exact methods in real-time operations, we introduce a scalable Genetic Algorithm (GA). Using a dual-vector encoding and feasibility-based decoding mechanism, the GA efficiently navigates the continuous search space, avoiding discretization-induced resolution loss while enforcing spatiotemporal constraints.

Extensive experiments demonstrate the GA operates within a strict 1-second limit, significantly outperforming conventional dispatching rules (FIFO, SPT, Greedy) in maximizing total economic value. Additionally, the GA ensures a well-balanced workload across the robotic fleet, mitigating uneven mechanical wear and enhancing long-term operational reliability in capacity-constrained environments.

Keywords: Sustainability and Green Automation, Modelling, Simulation and Validation of Cyber-physical Energy Systems, Optimization and Optimal Control
Abstract: 受碳峰值和中和目标驱动，工业园区加速了清洁能源转型。然而，绿色蒸汽和氢气的生产份额仍然较低，因为由于效率低下和可再生能源波动，经济性较低。高温热泵和热储能为经济高效的绿色氢气和蒸汽共生产提供了有前景的方式，但其集成规划在组件匹配、复杂度建模和计算可处理性方面面临挑战。本文提出了一种基于光伏驱动的绿色蒸汽-氢供给系统，结合多种蒸汽发电和热储存技术。基于温度-压力耦合及热存储占地与经济学之间的权衡，构建了最优规划模型，重新表述为混合整数线性规划问题。案ߴ

Keywords: AI-Based Methods, Collision Avoidance, Agent-Based Systems
Abstract: Stability is a cornerstone of legged robot performance, yet rigorous evaluation remains challenging. Beyond algorithm development, systematic testing is essential. However, conventional approaches are hindered by high variance in outcomes, the safety-critical cost of failures, and the prohibitive sample sizes required for statistically valid estimates under Monte-Carlo testing. This creates a persistent gap where unbiased performance estimates demand more trials than high fidelity standardized testing can afford. To address this, we introduce an importance sampling framework tailored for stability testing of legged robots. More importantly, we propose the use of template models, long established in locomotion control design but not previously applied to testing, as a principled means to construct effective importance distributions. Our approach enables accelerated and statistically sound stability tests, achieving substantial sample reduction while preserving unbiased estimates. Experiments with a Unitree G1 humanoid and a bipedal robot, Cassie, in MuJoCo demonstrate that template inspired methods provide superior efficiency.

Keywords: Discrete Event Dynamic Automation Systems, Reinforcement
Abstract: A unified formulation of controller synthesis for discrete event systems is generalized to include probabilistic systems as well. Controllers for Markov decision processes (MDPs) are formulated such that temporal logic properties hold with probability 1. The unification is based on mu-calculus and a minor extension of CTL*. MDPs are fundamentally nondeterministic, such that a single action can lead to multiple target states when probabilistic distributions are introduced. This nondeterminism cannot be handled by ordinary mu-calculus, but a simple extension that introduces temporary uncontrollable events solves this problem. The proposed MDP control strategy is also shown to be useful in reinforcement learning, where temporal logic properties can be guaranteed in a data-efficient manner.

Keywords: Smart Grids, Reinforcement
Abstract: Reinforcement learning (RL) has become a widely used approach for microgrid energy management. However, selecting an appropriate RL algorithm for a specific operating scenario is still largely based on trial and error, and no single algorithm consistently performs best across all conditions. This paper proposes a knowledge graph (KG)-based framework for automated RL algorithm selection in multi-scenario microgrid energy management. A domain-specific ontology is constructed to capture algorithm characteristics, scenario features, and their performance relationships. By introducing Rice's Algorithm Selection Problem (ASP) framework into RL-based energy management, the proposed method enables explainable and scenario-aware algorithm recommendations. The framework is validated through case studies on single-agent dispatch and multi-agent voltage regulation.

Keywords: Optimization and Optimal Control, Discrete Event Dynamic Automation Systems
Abstract: Policy-gradient (PG) methods are the most prevalent approaches in reinforcement learning. However, their theoretical understanding in finite-horizon Markov decision processes (MDPs) remains limited, primarily due to the nonconvexity of the performance objective and the dominance of backward induction in dynamic programming. This paper studies the monotonic, non-asymptotic global linear convergence of projected policy gradient (PPG) methods in finite-horizon MDPs under arbitrary constant step sizes. Motivated by the intrinsic connection between PG and policy iteration (PI), we first propose a new PI method for finite-horizon MDPs that enables simultaneous policy improvement across all decision stages. Under parallel computation, the proposed method exhibits more favorable scalability than classical dynamic programming, particularly for problems with large action spaces and long horizons. Subsequently, we establish the global optimality of first-order stationary points under natural policy parameterization using an analytical method, without invoking gradient dominance or Polyak–Łojasiewicz conditions. Building on this result, we show that PPG enjoys monotonic, non-asymptotic linear convergence for arbitrary step sizes. Numerical experiments corroborate our theoretical results.

Keywords: Optimization and Optimal Control, Manipulation Planning
Abstract: Retargeting human hand motions to dexterous robot hands from monocular video remains a challenging task due to the inherent noise, depth ambiguity, and frequent occlusions in vision-based hand pose estimation. Existing physics-based methods often fail when the input reference trajectory violates physical constraints, leading to unstable optimization, severe object penetration, or loss of contact. In this paper, we propose a robust optimization framework that explicitly addresses perception uncertainty to achieve physically plausible and contact-rich dexterous manipulation. By leveraging reliable object pose estimates (e.g., from FoundationPose) as strong anchors and treating hand keypoints as soft, uncertainty-weighted targets, our method recovers feasible robot trajectories even from noisy visual inputs. We introduce a contact-implicit refinement strategy that utilizes the geometric relationship between the hand and the object to dynamically correct hand pose errors. Furthermore, we integrate a physics-consistent post-processing stage to ensure temporal smoothness. We evaluate our approach on a self-collected monocular hand-object interaction dataset and the public OakInk benchmark using the Inspire Hand. Experimental results show clear gains over both kinematic IK and SPIDER, with substantially reduced penetration depth, improved contact stability, and smoother joint trajectories, bridging the gap between noisy visual perception and precise robotic control.

Keywords: Planning, Scheduling and Coordination, Human Factors and Human-in-the-Loop, Optimization and Optimal Control
Abstract: Multi-objective scheduling requires decision makers (DMs) to select a single optimal schedule from a set of Pareto optimal solutions, a process inherently subject to cognitive stochasticity. While traditional preference elicitation relies on simple binary choices, this limits the expression of nuanced human feedback. To address this, we propose a framework that leverages large language models to warm-start from open-ended natural language requests and extract rich supplementary information, such as preferences, strengths, bounds, and directional signals. This extracted data is seamlessly incorporated into a Gaussian Process surrogate model to dynamically handle decision noise and efficiently guide the search. Experiments demonstrate that our approach outperforms standard baselines, achieving over a 60% regret reduction.

Keywords: Collaborative Robots in Manufacturing, Intelligent and Flexible Manufacturing, Discrete Event Dynamic Automation Systems
Abstract: This work studies real-time dynamic scheduling in a robotic assembly cell, where multiple robots process jobs transported by a circular conveyor belt. Each robot also observes only a limited local region of the system; the problem naturally involves decentralized decision-making under partial observability. We formulate this problem as a decentralized partially observable Markov decision process and propose a graph-based multi-agent reinforcement learning approach for adaptive online scheduling. The proposed policy represents observable jobs and robots as graph embeddings and updates them through Transformer-based self-attention and cross-attention modules. By enabling all robots to make decisions simultaneously with shared parameters, the framework remains scalable with respect to the number of jobs, robots, and observable regions. Experimental results demonstrate that our approach outperforms conventional heuristic algorithms.

Keywords: Planning, Scheduling and Coordination, Factory Automation, Learning and Adaptive Systems
Abstract: Deep reinforcement learning (DRL) has achieved strong performance on the flexible job shop scheduling problem (FJSP), with existing methods typically trained on small-scale instances randomly generated from synthetic distributions. In real production environments, however, the incoming jobs are drawn from a predefined set of product types, each with fixed processing times and machine eligibilities. We exploit this prod uct mix information by fine-tuning a pretrained DRL scheduler on mix-sampled instances. On a 20×10 recipe validation set, mix-aware fine-tuning reduces makespan by 7.86% relative to the best zero-shot pretrained baseline and reaches the final performance of training from scratch in a small fraction of the training steps.

Keywords: Semiconductor Manufacturing, Manufacturing, Maintenance and Supply Chains
Abstract: With the rapid evolution of global semiconductor supply chains, maintaining efficient and stable fab-level operations is critical to manufacturing performance. As core processing platforms in advanced wafer fabrication, cluster tools face increasing scheduling complexity under high-mix, low-volume production. This study addresses the transient scheduling problem of dual-armed cluster tools processing two wafer types concurrently with explicit consideration of wafer residency time constraints (WRTCs). A mixed-integer programming (MIP) model is developed to optimize wafer release order and robot task sequence with the objective of minimizing total completion time. Numerical results show that dual-armed cluster tools consistently outperform single-armed configurations. Overall, the proposed approach provides a practical solution for enhancing operational efficiency and strengthening the operational resilience of dual-armed cluster tool scheduling under transient production conditions.

Keywords: Semiconductor Manufacturing, Inventory Management, Logistics
Abstract: This paper addresses a long-term outsourcing planning (LTOP) problem in semiconductor back-end manufacturing. Customer demand is specified for each product over time, typically at a daily resolution across the planning horizon. Given this demand, planners must decide how much to outsource, when to process, and where to route multiple products through outsourced bumping (BUMP) and electrical die sorting (EDS) stages. These decisions must respect operational constraints such as lead times, yield, and production rates, as well as financial constraints such as monthly spending limits on outsourcing expenditures. A key challenge arises because yield factors smaller than one are repeatedly applied across multiple processing stages and compounded with production rates, making downstream supply planning difficult when production units are indivisible. To address this issue, we propose a methodology that constructs implementable production plans based on guaranteed supply quantities while accounting for yield and production rates. We formulate the problem as an integer multi-commodity flow model, whose objective balances total shortage and deviation from monthly outsourcing budget targets. To solve large instances, we propose a fix-and-optimize matheuristic that iterates through three main steps: solving a linear programming relaxation, converting fractional flows into feasible integer flows, and applying partial variable fixing. Computational experiments on industry-inspired instances demonstrate that the proposed algorithm produces scalable and physically implementable plans under tightly coupled operational and financial constraints.

Keywords: Semiconductor Manufacturing, Planning, Scheduling and Coordination, Machine learning
Abstract: We address a scheduling problem for semiconductor equipment installation. Each equipment installation is treated as a project with a due date, where the project consists of multiple stages. Each stage contains multiple activities connected through complex precedence relations. Each activity has a processing time and must be assigned to one of its eligible teams, whose capabilities may affect processing time. The scheduling decisions involve (1) assigning activities to teams and (2) sequencing them within each team to minimize total tardiness under shared resource constraints. To address the scalability challenge, we propose an imitation learning (IL) approach that learns scheduling policies from high-quality constraint programming (CP) solutions on small instances. To the best of our knowledge, this is the first learning-based approach for large-scale semiconductor equipment installation scheduling. The trained policy consistently outperforms CP, dispatching rules, and reinforcement learning (RL) on large-scale instances.

Keywords: Agent-Based Systems, Machine learning, Reinforcement
Abstract: 3D multi-UAV pursuit-evasion games are a fundamental problem in cooperative autonomous systems. However, coordinating heterogeneous swarms, such as long-range Scouts and short-range Interceptors, in obstacle-rich 3D spaces remains challenging due to asymmetric kinematic limits, heterogeneous sensing capabilities, and obstacle-induced sensing occlusions. Traditional Heterogeneous Graph Neural Networks (HGNNs) structurally isolate parameter spaces for different types of agents, which can limit parameter sharing and data efficiency in continuous physical coordination tasks. To address this issue, we propose a lightweight type-aware framework, Type-Aware Graph Attention Multi-Agent Proximal Policy Optimization (TAG-MAPPO). TAG-MAPPO combines a shared graph backbone for universal kinematic representation with an early-modulation layer placed before message passing, allowing agent-type information to affect subsequent graph attention weights. Simulation results show that TAG-MAPPO achieves competitive capture performance while reducing episode length, collision rate, and energy cost compared with representative baselines. We further evaluate zero-shot transfer across different pursuer population sizes, showing structural adaptability to UAV dropouts and reinforcements without retraining.

Keywords: Automation Technologies for Smart Cities, Agent-Based Systems, Learning and Adaptive Systems
Abstract: Collaborative rogue UAV interception is critical for urban airspace security, yet remains a longstanding challenge due to the spatiotemporal asynchrony inherent in heterogeneous interdiction resources. Although learning-based approaches promise greater adaptability compared to traditional methods in complex scenarios, they remain susceptible to forced temporal alignment and often struggle to disentangle individual contributions during multi-agent collaboration. To address these problems, we propose an adaptive collaborative defense method named Asynchronous Counterfactual Credit Assignment Multi-Agent Reinforcement Learning (ACCA-MARL). First, we introduce a Duration-Aware Virtual Synchrony (DAVS) mechanism, which seamlessly aligns asynchronous operations by evaluating macro-actions with varying durations, ensuring robust temporal representation against mismatched trajectories. Second, we design a Counterfactual Causal Credit Assignment (C3A) module by leveraging do-calculus baselines. This module explicitly filters out spurious correlations, enabling the model to adaptively isolate the marginal causal contribution of each distributed defense node. Through extensive simulation experiments on a multi-target urban defense dataset, ACCA-MARL demonstrates superior performance over multiple evaluation metrics, including neutralization efficiency and interdiction resource depletion.

Keywords: Formal Methods in Robotics and Automation, Motion and Path Planning, Telerobotics and Teleoperation
Abstract: While generative AI has revolutionized digital content creation, extending this creative capability to physically embodied robots remains a significant challenge. Robotic painting, which accomplishes physical artworks by planning stroke trajectories, is a crucial direction for enhancing embodied creative capabilities. Existing methods, using imitation learning or Stroke-based Rendering (SBR), either heavily rely on extensive expert demonstrations with poor generalization or focus on mechanical pixel-level approximations that lack human-like skills. Inspired by human combinatorial generalization—the ability to create novel works by recombining previously acquired skills—we propose "Learn Few, Paint Many," a novel Real2Sim2Real computational experiment framework for embodied robotic painting. Our approach extracts object-specific painting skill primitives from limited expert data. For unseen painting tasks, the system conducts computational experiments in a virtual environment by semantically selecting painting primitives and iteratively optimizing their combination parameters under compositional and aesthetic constraints. Finally, it recommends the optimal painting scheme and guides the robot to execute the physical creation in the real world. Experimental results demonstrate that our method successfully achieves combinatorial generalization, synthesizing novel, high-quality artworks using only a minimal set of expert skills and outperforming existing methods.

Keywords: Motion and Path Planning, Intelligent Transportation Systems, Collision Avoidance
Abstract: Offline Multi-Agent Path Finding plans for multiple Unmanned Aerial Vehicles (UAVs) Unmanned Traffic Management (UTM) can fail during disturbed execution because stochastic delays break the joint timing assumptions that make nominal plans conflict-free. We present a closed-loop discrete UTM method that couples UTM-aware planning with online execution repair. It extends Lazy Constraints Addition Search for MAPF (LaCAM) with time-augmented joint states and successor filtering for temporal no-fly zones, heterogeneous protection footprints, and directional downwash risk envelopes. At execution time, Online Discrete Alignment and Replanning (ODAR) aligns observed UAV states to reference plans, applies bounded one-step repair and conflict-aware arbitration, and triggers replanning when local correction is insufficient. In discrete simulation with up to 300 UAVs, the proposed method preserves sub-second planning on benchmark maps. Under stochastic delays, offline execution and alignment-only variants leave measured execution conflicts, while full ODAR reduces the discrete conflicts to zero in the tested scenarios. Local conflict-set replanning matches global replanning in completion while using 5.1% to 21.0% of the global cumulative replanning time per case and 9.4% in aggregate.

Keywords: Reinforcement, AI-Based Methods, Cyber-physical Production Systems and Industry 4.0
Abstract: Dynamic flexible job shop scheduling (DFJSP) with job insertions is a representative challenge in modern manufacturing systems, where dispatching decisions must continuously adapt to evolving shop-floor states. In such event-driven settings, effective reinforcement learning hinges on state representations that can simultaneously capture (i) intra-job technological precedence and (ii) inter-job resource contention on shared machines. This paper proposes Structure-Aware Graph Feature Extraction with PPO (SAGFE-PPO), a learning-based dispatching framework for DFJSP under stochastic job insertions. SAGFE-PPO constructs an incremental operation precedence graph to encode stable process constraints under job insertions, and models real-time machine competition to characterize dynamic resource contention at each decision epoch. The resulting structure-aware embeddings are fused with pairwise operation–machine features and optimized using proximal policy optimization to select feasible assignments. SAGFE-PPO consistently achieves lower makespan than representative DRL schedulers on dynamically generated DFJSP instances with Poisson job insertions, while maintaining comparable inference latency. Gantt-chart analyses further reveal enhanced machine utilization and reduced schedule fragmentation under insertion disturbances.

Keywords: Agent-Based Systems, Optimization and Optimal Control, AI-Based Methods
Abstract: Dynamic processes in industry are typically run in time-varying environments with multiple conflicting objectives, leading to significant challenges in real-time optimization. Traditional scalarization-based approaches rely on fixed objective weights, and hence they fail when the working conditions change. Evolutionary multi-objective algorithms are capable of approximating the Pareto front, but their high computational cost prevents real-time applications. To address these issues, this paper proposes an objective-specific multi-agent reinforcement learning framework for dynamic multi-objective cooperative optimization. In the proposed method, each agent is assigned to one optimization objective and learns an individual policy, while an attention-based action fusion mechanism coordinates shared control decisions among agents. This design avoids manually defined static reward weights and improves adaptive trade-off capability under dynamic conditions. Experiments are performed with real-world data of cement clinker calcination using three objectives: minimization of unit heat consumption, minimization of unit electricity consumption, and maximization of clinker output. The results demonstrate that the proposed method outperforms the Weighted Sum, NSGA-II, and attention-based multi-agent baselines with respect to the quality of the Pareto solutions, training stability, and convergence speed. The proposed framework offers a powerful and scalable framework for real-world dynamic industrial multi-objective optimization.

Keywords: Human-Centered Automation, Collaborative Robots in Manufacturing, Intelligent and Flexible Manufacturing
Abstract: Collaborative assembly robots must adapt to diverse human partners while respecting the hard precedence constraints that govern assembly processes. Existing reinforcement learning (RL) approaches explore the full action space at every step, including structurally impossible actions, wasting training signal and slowing convergence. We present a framework combining task-graph-guided action masking with Proximal Policy Optimization (PPO), semantic voxel observations, and mixed-population training against parameterized human behavioral profiles. A directed acyclic graph encodes task prerequisites; a binary mask derived from the graph and the partner's concurrent actions constrains policy sampling to valid actions at each timestep, yielding up to 2.5× faster convergence than unmasked baselines. We evaluate on a two-phase block-stacking task. In Phase 1 (behavioral adaptation), a single adaptive policy achieves 100% task success across all human types, outperforming specialized per-profile policies by 9.5% in episode return, with specialized policies overfitting to predictable partners. In Phase 2 (flexible task allocation), the policy reaches 94% overall success; we identify and analyze a mutual-waiting failure mode rooted in training distribution bias and propose targeted mitigations. Together, these results establish task-graph-guided masking as a practical and scalable foundation for human-adaptive collaborative assembly.

Keywords: Motion and Path Planning, Actuation and Joint Mechanisms, Manufacturing, Maintenance and Supply Chains
Abstract: Accessing confined spaces such as storage tanks, process vessels and aircraft wings for performing inspection and maintenance tasks present significant safety risks and operational challenges for human workers. This paper introduces VORAC, a novel 7-DOF modular redundant robotic manipulator specifically designed for adapting limited entry and unstructured nature of these spaces. It features a unique kinematic configuration combining a prismatic joint, a twisting joint, and five oblique swivel (OS) joints. A key design feature is an internal tubular access path that enables the delivery of utilities from base to the end-effector. The system is guided by an end-effector-mounted RGB-D camera for 3D environmental perception. The paper presents Particle Swarm Optimization (PSO) based path planning algorithm that enables collision-free passage for the manipulator through provided waypoints. Vision-based surface feature extraction algorithms were developed for waypoint identification, specifically by detecting circular openings and geometrically modeling pipe bends from point cloud data. The effectiveness of the integrated system is validated with a functional prototype, demonstrating successful semi-autonomous navigation in emulated confined space scenarios.

Keywords: Robotics and Automation in Life Sciences
Abstract: Robotic automation is essential for scaling autonomous scientific discovery, where long-term experiments demand reliability, reproducibility, and high throughput. A routine, yet challenging, operation in these labs is rack insertion, a contact-rich task requiring millimeter-scale precision and high sensitivity to error. Existing platforms often depend on fiducial markers or fixed placements, which limits flexibility and scalability in human-centric lab environments. To address these limitations, we present a mobile robotic system for high-precision, vision-guided rack insertion into an automated liquid handler (Opentrons OT-2). The platform consists of a Franka Emika Panda arm mounted on a dockable trolley, enabling modularity and extended workspace reach across different laboratory stations. The core of our approach is a zero- shot 6D pose estimation pipeline followed by a closed-loop vision-based refinement stage that iteratively corrects residual misalignments. Evaluated across four rack types and six OT-2 deck slots, the system achieved an average insertion success rate of 83.1%, demonstrating a practical end-to-end solution for autonomous rack handling in laboratory environments.

Keywords: Motion and Path Planning, Reactive and Sensor-Based Planning, Autonomous Vehicle Navigation
Abstract: Narrow indoor logistics corridors create a tension in local navigation: reactive planners avoid unexpected obstacles but may oscillate or lose corridor consistency, while predictive trackers maintain geometric accuracy but become brittle when the planned corridor segment is obstructed. We propose a corridor-aware hybrid navigation architecture with three cou- pled components: a modified TEB local planner with corri- dor and footprint-clearance penalties, a constrained linearized MPC tracker with corridor-width-derived lateral bounds and bidirectional tracking for docking, and a feasibility-driven supervisor that switches between controllers using obstacle probing, corridor connectivity, and dwell-time hysteresis. The method is validated in Gazebo simulation and on a Clearpath Jackal robot across docking, undocking, and obstacle-avoidance scenarios. Results show corridor-consistent motion, reliable handover between reactive and predictive control, and safe recovery when passages become physically impassable.

Keywords: Learning and Adaptive Systems
Abstract: Aerial manipulators enable physical interaction in hard-to-reach environments; however, the combined problem of direct whole-body aerial manipulation under rapid arm motion, payload changes, and related unknown dynamic uncertainty remains a largely unsolved problem. We present a hierarchical control framework that combines Reinforcement Learning (RL) with an inner-loop dynamics estimator to address this problem. The RL outer loop maps desired 6-degrees-of-freedom (DOF) end-effector targets to coordinated whole-body commands, enabling direct task-driven control without relying on a fully accurate coupled dynamic model in the policy layer. An inner loop then tracks these commands while compensating for transient inertial shifts and uncertainty during execution via a dynamics estimator scheme without requiring system model knowledge. We validate the proposed approach on a custom quadrotor equipped with a 3-DoF manipulator through hardware experiments under varying payload conditions. Compared with RL+PID and RL+INDI+PID baselines, the proposed method reduces end-effector tracking error and improves task success rate across the tested hardware conditions. These results show that combining learned whole-body coordination with estimator-based low-level compensation improves the precision and robustness of aerial manipulation under changing operating conditions.

Keywords: Collaborative Robots in Manufacturing, Computer Vision for Manufacturing, Deep Learning in Robotics and Automation
Abstract: Small and medium enterprises (SMEs) often face challenges in adopting novel robotized concepts due to the cost and complexity of conventional automation technology. This paper presents the development of a low-cost robotic manipulation system, designed to support flexible assembly in SME manufacturing. The proposed approach considers the use of a commercial off-the-shelf collaborative robot, for which a low-cost end-effector is customized using additive manufacturing with an eye-in-hand industrial monocular camera. Vision-based object localization is performed using a YOLO11 detector, enabling the estimation of grasp poses for flexible robot manipulation. To compensate for the absence of depth information, the system uses the robot's force feedback for contact detection during grasping. The complete perception-control pipeline was implemented and initially validated through a simple automated assembly task involving multiple components. The results demonstrate the feasibility of the proposed approach as a cost-effective solution for smart robotic manipulation in SMEs manufacturing.

Keywords: Computer Vision for Transportation, Computer Vision in Automation, Deep Learning in Robotics and Automation
Abstract: Monocular 3D object detection is attractive for autonomous systems because it offers low-cost, easy-to-integrate perception, but recovering metric geometry from a single image remains ill-posed, especially under occlusion, truncation, and class imbalance. Recent methods improve depth and context reasoning, yet size and depth estimates remain coupled and unstable in ambiguous unified multi-class settings. We propose MonoPRIO, a prior-conditioned extension of MonoDGP that uses offline class-aware size prior banks and query-level adaptive prior routing to stabilise size estimation. MonoPRIO combines routed mixture priors with uncertainty-aware log-space size conditioning and Cluster Aligned Prior (CAP) regularisation on matched positives to keep predictions near plausible prototype manifolds. Prior construction is offline, enabling lightweight online inference without additional prior-feature extraction. On KITTI, MonoPRIO achieves the strongest fully reported unified multi-class result on the official KITTI benchmark test split and the best unified validation medians among compared methods. It improves unified accuracy over MonoDGP while preserving a near-identical compute budget, and a compact mechanism ablation shows that routed prior injection and CAP jointly improve the accuracy--size-fidelity trade-off.

Keywords: AI-Based Methods, Computer Vision in Automation, Machine learning
Abstract: Visual anomaly detection is a fundamental task in industrial automation. While existing approaches have achieved notable progress in identifying structural defects, the detection of logical anomalies remains relatively underexplored. In practice, structural and logical anomalies frequently co-occur in industrial workflows. Therefore, a solution capable of detecting both structural and logical anomalies is crucial for advancing comprehensive anomaly detection research. To address this limitation, we propose a unified framework, termed UniSLAD, which jointly addresses logical and structural anomalies without additional training, enabling a practical solution for dynamic industrial environments. First, we introduce a dual-feature extractor that synergistically integrates a Convolutional Neural Network (CNN) backbone for local texture perception with a Transformer backbone for global contextual reasoning, yielding richer and more comprehensive representations. Building on this foundation, we design dual-granularity feature representation modules. At the patch level, memory banks enhanced by the Mahalanobis Transform (MT) preserve representative features and support more discriminative anomaly scoring. At the image level, distribution maps are aggregated using Lower–Upper Mean (LUM) and Power Mean Pooling (PMP), yielding a more robust global representation than conventional average pooling. Extensive experiments on the two industrial benchmarks demonstrate that UniSLAD achieves competitive performance in comprehensive anomaly detection, achieving 99.4% and 93.1%, respectively. Furthermore, ablation studies verify the individual contributions and effectiveness of each proposed component.

Keywords: Computer Vision for Manufacturing, Computer Vision for Transportation, Computer Vision in Automation
Abstract: Monocular 6D pose estimation of metallic parts remains a difficult problem in industrial environments, where strong specular reflections and non-uniform illumination create a pronounced sim-to-real gap. The challenge is particularly critical for small and medium enterprises (SMEs), where new parts are introduced frequently and collecting annotated real data for each object is impractical.

We present a controlled stress-test study of sim-to-real transfer for reflective industrial parts, supported by a fidelity-aware synthetic dataset pipeline that integrates high-resolution texture recovery, realistic illumination modeling, and bounded domain randomization. Using three gearbox components and three extreme illumination setups, we provide 9,000 real test frames to evaluate how synthetic-only training generalizes under challenging factory conditions (Project page and dataset: https://synth-metal.github.io/).

The study shows that fidelity-aware synthetic data improves robustness compared to both default synthetic and real-only datasets, especially for highly reflective objects under cross-scene shifts. Furthermore, moderate dataset sizes with bounded variation are more effective than very large or unconstrained datasets. Reflective and less reflective objects also reveal complementary trends, suggesting that synthetic and real data are best used together. These findings expose the hardest cases of sim-to-real transfer and offer practical insights for efficient small-and-medium enterprise deployment.

Keywords: Swarms, Autonomous Agents, Robot Networks
Abstract: This paper investigates the problem of consensus optimization in networked unmanned aerial vehicle (UAV) systems. Departing from conventional graph-theoretic models that employ binary edge weights, we propose a more realistic framework where the wireless signal-to-noise ratio (SNR) of each communication link defines a continuous, probabilistic edge weight. This model accurately captures the non-uniform and stochastic nature of real-world wireless channels. Our primary contribution is a novel distributed reinforcement learning algorithm based on the Boltzmann-Gibbs distribution (BGRL). The system state is characterized by a vector of UAV transmit powers and their respective distances to a fixed ground node—parameters that directly govern the link SNR. Simulation results demonstrate that the proposed BGRL framework is robust to variations in channel distribution and achieves consensus approximately an order of magnitude faster than a centralized benchmark using the ADAM optimization method.

Keywords: Additive Manufacturing, Computer Vision for Manufacturing
Abstract: In modern manufacturing, implementing autonomous systems is crucial for ensuring production efficiency and quality stability. Wire Arc Additive Manufacturing (WAAM) is widely used in aerospace, shipbuilding, and defense due to its rapid fabrication of large-scale metal components and high material efficiency. However, arc instability and complex thermal histories can cause surface defects, including excessive oxidation, bead discontinuity, and waviness deposition, compromising reliability. To enable real-time quality management, we developed a GUI-based inspection program providing live defect visualization and diagnostics for WAAM components, supporting immediate decision-making in the WAAM process. Building on this, a precision vision inspection system was implemented using YOLOX on multi-resolution CMOS camera images of additively manufactured components. To improve generalization under limited data, data augmentation techniques including Mosaic and MixUp were applied. Quantitative comparison across YOLOX scales showed that YOLOX-X achieved the highest performance of 94.69% under augmented conditions, demonstrating its suitability for WAAM precision inspection. The proposed system supports operator-driven process parameter adjustment via the GUI upon defect detection, and reduces the subjectivity of visual inspection while reducing non-destructive testing costs.

Keywords: Logistics, Planning, Scheduling and Coordination, Agent-Based Systems
Abstract: The growing demand for small-lot, multi-variety e-commerce deliveries has intensified operational workloads in warehouse picking operations. Goods-to-Person (GTP) systems employing Autonomous Mobile Robots (AMRs) have been widely adopted to address these challenges. While substantial research has focused on task allocation and path planning for AMRs, the problem of order allocation—assigning ordered items to workstations—has received relatively little attention, despite its significant influence on system throughput. This paper proposes a novel order allocation method for dynamic GTP systems based on two-sided matching theory. Specifically, we extend the Deferred Acceptance algorithm to balance AMR transport efficiency and workstation utilization, leading to significant improvements in system throughput. Extensive simulations demonstrate that the proposed approach consistently improves throughput compared with conventional baseline methods, particularly under rapidly changing order arrivals. In addition, the computation time remains sufficiently low for real-time deployment.

Keywords: Planning, Scheduling and Coordination, Optimization and Optimal Control
Abstract: ，高效的任务分 7197; 对于改进至关ດ ; 5;要 操作效率和系统 1283;定性 自动化生产和物 7969;系统。传统 分配模型主要侧 7325;于最小化分配 忽视工作负荷平 4913;的成本，可能ल ;8;致 资源过载和运营 9942;颈。前往 针对这一问题，64 12;文研究一个多௤ ;6;标 负载均衡的广义 5351;派问题 （MOGAP-LB），其中总 6171;值成本和 最大负载占用率 1516;时被最小化。为了解决这个问 9064;，采用了问题ल ;8;向的帕累托 提出了进化算法 5288;PEA）。拟议 方法整合了三个 0114;补的修复规则 在容量限制下高 5928;恢复可行性 同时促进工作量 0340;均衡分配。在加法，基于εÊ 48;Ó 83;的选择性局๽

Keywords: Planning, Scheduling and Coordination, Optimization and Optimal Control
Abstract: A white light interferometer is a sensor suitable for automating inspections of components requiring high measurement accuracy. These sensors pose specific challenges for viewpoint planning that have not yet been addressed in the literature. This paper presents a novel approach to solving the view planning problem using gradient-based optimization. The goal is to determine a high-quality set of viewpoints for complete surface coverage. A surrogate optimization problem is formulated to maximize the coverage given a certain number of viewpoints. By iteratively increasing the number of viewpoints and re-optimizing all previous viewpoints together with the new one, a set of viewpoints achieving complete coverage is obtained. The proof of concept is demonstrated on representative test structures.

Keywords: Probability and Statistical Methods, Learning and Adaptive Systems, Big-Data and Data Mining
Abstract: Reliable prediction of key product properties in industrial processes is often hindered by the scarcity of labeled data and non-uniform noise distributions. To address these challenges, this paper proposes a semi-supervised regression framework based on Manifold Adaptive Conformal Prediction (MACP). The core of MACP lies in a two-stage model architecture that leverages the geometric manifold of extensive unlabeled data to train two independent Laplacian-regularized estimators. These estimators simultaneously identify the conditional mean and characterize the heteroscedastic error distribution, enabling the model to adaptively adjust Prediction Intervals (PIs) according to local data density and noise levels. A key theoretical contribution is that MACP provides distribution-free, finite-sample coverage guarantees, ensuring statistical reliability without stringent distributional assumptions. Experimental evaluations on industrial steel strip mechanical property datasets demonstrate that, compared to state-of-the-art methods, the proposed method consistently achieves the target coverage with significantly narrower prediction intervals.

Keywords: Reinforcement, Deep Learning in Robotics and Automation, AI-Based Methods
Abstract: Reinforcement learning in discrete-continuous hybrid action spaces presents fundamental challenges for robotic manipulation, where high-level task decisions and low-level joint-space execution musthref be jointly optimized. Existing approaches either discretize continuous components or relax discrete choices into continuous approximations, which suffer from scalability limitations and training instability in high-dimensional action spaces and under domain randomization. In this paper, we propose Hybrid TD3, an extension of Twin Delayed Deep Deterministic Policy Gradient (TD3) that natively handles parameterized hybrid action spaces in a principled manner. We conduct a rigorous theoretical analysis of overestimation bias in hybrid action settings, deriving formal bounds under twin-critic architectures and establishing a complete bias ordering across five algorithmic variants under synchronized Gaussian error assumptions. Building on this analysis, we introduce a weighted clipped Q-learning target that marginalizes over the discrete action distribution, achieving equivalent bias reduction to standard clipped minimization while improving policy smoothness. Experimental results demonstrate that Hybrid TD3 achieves superior training stability and competitive performance against state-of-the-art hybrid action baselines. The resources are available at https://github.com/thanhnguyencanh/HybrAJ.

Keywords: Reinforcement, Learning and Adaptive Systems, Optimization and Optimal Control
Abstract: Multiobjective Markov decision processes (MOMDPs) are widely used in complex scenarios, where multiple conflicting objectives need simultaneous optimization. Multiobjective reinforcement learning (MORL) algorithms have emerged as a powerful framework to seek a complete Pareto front (PF) with different compromise policies that can balance the multiple conflicting objectives. This paper proposes a multiobjective reinforcement learning algorithm based on decomposition and adaptive weight (MORL/D-AW). MORL/D-AW transforms a given MOMDP into a set of single-objective subproblems, each of which is a standard Markov decision process (MDP), and subsequently applies the proximal policy optimization (PPO) algorithm to solve each subproblem. Since the true PF is usually complex and even unknown in practical problems, we propose a weight adaptive strategy to adjust weight vectors periodically. This enables the adaptive redistribution of subproblem weight vectors to achieve more uniform solutions, leading to the enhancement of population diversity. Meanwhile, we adopt a parameter-transfer mechanism to facilitate cooperation among subproblems, which can improve the convergence of MORL/D-AW. Experimental results on typical benchmarks demonstrate that MORL/D-AW outperforms state-of-the-art MORL algorithms in terms of the diversity and quality of the obtained solutions.

Keywords: Reinforcement, Machine learning, Manufacturing, Maintenance and Supply Chains
Abstract: This paper investigates the dynamic Multi-period Two-Dimensional Non-Guillotine Cutting Stock Problem with usable leftover (MP-2D-NGCSP) in the context of steel production. In this problem, slabs and orders arrive dynamically in batches over discrete time steps, where only information from the current period is observable. Under constant volume conditions, the slab is rolled into a mother plate, which is then cut to produce the order plates. In each period, planners must allocate slabs to orders while ensuring all current orders are dispatched. The primary objective is to minimize the total occupied mother plate area across all periods. To solve this problem efficiently, we formulate the MP-2D-NGCSP as a Markov Decision Process (MDP) and propose a hybrid framework combining Deep Reinforcement Learning with Integer Programming (DRLIP). The DRLIP decomposes each time step into a dynamic selection phase and a static allocation phase. In the former, a heterogeneous graph encoder, a position decoder, and an order decoder are utilized to output selected positions and their corresponding sequences of orders. In the latter, a two-dimensional knapsack problem with coupling is formulated as an Integer Programming (IP) model to optimize order subsets based on the order decoders’ outputs. A heuristic algorithm is designed to solve the IP model. Experimental results demonstrate that, compared to existing methods for solving the MP-2D-NGCSP, our DRLIP occupies less slab area.

Keywords: AI-Based Methods, Cognitive Automation, Agent-Based Systems
Abstract: Finite element method (FEM) stress analysis is widely used in engineering design and maintenance evaluation. However, practical use of FEM stress analysis results still relies on expert interpretation. Experienced analysts apply tacit reasoning procedures that integrate visual stress observations with geometric and physical constraints, making FEM interpretation difficult to model as a simple image recognition task. In this paper, we propose a cognition-grounded reasoning framework for modeling expert interpretation of FEM stress analysis using multimodal large language models. Through interviews with structural analysts and analysis of practical review workflows, we extract expert tacit knowledge and formalize it as a four-stage cognitive process. Based on this model, we introduce a Co-Agent architecture in which each reasoning stage is implemented as a role-specialized vision-language agent. To evaluate the approach, we construct a dataset for expert interpretation of piping stress analysis results derived from practical FEM review workflows. Experimental results show that structuring expert reasoning procedures improves interpretation accuracy compared with direct prompting, while the proposed multi-agent architecture further improves both prediction accuracy and reasoning stability.

Keywords: AI-Based Methods, Machine learning
Abstract: As smart industry evolves toward human-centric collaboration, large language models are increasingly expected to function as cognitive agents that actively support complex reasoning and cooperative decision-making. However, although existing search-based reasoning methods can improve performance on complex reasoning tasks, their high computational cost and latency make them unsuitable for real-time collaborative decision-making. To address this limitation, we propose COMPASS (COMpressed Policy for Action-Selective Search), an efficient reasoning framework designed for complex knowledge reasoning and collaborative decision-making scenarios. COMPASS employs offline Monte Carlo Tree Search to obtain high-quality reasoning trajectories and compresses the embedded high-value decision patterns into a action selector, which then operates jointly with a language reasoning model to enable structured inference without explicit online search. Experimental results on three benchmarks (GSM8K, GSM-Hard, and StrategyQA) across four open-source LLMs show that COMPASS consistently outperforms CoT-based and self-consistency methods, while achieving performance comparable to search-based methods with token consumption reduced by approximately 95%-97%. Further analysis indicates that COMPASS can learn stable high-level decision patterns, enabling efficient and goal-directed reasoning in complex multi-step tasks.

Keywords: Human Factors and Human-in-the-Loop, Computer Vision for Manufacturing, AI-Based Methods
Abstract: Automated defect inspection in manufacturing often encounters edge cases where machine predictions are unreliable and human judgment becomes necessary. The CHIPS framework, which couples few-shot learning with the Segment Anything Model(SAM) through automated prompt generation, addresses this by routing low-confidence cases to human operators. However, operators currently receive no decision support during intervention, and a measurable performance gap between experienced and novice inspectors has been observed. This work-in-progress paper proposes augmenting CHIPS with a Large Language Model(LLM)-based advisory module that provides guidance adapted to each operator's skill level in real time. The module takes visual context from the inspection pipeline together with an evolving operator profile as input and tailors its recommendations accordingly. Prior CHIPS evaluation data documented a 4.5% Intersection-over-Union gap between experts and novices on challenging cases; we argue that targeted guidance can narrow this gap substantially. We present the proposed architecture, discuss deployment considerations for industrial settings, and outline a validation plan.

Keywords: Intelligent and Flexible Manufacturing, Robust Manufacturing, Manufacturing, Maintenance and Supply Chains
Abstract: This paper delves into the boundary synchronization control problem for multi-tiered re-entrant manufacturing networks (RMNs). A hierarchical RMN architecture, consisting of the production line layer, the manufacturing layer, and the workshop layer, is constructed. The system dynamics within this architecture are governed by hyperbolic partial differential equations. The proposed modeling framework explicitly integrates the impacts of time-varying transport delays, boundary input delays, and external disturbances. To address these challenges, a distributed boundary synchronization control strategy is developed, using local and neighboring boundary measurements. By employing Lyapunov-Krasovskii functionals, sufficient conditions are derived to ensure H_infty disturbance attenuation of the closed-loop RMN system. Numerical simulations are provided to validate the effectiveness of the proposed approach.

Keywords: Cyber-physical Production Systems and Industry 4.0, Planning, Scheduling and Coordination, Task Planning
Abstract: Engineers designing production systems need to verify that a given layout supports all required production sequences. Automated planning techniques can answer such questions, but formulating the required planning problems in the Planning Domain Definition Language (PDDL) demands specialized expertise that production engineers typically lack. Asset Administration Shells (AAS) have emerged as the standardized Digital Twin for industrial assets in Industry 4.0. We show that AAS capability models, structured using four established Industry 4.0 standards (VDI3682 for process descriptions, IEC61360-1 for semantic property qualification, IDTA02011 for type hierarchies, and IDTA02016 for instance descriptions), contain sufficient information to generate complete PDDL problems automatically. Unlike prior work that introduced PDDL-specific submodels, our approach derives all planning elements from domain-level descriptions of resource functions, so-called capabilities, allowing engineers to model capabilities without any exposure to PDDL syntax or planning concepts. Our extraction algorithm transforms distributed Multi-AAS architectures into complete PDDL planning problems. We validate the approach on AAS models of a laboratory production system, comparing four layout variants using optimal planning to demonstrate how engineers can systematically explore design trade-offs by modifying the AAS model and regenerating the planning domain.

Keywords: Optimization and Optimal Control, Robust/Adaptive Control, Learning and Adaptive Systems
Abstract: 本文考虑了一个模型预测控制（MPC）问题，适用于在有限数据下具有不确定成本权重参数的离散时间线性系统。引入了以经验分布为中心的Wasserstein模糊性集合，并开发了一个基于分布稳健优化的MPC（DRO-MPC）框架。假设不确定性仅影响目标函数，而系统动态和输入约束则保持确定性。研究表明，DRO的最优值提供了有限样本高置信度的上层证书，证明了真实样本外有限视界预测成本，并且随着样本量增加，逐渐收敛至真实随机MPC问题的最优值。递归可行性和名义渐近稳定性也在可行集合上被建立。数值结果表明，所提出的DR

Keywords: Semiconductor Manufacturing, Factory Automation, Intelligent and Flexible Manufacturing
Abstract: Accurately evaluating manufacturing capacity in semiconductor fabrication is significantly complicated by Queue Time Constraints (QTCs), which are typically ignored by traditional static planning methods. This paper proposes a novel capacity evaluation method to identify the capacity subject to predefined QTCs and violation thresholds. We introduce a capacity decomposition approach that categorizes the capacity of a machine group into actual QTC-feasible capacity, additional capacity required for QTCs, and redundant capacity. This decomposition provides managerial insights by explaining how and where QTCs consume capacity. To quantify the impact of QTCs, we develop an integrated approach combining a simulation module with an iterative configuration adjustment algorithm. Specifically, the simulation estimates stochastic waiting times and evaluates QTC violation rates, while the algorithm determines the optimal system configuration capable of controlling all QTCs subject to predefined violation thresholds. Numerical experiments demonstrate the effectiveness of the proposed method in quantifying the impacts of QTCs on capacity.

Keywords: Semiconductor Manufacturing, Planning, Scheduling and Coordination, Intelligent and Flexible Manufacturing
Abstract: We consider complex job shops with batch processing machines, sequence-dependent setup times, time constraints between consecutive operations, and reentrant process flows. Such constraints are typical in semiconductor wafer fabrication facilities (wafer fabs). A genetic programming approach is designed to discover appropriate machine group- specific dispatching rules for a high on-time delivery per¬formance, large throughput, and small cycle times. Simulation experiments are conducted for a model of a large-scaled wafer fab. The results demonstrate that high-quality dispatching rules are learned.

Keywords: Process Control, Semiconductor Manufacturing
Abstract: In semiconductor manufacturing, production scheduling, quality control, and maintenance decisions are tightly coupled. Limited inspection capacity and equipment degradation make it difficult to ensure product quality while maintaining high efficiency.

This paper proposes an integrated scheduling approach that incorporates sampling and maintenance decisions through threshold-based policies driven by system state. The problem is modeled as a job shop scheduling problem extended with machine degradation and inspection constraints. A genetic algorithm is used to optimize job sequences, while sampling and maintenance are dynamically triggered. Simulation-based evaluation captures both efficiency and risk. Computational experiments show that the approach reduces production risk while preserving scheduling performance.

Keywords: Planning, Scheduling and Coordination, Machine learning, Semiconductor Manufacturing
Abstract: Demand forecasting problems for semiconductor supply chains are studied. We consider different forecasting sources, i.e. forecasts provided by experts, machine learning, or by statistical forecasting methods and different leading indicator values to compute a combined h-step ahead forecast. We introduce a framework based on nonlinear optimization to set internal parameters for the individual forecasting schemes and to blend the forecasts and the leading indicator values. Computational experiments are carried out for a data set of a large semiconductor manufacturer in Europe. The obtained results demonstrate that the framework is able to compute high-quality forecasts.

Keywords: Planning, Scheduling and Coordination, Semiconductor Manufacturing, Optimization and Optimal Control
Abstract: This study addresses a parallel machine scheduling problem for heat treatment processes in Organic Light-Emitting Diode (OLED) display manufacturing, motivated by a real world scheduling problem at one of the largest display manufacturing companies in the Republic of Korea. Heat treatment is a process in which thermal energy is applied to workpieces to achieve the required performance, and temperature control plays a critical role in ensuring process quality. Accordingly, each job must strictly satisfy its lower and upper processing temperature bounds and receive a sufficient level of thermal energy to achieve the required process effect. A key feature of the problem is that the processing temperature of each job is treated as a decision variable. The processing temperature determined for each job defines its processing time, while the setup time is induced by temperature transitions between consecutive jobs. Thus, processing time and setup times are determined by processing temperature and sequencing decisions. The objective is to minimize the makespan. To solve the problem, this study proposes an exact algorithm based on a Branch and Bound (B&B). In particular, strengthened lower bounds are developed to reflect temperature dependent processing times and the setup burden caused by temperature transitions. These bounds reduce the search space and enable the proposed algorithm to obtain optimal solutions efficiently.

Keywords: Planning, Scheduling and Coordination, Logistics, AI-Based Methods
Abstract: Efficient logistics scheduling is essential for synchronizing production and on-site operations in off-site construction (OSC). However, existing static scheduling approaches lack adaptability to dynamic uncertainties and fail to support real-time coordination between manufacturing and site activities. To address this limitation, this study proposes an adaptive two-echelon WAN–LAN scheduling framework for prefabricated module delivery and on-site receiving, which integrates an optimization model with a large language model (LLM)-based agent. The baseline schedule is generated via the WAN–LAN optimization model. Specifically, the WAN echelon optimizes factory loading and truck dispatching decisions, whereas the LAN echelon dynamically rearranges on-site module arrivals to support just-in-time installation. These two interconnected stages aim primarily to maximize valid hoisting throughput. To tackle real-time disruptions, the LLM-based agent monitors system states, detects abnormal events, revises scheduling schemes, and delivers interpretable feedback to field operators, sustaining consistent coordination between factories and construction sites. Experimental case studies verify that the proposed framework enhances scheduling synchronization and robustness amid dynamic disturbances. The results prove its practical applicability to real-world OSC logistics.

This study delivers two key contributions: (1) a two-echelon WAN–LAN scheduling model that jointly optimizes factory dispatching and on-site lifting sequencing; (2) an LLM-driven adaptive decision mechanism capable of real-time disturbance mitigation with human-interpretable feedback.

Keywords: AI-Based Methods, Modelling, Simulation and Validation of Cyber-physical Energy Systems, Deep Learning in Robotics and Automation
Abstract: Multivariate long-sequence time series forecasting (MLSTF) is critical in applications such as electricity consumption planning, where accurate prediction requires models to capture complex temporal dependencies over extended horizons while remaining computationally efficient. Current MLSTF methods are largely built upon Transformer architectures, which, despite their strong performance, suffer from quadratic complexity and high memory usage due to dense self-attention, limiting scalability for long input sequences. To address this issue, we propose GraphReformer, a novel forecasting architecture that replaces the encoder-side dense self-attention module with NRI-based implicit relational inference. GraphReformer models temporal dependencies through probabilistic relational graphs and incorporates a Variational Information Bottleneck (VIB) to preserve prediction-relevant information while suppressing redundancy and noise. Based on the inferred relational structure, the model further performs time-step aggregation to capture both local and long-range dependencies efficiently. Experiments on representative long-sequence forecasting benchmarks demonstrate that GraphReformer achieves competitive or superior forecasting accuracy, yielding an average MSE reduction of approximately 24% compared with strong baseline methods. Meanwhile, it substantially cuts computational and memory costs, demonstrating that relational inference is a promising alternative to dense self-attention for large-scale time series forecasting.

Keywords: Energy and Environment-aware Automation, Renewable Energy Sources, Sustainability and Green Automation
Abstract: Geothermal-based integrated energy systems (GIES) provide a promising approach for space heating. However, shallow ground-source heat pump (S-GSHP) systems and medium-deep ground-source heat pump (MD-GSHP) systems have inherent limitations when used alone. S-GSHP systems suffer from soil temperature attenuation in cold regions, which negatively impacts the operation stability, while MD-GSHP systems are mainly used for heating only, requiring improvements to the annual utilization. To overcome the limitations and enhance overall system performance, this paper investigates the optimal coordination of geothermal-based integrated energy systems coupling medium-deep and shallow ground-source heat pumps (MD-SGSHP), and the complex operation constraints and soil dynamic characteristics are considered to precisely describe energy conversion and geothermal energy utilization during operation. A successive subproblem solving-based optimization (SSSO) algorithm is developed to address the computational challenge. Numerical results demonstrate that the SSSO algorithm achieves a near-optimal solution within finite computational cost, significantly enhancing computational efficiency. Furthermore, while ensuring the sustainable utilization of both medium-deep and shallow geothermal energy, the proposed system reduces annual operation cost by approximately 5%.

Keywords: Optimization and Optimal Control, Planning, Scheduling and Coordination, Cloud Computing For Automation
Abstract: The rapid growth of artificial intelligence (AI) applications is making geographically distributed AI data centers (AIDCs) an important class of large and potentially flexible electricity loads. These loads exhibit high power demand and temporal variability, while their geographical distribution enables part of the aggregate AIDC load to be redistributed across multiple sites. Meanwhile, increasing wind and photovoltaic (PV) generation introduces significant uncertainty into power system operations. However, conventional unit commitment (UC) models typically treat computing loads as inelastic demand and cannot effectively coordinate AIDC load flexibility with generation scheduling.

To address this issue, this paper proposes a multi-timescale coordinated scheduling approach for power systems and geographically distributed AIDCs under renewable uncertainty. The framework determines weekly day-ahead UC decisions with network constraints and performs 15-min rolling economic dispatch (ED) while coordinating AIDC load migration, deferral, and curtailment. A two-stage robust optimization model with a budgeted uncertainty set is developed to hedge against wind and PV uncertainty. Case studies on the IEEE 39-bus system with data-driven AIDC load profiles demonstrate that the proposed approach reduces renewable curtailment and system operating costs under renewable-rich operating conditions with renewable uncertainty.

Keywords: Sustainability and Green Automation, Modelling, Simulation and Validation of Cyber-physical Energy Systems, Power and Energy Systems automation
Abstract: The industrial sector is a major source of carbon emissions, and zero-carbon industrial parks with a high proportion of renewable energy integration have become an important pathway for industrial carbon reduction. However, wind and photovoltaic power exhibit significant long-term randomness, making it difficult for planning based on a single historical year to support long-term investment decisions. To address this issue, this paper proposes a long-term optimal planning method for a hydrogen-based industrial energy system considering renewable energy randomness. A multi-stage planning model is established with phased load growth and staged equipment construction. Meanwhile, low-output events of wind and photovoltaic power are extracted from historical data, and a non-homogeneous Poisson process combined with an Itô process is used to generate long-term renewable energy scenarios. Based on the simulated scenarios, pessimistic, moderate, and optimistic cases are selected for planning analysis. The results show that renewable randomness significantly affects long-term investment behavior. Compared with the optimistic scenario, the discounted investment cost of the pessimistic scenario increases by 7.5%, while the total installed storage capacity increases by 24.3%. In addition, pessimistic renewable conditions lead to more equipment investment in the early stages.

Keywords: Smart Grids, Planning, Scheduling and Coordination, Reinforcement
Abstract: Toward global carbon neutrality, power systems with high penetration of renewable energy sources (RESs) face significant operational challenges due to the inherent uncertainty of RESs. Meanwhile, data centers, with flexible workloads and rapidly growing energy consumption, offer promising potential as flexible loads for supply–demand balancing. This paper studies the interaction between electricity retailers and data centers. An incentive-based Stackelberg game framework is developed to model hierarchical coordination under uncertainty and asymmetric information. A two-phase reinforcement learning-based solution is proposed as an approximate policy learning approach to accelerate equilibrium computation and enable real-time optimization. Preliminary experiments are designed to validate the modeling and algorithmic advantages of the proposed framework.

Keywords: Modelling, Simulation and Validation of Cyber-physical Energy Systems, Renewable Energy Sources, Power and Energy Systems automation
Abstract: A vertical-domain large language model named POMLLM is constructed to integrate equipment mechanisms, scheduling criteria, and engineering experience, thereby generating structured expert knowledge. A hybrid load prediction model is developed to improve forecasting accuracy under small-sample and complex weather conditions. A dual time-scale scheduling framework is established, where hydrogen storage is used for cross-seasonal energy regulation and battery storage for intra-day power balancing. Considering economic efficiency, carbon emissions, and thermal comfort, the model is solved using the Pareto algorithm and rolling optimization. Simulation results demonstrate that the proposed strategy reduces annual operating cost by 4.5%–21.64%, cuts hydrogen heat waste by 15.93%, decreases carbon emissions by 26.38%, and improves user comfort by 28.6%, providing a feasible solution for low-carbon and intelligent operation of building integrated energy systems.

Keywords: Failure Detection and Recovery, Probability and Statistical Methods, Robust/Adaptive Control
Abstract: In this paper, a novel approach to designing stealthy Denial-of-Service attacks, capable of compromising the on-line operations of constrained control systems equipped with resilient control units, is presented. The idea consists of fully exploiting the knowledge that the malicious agent has about the plant model, controller capabilities and statistical properties of the communication medium to elude detection mechanisms. To this end, a crucial role is performed by the Kullback-Leibler Divergence that is in charge of measuring the dissimilarity between two probability density functions. In particular, this operator is used by a malicious agent to build up a proper artificial distribution capable of misleading the resilient control architecture and, as a consequence, to deteriorate the overall control performance by keeping a satisfactory level of stealthiness.

Keywords: Optimization and Optimal Control
Abstract: High-dimensional Expensive Mixed-Variable Optimization Problems (HEMVOPs) are widely prevalent in the design of complex engineering systems, such as Unmanned Aerial Vehicle (UAV) communications and smart grids. These problems are notoriously challenging due to the curse of dimensionality, strong coupling between continuous and discrete variables, and prohibitive cost of each physical simulation evaluations. To address these issues, this work proposes an Ensemble Surrogate-assisted Hybrid Search framework (ESHS). First, a data-driven importance-aware dimension expansion mechanism is proposed to relieve the curse of dimensionality. In addition, an adaptive heterogeneous ensemble model is constructed to reduce simulation cost. Extensive experiments are conducted on two highly complex real-world UAV-assisted network communication problems. The results demonstrate that, under strictly limited evaluation budgets, ESHS outperforms existing state-of-the-art algorithms in both convergence efficiency and global search accuracy.

Keywords: Smart Grids, Planning, Scheduling and Coordination, Diagnosis and Prognostics
Abstract: Power facility inspection is a critical guarantee for the safe and stable operation of power systems and the development of smart grids. Traditional inspection practices are predominantly manual, and while advanced technologies, including Unmanned Aerial Vehicle (UAV), Large Language Model (LLM), and edge computing, have been applied individually, their fragmented deployment hinders the overall intelligent upgrading of power facility inspection. To address this issue, this paper presents a new digital twin-enabled closed-loop intelligent workflow for power facility inspection. With the digital twin platform as the core hub, the workflow integrates diverse edge technologies and intelligent algorithms, and is structured into five interconnected stages that achieve full-process automation, from the parsing of unstructured inspection demands to the reasoning of autonomous maintenance strategies. LLM performance evaluations and a demonstrative system are conducted to validate the framework’s feasibility and effectiveness. The key contribution of this work is the systematic integration of fragmented technologies to form a data-driven intelligent inspection closed loop, which upgrades digital twins from a mere visualization tool to an intelligent decision-making hub. This research provides practical technical reference for smart grid operation and maintenance, and effectively enhances inspection efficiency as well as the resilience of power infrastructure.

Keywords: Deep Learning in Robotics and Automation, Computer Vision for Manufacturing, Computer Vision in Automation
Abstract: Object-level 3D reconstruction is crucial for industrial assembly applications such as digital modeling, virtual inspection, and process analysis. Recently, 3D Gaussian Splatting (3DGS) has emerged as a promising technique for high-fidelity scene reconstruction. However, when applying 3DGS to segmented object images, visual artifacts and noisy structures often emerge, degrading the quality of the reconstructed components. To address this, we propose an object-centric 3DGS framework tailored for these scenarios. Specifically, SAM2-generated masks are utilized to isolate individual components. To enhance quality and suppress noise, a spatial boundary constraint is introduced to restrict Gaussian distribution within the estimated scene extent, preventing primitives from drifting. Furthermore, a floater pruning strategy is applied during training to eliminate degenerate Gaussians, which significantly improves geometric consistency. Experiments on a custom assembly dataset indicate that our approach outperforms existing baselines, producing much cleaner reconstructions. This work provides a highly effective solution for high-fidelity object-level reconstruction in manufacturing contexts.

Keywords: Deep Learning in Robotics and Automation
Abstract: Nonlinear Model Predictive Control (NMPC) achieves high-accuracy trajectory tracking for mobile robots but is often too computationally demanding for resource-constrained platforms. We present a learning-based predictive tracking controller that imitates an expert NMPC policy to achieve real-time performance on differential-drive robots. Our approach combines (a) state-perturbation augmentation with relative-error states to enrich the expert dataset and (b) a lightweight neural architecture based on 1-D convolutions with residual blocks and a physics-aware loss informed by robot kinematics. Expert state--action pairs are collected in a high-fidelity, physics-enabled simulator. Finally, we evaluate the resulting policy in a real-time experiment using a Quanser QBOT differential-drive platform, demonstrating tracking accuracy comparable to NMPC while significantly reducing computation time.

Keywords: Machine learning, AI-Based Methods, Simulation and Animation
Abstract: New construction in densely populated urban areas inevitably disturbs existing buildings. Real-time assessment of the mechanical response of disturbed structures is crucial for construction safety, but traditional finite element analysis is difficult to meet real-time monitoring requirements due to its computationally intensive nature. This article proposes a data-driven surrogate model framework and integrates it into an automated finite element assembly line to achieve fast, accurate prediction of structural force-displacement response. Taking reinforced concrete shear walls as the research object, 25 sets of orthogonal experimental conditions were generated using parameterized finite element models, yielding a dataset of 1025 samples. Five regression models were used for training and comparison, and model performance was systematically evaluated across dimensions such as prediction accuracy, feature importance, and sensitivity to sampling interval. The results indicate that Gaussian process regression has the highest accuracy, but is highly sensitive to sampling intervals; gradient boosting and random forests have lower accuracy but are more robust to sparse sampling. Inference times for all surrogate models are under 1 ms per prediction on standard hardware, compared to several hours per FEM simulation. The framework proposed in this study can provide an effective tool for rapid safety assessment of existing buildings affected by construction disturbances.

Keywords: Diagnosis and Prognostics, AI-Based Methods, Machine learning
Abstract: Neural architecture search (NAS) is increasingly important for reducing manual trial-and-error in deep fault diagnosis model design. This demand is even more pronounced in system-level compound fault diagnosis, where heterogeneous subsystems, cross-component coupling, and unseen fault compositions make architecture design particularly difficult. To address this challenge, we propose a large language model (LLM) guided multi-model collaborative NAS framework for fault diagnosis with compositional generalization. The framework adopts a bilevel optimization strategy: the inner loop learns subsystem-specific representations with dedicated encoders and a global fusion module, while the outer loop uses a LLM to refine architecture configurations based on structured validation feedback under constrained one-step updates. Experiments on the BJTU-RAO metro bogie system-level dataset show that the proposed method outperforms representative baselines and improves compositional generalization accuracy (CGA) over stochastic search, achieving 87.42% CGA versus 83.24% for random-search NAS.

Keywords: Computer Vision for Manufacturing, Zero-Defect Manufacturing
Abstract: In industrial Unsupervised Anomaly Detection (UAD), multi-illumination imaging provides critical cues for revealing subtle defects on complex surfaces that may remain hidden under a single light source. However, existing UAD methods typically process images captured under different lighting conditions independently, without explicitly modeling their underlying correlations, making them vulnerable to extreme specular reflections and self-shadowing. To address this limitation, we propose Synergistic Appearance-Geometry Reconstruction (SAGR), a cross-modal framework for multi-illumination anomaly detection. SAGR jointly models photometric appearance and PS-derived surface normal cues, and further introduces a Cross-Modal Consistency Alignment (CMCA) module to capture their underlying cross-modal correspondence. By exposing discrepancies that violate this learned correspondence, SAGR enables more reliable anomaly detection and localization. Experiments on the M^2AD benchmark show that SAGR achieves superior detection and localization performance with strong robustness to complex lighting and geometric variations.

Keywords: Motion and Path Planning, Industrial and Service Robotics, Formal Methods in Robotics and Automation
Abstract: Robust visual servoing during robotic grasping remains challenging, as approach-time occlusion frequently disrupts closed-loop control. To address this, we present an active observation framework in which an observer arm continuously repositions an RGB-D camera to maintain visibility of both the target and the gripper throughout the grasp. Occlusion during the final approach remains challenging: fixed eye-to-hand lacks a recovery mechanism once the target is hidden, while eye-in-hand suffers from self-occlusion and sensitivity to hand-eye calibration. The observer actively follows the manipulation and, when visibility degrades, explores alternative viewpoints to restore a shared target-gripper view. Because both objects are detected in a single RGB-D image, the resulting Relative Servo Error avoids separately transformed target and gripper estimates. An active two-phase viewpoint policy maintains measurement precision throughout the approach. We evaluate on a dual 6-DoF platform using open-vocabulary detection without task-specific training data. Across 50 randomized trials per condition, the system achieves 94% grasp success versus 78% for a reactive-viewpoint baseline and 64% for a fixed camera. On triggered severe-occlusion trials, recovery takes 1.8 ± 0.6 s on average.

Keywords: Planning, Scheduling and Coordination
Abstract: In aircraft manufacturing, the final assembly line is strictly limited by the physical layout and pulse production mode. This paper investigates the assembly line balancing problem considering complex spatial area constraints. A bi-objective optimization model is established to minimize both cycle time (CT) and the smoothness index (SI). To solve this, a hybrid NSGA-II algorithm is developed, integrating a rigorous spatial-aware decoding mechanism to handle physical capacity limitations and resolve spatial interference. Based on a real-world case comprising 76 tasks, the efficacy of the proposed approach is validated. The numerical results demonstrate that the algorithm consistently achieves highly balanced schedules. This work provides a practical, highly stable decision support tool for complex assembly line balancing under strict spatial constraints.

Keywords: Machine learning, Manufacturing, Maintenance and Supply Chains
Abstract: With the development of industrial 4.0, smart manufacturing equipment generates large volumes of sensor data. However, identifying abnormal part-processing behaviors from these signals remains challenging in many applications due to frequent mode switching, variable cycle durations, and limited labeling in practice. This paper develops a hidden semi-Markov model (HSMM)-based approach for industrial time series data analysis to automatically classify production behaviors during production activities. After training the HSMM, we apply density-based clustering to segment durations within each production run and transition state to identify outliers as the anomalies and use the state sequence to capture spike type events. A case study on real industrial data demonstrates that the proposed method can effectively separate operation modes and detect multiple anomaly types.

Keywords: Smart Home and City, Intelligent Transportation Systems, Sustainability and Green Automation
Abstract: Urban congestion-mitigation planning increasingly requires data pipelines that connect heterogeneous sensing inputs, spatial analysis, and interpretable decision support. This paper treats urban transportation as a cyber-physical system (CPS) and proposes a large language model (LLM)- assisted framework for evaluating green transportation infrastructure. Heterogeneous traffic, road, transit, and demographic data are organized into a CPS pipeline in which sensing- derived records form the perception layer, LLM-assisted feature enhancement supports clustering, Multiscale Geographically Weighted Regression (MGWR) provides regional feedback, and an Entropy Weight Method (EWM)-Technique for Order Preference by Similarity to Ideal Solution (TOPSIS) model provides software-layer plan ranking. The framework is instantiated in a Baltimore case involving three candidate plans: bus-service expansion, bicycle-lane expansion, and congestion pricing. In the proposed workflow, the LLM is not treated as an independent decision maker; instead, it performs lightweight record repair and semantic corridor encoding before spatial clustering, thereby improving the quality of the downstream congestion-analysis input. The empirical results show that the semantic enhancement stage yields more coherent sub- regions and stronger regression performance, while the final decision layer consistently identifies the bicycle-lane plan as the preferred option under both spatial and multi-criteria evaluation. These results support the use of LLM-assisted CPS analytics as a structured decision framework for congestion- mitigation planning rather than as a stand-alone replacement for domain models.

Keywords: Deep Learning in Robotics and Automation, Reinforcement, Motion Control
Abstract: Unmanned aerial vehicles (UAVs) require accurate and robust position tracking in the presence of environmental disturbances. Although outer-loop position control is typically implemented using PID regulators, manual gain tuning is timeconsuming and often leads to suboptimal performance, especially under wind uncertainty and varying operating conditions. In contrast to most existing studies that focus on attitude control adaptation, this paper proposes a reinforcement learning-based adaptive gain tuning approach for position-loop PID control of UAVs without manual intervention. A Deep Deterministic Policy Gradient algorithm is employed and evaluated under two scenarios: nominal conditions without wind and disturbed conditions with wind uncertainty. Simulation results demonstrate that the learned gains consistently outperform hand-tuned PID controllers in terms of trajectory tracking accuracy and transient response. In addition, three different approaches are investigated to further validate the effectiveness and robustness of the proposed framework.

Keywords: Optimization and Optimal Control, AI-Based Methods, Modelling, Simulation and Validation of Cyber-physical Energy Systems
Abstract: Vector Ordinal Optimization (VOO) is a promising approach for tackling multi-objective simulation optimization problems. However, existing theoretical analyses of VOO are confined to providing ordinal performance guarantees within the evaluated sample set. Consequently, it remains difficult to ascertain whether the good enough solutions identified within this local subset remain globally competitive across the entire decision space. To address this issue, we propose a novel VOO framework based on the Normalized Domination Count (VOO-NDC). The proposed method establishes a rigorous Global Ordinal Performance Guarantee by deriving an analytically computable theoretical lower bound for the Global Alignment Probability. Furthermore, by extending the VOO-NDC framework, we develop a universal method to evaluate the global ordinal performance of any arbitrarily given candidate set. Numerical experiments on standard ZDT benchmark functions and unit commitment instances in power systems demonstrate the effectiveness and practical value of the proposed methods.

Keywords: Renewable Energy Sources, Optimization and Optimal Control
Abstract: With the integration of large amounts of uncertain renewable energy into the grid, the energy system is evolving toward large-scale distributed new-energy systems. This transition gives rise to numerous stochastic dynamic decision-making problems in energy operation and management. Among these, maintaining load balance and ensuring the operational safety and stability of power grids are of critical importance.

This study focuses on energy management within microgrids, aiming to schedule the charging and discharging of energy storage devices to reduce power fluctuations and improve the availability and utilization of renewable generation. The problem is modeled as a mean-variance optimization problem in a finite-horizon discrete-time Markov decision process (MDP), optimizing a mean-variance trade-off of cumulative rewards.

However, this problem has long posed challenges, as the non-separability of variance renders dynamic programming ineffective. By introducing pseudo-mean and pseudo-variance, the original problem is reformulated as a bilevel optimization problem: the outer level is a single-parameter optimization for pseudo-mean, and the inner level is a pseudo-mean variance optimization, which can be linked to a standard MDP through state augmentation. To efficiently solve this bilevel problem, we propose a global optimization algorithm based on policy-dominated iterative interval elimination. Numerical experiments based on distribution network data validate the effectiveness of the proposed method. The research outcomes provide valuable solutions for the energy management challenges of networked microgrid systems in distribution networks.

Keywords: Agent-Based Systems, Intelligent Transportation Systems, AI-Based Methods
Abstract: Generative AI models, notably Large Language Models (LLMs), have excelled in open-ended content creation across text, images, and video, where evaluation is largely perceptual and tolerant to occasional imperfections. However, applying these models to engineering design remains challenging because engineering practice is governed by a fundamentally different standard: a design is acceptable only if it strictly satisfies code-based constraints. To bridge this gap, we propose OptiForm-Agent, a novel architecture that shifts the role of LLMs from direct artifact generators to rigorous mathematical formulators. Instead of directly generating design artifacts, OptiForm-Agent leverages LLMs to translate design codes, requirements, and preferences into mathematical constraints, delegating the final generation to numerical solvers to guarantee strict compliance. Moreover, direct translations of engineering semantics often yield computationally intractable formulations. To ensure solvability, we introduce an Adaptive Reformulation Mechanism that dynamically analyzes problem structure and automatically applies strategies such as linearization, variable substitution or decomposition to produce executable, solver-friendly models. We validate our method on highway route alignment, a highly constrained and representative engineering design domain. Experiments on real-world terrain scenarios demonstrate that OptiForm-Agent produces solver-optimized alignments with zero inspected hard-constraint violations in the evaluated scenarios, while direct LLM generation fails to reliably satisfy engineering requirements.

Keywords: Smart Grids, Distributed Generation and Storage, Modelling, Simulation and Validation of Cyber-physical Energy Systems
Abstract: Buildings are increasingly evolving from passive energy consumers into proactive energy hubs through the integration of on-site renewable generation, electrical storage systems (ESS), and flexible load management. This enables developing energy-efficient, low-carbon and cost-effective smart buildings. However, achieving the objective requires effective coordination of the heterogeneous energy resources in buildings to fully unlock the operational flexibility. This task is particularly challenging for commercial buildings where the HVAC system constitutes the dominant flexible load and shows substantial operational flexibility. The operation of the HVAC system is governed by multi-zone thermal dynamics that exhibit strong temporal and spatial couplings, making the joint optimization of HVAC operation and other flexible resources computationally challenging. To address this challenge, this paper proposes a surrogate model assisted DRL framework for smart building integrated energy management. The framework incorporates a newly developed low-dimensional and physically meaningful surrogate model of multi-zone HVAC control into the integrated energy management. The surrogate model provides a compact and informative state representation, significantly reduces the dimensionality of state and action space, and enables a low-dimensional state-transition model for efficient policy learning of DRL agents. Case studies based on real-world data demonstrate that the proposed DRL significantly improves convergence rate compared with conventional DRL training on raw high-dimensional state-space model. Moreover, the results show that the substantial operational flexibility of multi-zone HVAC system can be effectively exploited to reduce building electricity cost.

Keywords: AI-Based Methods, Machine learning
Abstract: Proactive Quality of Service (QoS) estimation is crucial for reliable, low-latency 6G vehicular networks. This paper proposes a data-driven framework for generating predictive QoS heatmaps, significantly reducing the over- head of continuous channel state monitoring. Leveraging an OMNeT++/Simu5G and Simulation of Urban Mobility co-simulation, a high-fidelity dataset of urban spatiotemporal network dynamics is generated. A Random Forest regressor is trained using only two-dimensional vehicular geographical coordinates to predict a spatially aggregated QoS score, explicitly integrating normalized Signal-to-Interference-plus-Noise Ratio, Throughput, Medium Access Control delay, and Packet Loss Rate. Experimental evaluation compares Random Forest against standard regression baselines, showing that the model provides a competitive trade-off between prediction accuracy, non-linear spatial modeling capability, and computational efficiency. Ultimately, this predictive topography serves as a foundational knowledge base for 6G Digital Twins, enabling Deep Reinforcement Learning based network-aware vehicular routing.

Keywords: AI-Based Methods, Intelligent and Flexible Manufacturing, Collaborative Robots in Manufacturing
Abstract: Multi-robot task allocation and scheduling (MRTAS) in aerospace manufacturing is highly challenging due to the coupled physical, logical, and resource constraints inherent in complex sequential operations. Traditional exact algorithms face scalability limitations, whereas manual constraint programming (CP) modeling requires extensive domain expertise and lengthy development cycles, hindering their adaptation to dynamic production requirements. To address these challenges, a retrieval-augmented, self-evolving framework based on in-context learning and streaming large language models (LLMs), referred to as MRTAS-ICL-StreamLLM, is developed to autonomously generate high-quality CP models for MRTAS problems. The proposed framework comprises three synergistic components: a domain-adaptive retrieval mechanism that dynamically selects relevant demonstration examples via hybrid similarity measures combining lexical and structural features; an iterative constraint refinement module that employs temperature annealing to progressively generate domain-specific optimization rules while balancing pattern exploitation and novel exploration; and a code execution feedback loop that validates the generated models against actual solver performance. This closed-loop architecture transforms LLMs into robust modeling assistants capable of continuous learning from execution outcomes. Experimental validation in aircraft skin manufacturing scenarios demonstrates that MRTAS-ICL-StreamLLM generates accurate and executable CP models while adaptively improving solution quality through iterative refinement, thereby offering an effective approach to automated and adaptive CP modeling for complex industrial scheduling environments.

Keywords: Intelligent and Flexible Manufacturing, Computer Vision in Automation, Zero-Defect Manufacturing
Abstract: In high-mix low-volume manufacturing, quality inspection faces the double open-set risk: the simultaneous emergence of unknown textures and unseen defects. Conventional raw feature modeling fails to generalize across these specification shifts without extensive retraining. To address this, this paper proposes a novel framework: residual-based unified anomaly detection (RUAD). Unlike traditional raw feature modeling, RUAD shifts the paradigm to relative residual modeling. By retrieving prototype and subtracting texture-specific elements, RUAD explicitly decouples anomaly semantics from the texture background. This operation forces diverse normal textures to collapse into a unified, zero-centered residual distribution, while anomalies remain topologically distinct. Subsequently, anomalies are identified based on the low-density hypothesis. Extensive experiments demonstrate that RUAD achieves superior performance, offering a robust and unified solution for inspecting various textured products in high-mix low-volume automation.

Keywords: Compliant Assembly, Intelligent and Flexible Manufacturing, Swarms
Abstract: High amplitude/phase consistency across parallel radio frequency channels is essential for reliable coherent synthesis in array-based electronic system, yet unavoidable component dispersions often introduce significant inter-channel mismatches after assembly. This paper investigates the multi-channel assembly consistency optimization (MCACO) problem, where pre-tested heterogeneous components are assigned to multiple channels to improve consistency over multiple verification points. The optimization objective of amplitude/phase consistency is reflected by introducing a virtual reference baseline and a hierarchical objective function. And to efficiently solve MCACO problem, a knowledge-guided artificial bee colony (KGABC) algorithm that embeds domain-guided strategies into the employed, onlooker, and scout phases is proposed. Specifically, it combines controlled random perturbations, a repair-oriented neighborhood search driven by verification-point-centered complementary pairing, and a partial restart mechanism for stagnation recovery. Numerical experiments on 12 instances under a fixed runtime budget show that KGABC consistently outperforms the baselines, and the ablation study further confirms the effectiveness of the proposed knowledge-guided strategy. Moreover, the advantage becomes more pronounced as the instance scale increases, where KGABC reduces the number of unqualified channels by over 80% on medium-to-large instances.

Keywords: Intelligent and Flexible Manufacturing, Manufacturing, Maintenance and Supply Chains, AI-Based Methods
Abstract: Flexible job shop scheduling has been widely investigated in manufacturing systems. However, most existing studies implicitly assume that all required materials are available when operations become ready for processing. In practical production environments, materials are often delivered in dispersed batches, which may interrupt operation execution and lead to infeasible schedules. In this paper, flexible job shop scheduling with material kitting constraints is investigated, where operation feasibility is jointly determined by machine availability, technological precedence, and material availability over time. A material-aware scheduling framework is developed, in which a decoding mechanism explicitly incorporates material arrival information and dynamically adjusts operation start times. An insertion mechanism is further introduced to utilize material waiting windows and reduce idle machine time. To enhance solution quality under strict feasibility requirements, a cooperative search strategy is employed, in which feasibility-oriented scheduling is supported by auxiliary exploration under relaxed material constraints. Numerical experiments demonstrate that the proposed MASF consistently achieves superior or competitive makespan performance compared with baseline scheduling algorithms.

Keywords: Planning, Scheduling and Coordination, Intelligent and Flexible Manufacturing
Abstract: Distributed job-shop scheduling in continuous production is increasingly challenged by hybrid orders, where dynamically arriving local orders and cross-factory collaborative orders compete for shared resources while following different decision workflows. However, most existing studies assume either purely local orders or fully transferable orders, limiting their applicability in realistic distributed manufacturing environments. To address this gap, this paper formulates a dynamic distributed job-shop scheduling problem with hybrid orders under stochastic arrivals, where each order is modeled as a multi-operation job with precedence constraints. To handle the heterogeneous decision structure of hybrid-order scheduling, we propose a hierarchical multi-agent reinforcement learning method that coordinates enterprise-level factory assignment for global jobs, factory-level machine assignment for hybrid jobs, and machine-level job sequencing. Based on a discrete-event simulation model and PPO training, computational experiments are conducted against composite rule-based scheduling strategies and two non-hierarchical multi-agent methods. The results show that the proposed method generally outperforms the other methods. Moreover, the possible reasons for its superiority over the non-hierarchical multi-agent methods are further analyzed.