Abstract
As collaborative robots increasingly redefine industrial automation, understanding the factors that drive their adoption is essential to operations management. This study examines the main drivers of collaborative robot adoption in the Brazilian manufacturing sector by combining theory-driven framing with a machine learning classification approach. It was developed a Random Forest classifier to identify the strongest predictors of cobot adoption and to rank their relative importance. Data were collected from a sample of respondents-primarily managers and chief executive officers-representing 300 industrial companies. Grounded in the Technology-Organization-Environment (TOE) framework and complemented by Diffusion of Innovations (DoI) and Institutional (INT) perspectives, the analysis shows that technological advantages, namely space efficiency, cost reduction, and ease of integration, are critical drivers of adoption. Organizational factors, including proactive managerial involvement and alignment with an innovation-oriented culture, significantly increase the likelihood of collaborative robot uptake. The model demonstrated robust predictive performance and produced interpretable variable importance scores that confirm the relative influence of technological and managerial factors. These findings provide a structured lens for understanding and guiding managerial decision-making on cobot adoption and translate into practical recommendations for managers.

Abstract
Ensuring the security of Industrial Control Systems (ICS) is increasingly critical due to increasing connectivity and cyber threats. Traditional security measures often fail to detect evolving attacks, necessitating more effective solutions. This paper evaluates machine learning (ML) methods for ICS cybersecurity, using the ICS-Flow dataset and Optuna for hyperparameter tuning. The selected models, namely Random Forest (RF), AdaBoost, XGBoost, Deep Neural Networks, Artificial Neural Networks, ExtraTrees (ET), and Logistic Regression, are assessed using macro-averaged F1-score to handle class imbalance. Experimental results demonstrate that ensemble-based methods (RF, XGBoost, and ET) offer the highest overall detection performance, particularly in identifying commonly occurring attack types. However, minority classes, such as IP-Scan, remain difficult to detect accurately, indicating that hyperparameter tuning alone is insufficient to fully deal with imbalanced ICS data. These findings highlight the importance of complementary measures, such as focused feature selection, to enhance classification capabilities and protect industrial networks against a wider array of threats.

Abstract
Glaucoma is one of the leading causes of irreversible blindness worldwide, affecting millions of people, often silently and progressively. Early diagnosis is crucial to slow its progression, but it remains challenging due to the need for manual analysis of large volumes of retinal images by trained specialists. In this context, automatic detection systems based on deep learning offer a promising opportunity to facilitate and accelerate the diagnostic process, providing scalability and high accuracy. This work presents the development of an automatic method for optic disc and optic cup segmentation in retinal fundus photographs, aiming to support early glaucoma detection. The proposed methodology is based on convolutional neural networks (CNNs), specifically an enhanced U-Net architecture with a ResNet50 backbone, incorporating attention mechanisms and data augmentation strategies to improve segmentation accuracy. The model was trained and validated using the REFUGE dataset, which contains high-quality fundus images with manual annotations of the disc and cup regions. Experimental results demonstrate that the developed model achieved an average Dice coefficient of 0.937 for optic disc segmentation and 0.828 for optic cup segmentation. Analysis of the cup-to-disc ratio (CDR) yielded mean values of VCDR = 0.497 ± 0.059, ACDR = 0.252 ± 0.060, and mean CDR = 0.375 ± 0.058, with 55.0% of cases classified as low risk, 43.3% as moderate risk, and 1.7% as high risk for glaucoma. These results highlight the potential of the proposed method as an assistive tool for automated glaucoma screening. © 2025 The Authors. Published by Elsevier B.V.

Abstract
Product design is a complex and iterative process that requires the balance of multiple constraints, such as material selection, manufacturability, regulatory compliance, and structural integrity, among others. Traditional design workflows follow a human-driven approach, limiting efficiency, adaptability, and the ability to quickly respond to evolving limitations. This paper introduces an agentic approach to product design, leveraging multi-agent systems to distribute and automate design tasks dynamically. To demonstrate this methodology, a hypothetical enclosure design is used as a guiding example, demonstrating how agents interact to generate product specifications, select materials, validate structural properties, assess manufacturability, and perform other relevant tasks throughout the design process. To implement this framework, CrewAI is utilized as an agent coordination system that enables the structured definition of roles and execution of tasks for autonomous agents. In the final section, a case study is presented, focusing on the design of a parallelepiped enclosure, applying the proposed framework in a simulated environment. Our findings highlight the advantages of agent-based collaboration in product design, showcasing its potential to optimize workflows, reduce development time, and improve adaptability to changing requirements.

Abstract
Road safety has become an increasingly important concern and the integration of Advanced Rider Assistance Systems and Advanced Driver Assistance Systems plays a crucial role in preventing accidents. This work proposes a computer vision pipeline to automatically detect hazardous road anomalies—loose gravel, potholes, and puddles—from a motorcycle-mounted camera, targeting real-time operation on embedded edge devices. A hybrid dataset of 28764 annotated images was created by combining real-world photos, Blender-rendered synthetic scenes, and AI-generated images to improve diversity and coverage. Multiple state-of-the-art object detectors were trained and benchmarked, including the YOLOv5/7/11/12 families and the transformer-based RT-DETR architecture. While the RT-DETR model achieved the highest precision overall, its computational complexity and heavy resource requirements limited its suitability for real-time deployment on low-cost embedded platforms. Conversely, the YOLOv11n model demonstrated the best accuracy–efficiency trade-off, reaching mAP@0.5 = 0.872 at 320$\times$320 with 0.045 s/frame on a Jetson Nano, while lighter variants remained viable on Raspberry Pi boards. Across classes, gravel was the most reliably detected, and operating points around a confidence threshold of $\tau \approx 0.31$ yielded balanced F1 scores up to 0.82. Although results show that automatic road-condition monitoring on affordable hardware is feasible, the prototype has not yet undergone on-road field trials. It does not include an integrated rider alert module or energy-use assessment. These gaps define the immediate roadmap for deployment.

Abstract
Eye diseases, such as glaucoma, diabetic retinopathy, and age-related macular degeneration, drive the growing need for reliable and scalable analyses of fundus and optical coherence tomography (OCT) images. Deep learning performs strongly in ocular structure segmentation. However, it typically relies on dense pixel-wise annotations, which are costly and difficult to obtain at scale. Weakly supervised learning (WSL) can reduce this burden by leveraging coarse labels, limited strong annotations, and unlabeled data. This systematic umbrella review synthesizes survey and review articles on weakly supervised deep learning for image segmentation, with a focus on ocular imaging (fundus and OCT/OCTA). After analyzing twenty-one secondary studies, the main finding reveals an empty intersection: WSL-focused segmentation surveys are often modality-agnostic. Conversely, ocular reviews are predominantly fully supervised and seldom offer quantitative evidence on annotation-effort savings or direct comparisons between weak and fully supervised methods on identical datasets. Across the included reviews, label-efficient strategies cluster around CAM/MIL formulations, sparse supervision (points/scribbles/boxes), pseudo-labelling/self-training, and semi-/self-supervised learning, implemented mainly with U-Net/DeepLab families and increasingly Transformer or hybrid backbones. These results provide a structured map of available WSL mechanisms and, critically, identify reproducible reporting gaps that currently prevent fair benchmarking in ocular segmentation. Therefore, this review supports the development of ocular-specific benchmarks and minimum reporting practices that link segmentation performance to annotation effort.