Aldo Franco Dragoni

Reviewing documents against natural-language requirements is hard when those requirements demand interpretation, contextual evidence, and non-trivial judgment. It is even harder when the review is to be automated reliably and made auditable. In domains like education, where teaching materials must align with pedagogical models and evolving expectations, including the rise of Large Language Models (LLMs), such decisions cannot be reduced to mechanical checks. When an LLM assigns a verdict directly, interpretation, evidence selection, and judgment collapse into a single opaque step, making errors hard to localize and outcomes hard to audit. AiDER separates interpretation from judgment under a deliberate 2B–4B parameter constraint. Each natural-language requirement is compiled into an executable Common Expression Language (CEL) rule, and static analysis of that rule produces a bounded-evidence scaffold that constrains what the model may extract from the document. The verdict is then computed deterministically by applying the rule to the extracted evidence (no LLM call at this step), yielding an inspectable trace from requirement to rule, rule to evidence, and evidence to verdict. We evaluate AiDER on four small models (Gemma 4 and Qwen3.5, at 2B and 4B) under three compilation strategies: human-written reference rules, direct compilation, and validation-loop repair. The evaluation uses 200 requirement-document pairs from a higher-education course pack. Across three performance dimensions (syntactic validity, semantic fidelity, and document-level agreement), direct compilation often yields valid but semantically divergent rules, whereas validation-driven repair substantially improves both validity and fidelity across all four models. An ablation shows that the static scaffold alone encodes most of the extraction intent: LLM-generated extraction contracts add little when the rule is well-specified, but matter more when it is ambiguous.

Interest in the adoption of autonomous vehicles (AVs) continues to grow. It is essential to design new software architectures that meet stringent real-time, safety, and scalability requirements while integrating heterogeneous hardware and software solutions from different vendors and developers. This paper presents a lightweight, modular, and scalable architecture grounded in Service-Oriented Architecture (SOA) principles and implemented in ROS 2 (Robot Operating System 2). The proposed design leverages ROS 2’s Data Distribution System-based Quality-of-Service model to provide reliable communication, structured lifecycle management, and fault containment across distributed compute nodes. The architecture is organized into Perception, Planning, and Control layers with decoupled sensor access paths to satisfy heterogeneous frequency and hardware constraints. The decision-making core follows an event-driven policy that prioritizes fresh updates without enforcing global synchronization, applying zero-order hold where inputs are not refreshed. The architecture was validated on a 1:10-scale autonomous vehicle operating on a city-like track. The test environment covered canonical urban scenarios (lane-keeping, obstacle avoidance, traffic-sign recognition, intersections, overtaking, parking, and pedestrian interaction), with absolute positioning provided by an indoor GPS (Global Positioning System) localization setup. This work shows that the end-to-end Perception–Planning pipeline consistently met worst-case deadlines, yielding deterministic behaviour even under stress. The proposed architecture can be deemed compliant with real-time application standards for our use case on the 1:10 test vehicle, providing a robust foundation for deployment and further refinement.

Many “Industry 4.0” applications rely on data-driven methodologies such as Machine Learning and Deep Learning to enable automatic tasks and implement smart factories. Among these applications, the automatic quality control of manufacturing materials is of utmost importance to achieve precision and standardization in production. In this regard, most of the related literature focused on combining Deep Learning with Nondestructive Testing techniques, such as Infrared Thermography, requiring dedicated settings to detect and classify defects in composite materials. Instead, the research described in this paper aims at understanding whether deep neural networks and transfer learning can be applied to plain images to classify surface defects in carbon look components made with Carbon Fiber Reinforced Polymers used in the automotive sector. To this end, we collected a database of images from a real case study, with 400 images to test binary classification (defect vs. no defect) and 1500 for the multiclass classification (components with no defect vs. recoverable vs. non-recoverable). We developed and tested ten deep neural networks as classifiers, comparing ten different pre-trained CNNs as feature extractors. Specifically, we evaluated VGG16, VGG19, ResNet50 version 2, ResNet101 version 2, ResNet152 version 2, Inception version 3, MobileNet version 2, NASNetMobile, DenseNet121, and Xception, all pre-trainined with ImageNet, combined with fully connected layers to act as classifiers. The best classifier, i.e., the network based on DenseNet121, achieved a 97% accuracy in classifying components with no defects, recoverable components, and non-recoverable components, demonstrating the viability of the proposed methodology to classify surface defects from images taken with a smartphone in varying conditions, without the need for dedicated settings. The collected images and the source code of the experiments are available in two public, open-access repositories, making the presented research fully reproducible.

Leukemia is a blood cancer characterized by leukocyte overproduction. Clinically, the reference for acute lymphoblastic leukemia diagnosis is a blood biopsy that allows obtain microscopic images of leukocytes, whose early-stage classification into leukemic (LEU) and healthy (HEA) may be disease predictor. Thus, the aim of this study is to propose an interpretable artificial neural network (ANN) for leukocyte classification to timely diagnose acute lymphoblastic leukemia. The “ALL_IDB2” dataset was used. It contains 260 microscopic images showing leukocytes acquired from 130 LEU and 130 HEA subjects. Each microscopic image shows a single leukocyte that was characterized by 8 morphological and 4 statistical features. An ANN was developed to distinguish microscopic images acquired from LEU and HEA subjects, considering 12 features as inputs and the local-interpretable model-agnostic explanatory (LIME) algorithm as an interpretable post-processing algorithm. The ANN was evaluated by the leave-one-out cross-validation procedure. The performance of our ANN is promising, presenting a testing area under the curve of the receiver operating characteristic equal to 87%. Being implemented using standard features and having LIME as a post-processing algorithm, it is clinically interpretable. Therefore, our ANN seems to be a reliable instrument for leukocyte classification to timely diagnose acute lymphoblastic leukemia, guaranteeing a high clinical interpretability level.

Although face recognition technology is currently integrated into industrial applications, it has open challenges, such as verification and identification from arbitrary poses. Specifically, there is a lack of research about face recognition in surveillance videos using, as reference images, mugshots taken from multiple Points of View (POVs) in addition to the frontal picture and the right profile traditionally collected by national police forces. To start filling this gap and tackling the scarcity of databases devoted to the study of this problem, we present the Face Recognition from Mugshots Database (FRMDB). It includes 28 mugshots and 5 surveillance videos taken from different angles for 39 distinct subjects. The FRMDB is intended to analyze the impact of using mugshots taken from multiple points of view on face recognition on the frames of the surveillance videos. To validate the FRMDB and provide a first benchmark on it, we ran accuracy tests using two CNNs, namely VGG16 and ResNet50, pre-trained on the VGGFace and VGGFace2 datasets for the extraction of face image features. We compared the results to those obtained from a dataset from the related literature, the Surveillance Cameras Face Database (SCFace). In addition to showing the features of the proposed database, the results highlight that the subset of mugshots composed of the frontal picture and the right profile scores the lowest accuracy result among those tested. Therefore, additional research is suggested to understand the ideal number of mugshots for face recognition on frames from surveillance videos.

The developments in the internet of things (IoT), artificial intelligence (AI), and cyber-physical systems (CPS) are paving the way to the implementation of smart factories in what is commonly recognized as the fourth industrial revolution. In the manufacturing sector, these technological advancements are making Industry 4.0 a reality, with data-driven methodologies based on machine learning (ML) that are capable of extracting knowledge from the data collected by sensors placed on production machines. This is particularly relevant in plastic injection molding, with the objective of monitoring the quality of molded products from the parameters of the production process. In this regard, the main contribution of this paper is the systematic comparison of ML techniques to predict the quality classes of plastic molded products, using real data collected during the production process. Specifically, we compare six different classifiers on the data coming from the production of plastic road lenses. To run the comparison, we collected a dataset composed of the process parameters of 1451 road lenses. On such samples, we tested a multi-class classification, providing a statistical analysis of the results as well as of the importance of the input features. Among the tested classifiers, the ensembles of decision trees, i.e., random forest and gradient-boosted trees (GBT), achieved 95% accuracy in predicting the quality classes of molded products, showing the viability of the use of ML-based techniques for this purpose. The collected dataset and the source code of the experiments are available in a public, open-access repository, making the presented research fully reproducible.

Following the growing availability of video surveillance cameras and the need for techniques to automatically identify events in video footages, there is an increasing interest towards automatic violence detection in videos. Deep learning-based architectures, such as 3D Convolutional Neural Networks, demonstrated their capability of extracting spatio-temporal features from videos, being effective in violence detection. However, friendly behaviours or fast moves such as hugs, small hits, claps, high fives, etc., can still cause false positives, interpreting a harmless action as violent. To this end, we present three deep-learning based models for violence detection and test them on the AIRTLab dataset, a novel dataset designed to check the robustness of algorithms against false positives. The objective is twofold: on one hand, we compute accuracy metrics on the three proposed models (two are based on transfer learning and one is trained from scratch), building a baseline of metrics for the AIRTLab dataset; on the other hand, we validate the capability of the proposed dataset of challenging the robustness to false positives. The results of the proposed models are in line with the scientific literature, in terms of accuracy, with transfer learning-based networks exhibiting better generalization capabilities than the trained from scratch network. Moreover, the tests highlighted that most of the classification errors concern the identification of non-violent clips, validating the design of the proposed dataset. Finally, to demonstrate the significance of the proposed models, the paper presents a comparison with the related literature, as well as with models based on well-established pre-trained 2D Convolutional Neural Networks 2D CNNs. Such comparison highlights that 3D models get better accuracy performance than time distributed 2D CNNs (merged with a recurrent model) in processing the spatio-temporal features of video clips. The source code of the experiments and the AIRTLab dataset are available in public repositories.