The accumulation of litter in natural environments poses significant ecological and social challenges, motivating the development of automated solutions for litter detection. However, collecting and centrally aggregating large-scale annotated datasets for training object detectors often raises privacy and ownership concerns. In this work, we propose a Federated Learning (FL) framework to train a lightweight litter detection model based on the YOLO architecture, which enables collaborative model development without requiring centralized access to raw data. Each participating client locally trains the model on site-specific datasets collected in the wild, and only model updates are shared with a central server for aggregation. We compare and contrast different FL process configurations involving mixed and heterogeneous training datasets built starting from two commonly used benchmark datasets collected across different locations and having very different visual data distributions, i.e. TACO and PlastOPol. Experimental results show that the federated model, trained across these non-IID data distributions, achieves superior generalization in cross-dataset evaluation compared to the corresponding centrally trained models.

In healthcare systems, practitioners are responsible for making decisions when a patient’s health, or even life, are at stake. Real-time data-driven modeling, analysis, and prediction approaches, such as the Digital Shadow (DS) paradigm, inform and support decision-makers in such critical situations. We introduce GENGAR, a DS-based methodology to identify critical scenarios in a patient-device-physician (PDP) triad with human agents and cyber-physical devices interacting under uncertainty. The proposed solution relies on automata-based modeling and formal analysis techniques to predict and inform the practitioner of critical contingencies that may compromise patient safety, enhancing the system’s dependability. In particular, it leverages automata learning to infer and evolve a realistic patient model from clinical logs. GENGAR then exploits mutational and searchbased fuzzing to generate scenarios and detect failure cases, i.e., those violating predefined dependability requirements. Failure scenarios are then filtered using qualitative criteria on clinical plausibility, yielding up to $60 \%$ realistic cases.

The paper addresses the problem of estimating the energy consumed by production resources in manufacturing so that alternative process designs can be compared in terms of energy expenditure. In particular, the proposed methodology focuses on Computer Numerical Controlled (CNC) machining centers. Classical approaches to energy modeling require high expertise and large development effort since, for example, data acquisition is resource-specific and must be repeated frequently to avoid obsolescence. An automated and flexible data-driven methodology is designed in this work. A data-driven method is employed to learn a hybrid and stochastic model of a CNC machining center’s energetic behavior. The learned model is used to provide offline energy consumption estimates of simulated part-programs before the actual execution of the cutting. Numerical results show the performance of the proposed method on a set of case studies. The methodology is also applied to a real industrial application, including data collected during machine production.Note to Practitioners—This article provides a flexible and autonomous data-driven approach to building models representing the energetic behavior of production resources, particularly CNC machining centers. The learned models can predict machine energy consumption while executing complex part-programs. The algorithm uses data that are commonly acquired by contemporary machine monitoring systems and does not require ad-hoc experimental tests for training. Specifically, it requires the spindle rotary speed signal, part load/unload signal, and spindle (or machine) power signal during the learning phase, whilst the estimation phase uses only the load/unload and spindle speed simulated signals.

Language is a deep-rooted means of perpetration of stereotypes and discrimination. Large Language Models (LLMs), now a pervasive technology in our everyday lives, can cause extensive harm when prone to generating toxic responses. The standard way to address this issue is to align the LLM, which, however, dampens the issue without constituting a definitive solution. Therefore, testing LLM even after alignment efforts remains crucial for detecting any residual deviations with respect to ethical standards. We present EvoTox, an automated testing framework for LLMs’ inclination to toxicity, providing a way to quantitatively assess how much LLMs can be pushed towards toxic responses even in the presence of alignment. The framework adopts an iterative evolution strategy that exploits the interplay between two LLMs, the System Under Test (SUT) and the Prompt Generator steering SUT responses toward higher toxicity. The toxicity level is assessed by an automated oracle based on an existing toxicity classifier. We conduct a quantitative and qualitative empirical evaluation using five state-of-the-art LLMs as evaluation subjects having increasing complexity (7–671B parameters). Our quantitative evaluation assesses the cost-effectiveness of four alternative versions of EvoTox against existing baseline methods, based on random search, curated datasets of toxic prompts, and adversarial attacks. Our qualitative assessment engages human evaluators to rate the fluency of the generated prompts and the perceived toxicity of the responses collected during the testing sessions. Results indicate that the effectiveness, in terms of detected toxicity level, is significantly higher than the selected baseline methods (effect size up to 1.0 against random search and up to 0.99 against adversarial attacks). Furthermore, EvoTox yields a limited cost overhead (from 22% to 35% on average).This work includes examples of toxic degeneration by LLMs, which may be considered profane or offensive to some readers. Reader discretion is advised.

Abstract Developing robotic applications with human–robot interaction for the service sector raises a plethora of challenges. In these settings, human behavior is essentially unconstrained as they can stray from the plan in numerous ways, constituting a critical source of uncertainty for the outcome of the robotic mission. Application designers require accessible and reliable frameworks to address this issue at an early development stage. We present a model-driven framework for developing interactive service robotic scenarios, allowing designers to model the interactive scenario, estimate its outcome, deploy the application, and smoothly reconfigure it. This article extends the framework compared to previous works by introducing an analysis of the impact of human errors on the mission’s outcome. The core of the framework is a formal model of the agents at play—the humans and the robots—and the robotic mission under analysis, which is subject to statistical model checking to estimate the mission’s outcome. The formal model incorporates a formalization of different human erroneous behaviors’ phenotypes, whose likelihood can be tuned while configuring the scenario. Through scenarios inspired by the healthcare setting, the evaluation highlights how different configurations of erroneous behavior impact the verification results and guide the designer toward the mission design that best suits their needs.

Robots are currently mostly found in industrial settings. In the future, a wider range of environments will benefit from their inclusion. This calls for the development of tools that allow professionals to set up dependable robotic applications in which people productively interact with robots aware of their needs. Given the co-existence of humans and robots, the precise analysis-e.g., through formal verification techniques-of properties related to aspects such as human needs and physiology is of paramount importance. In this paper, we present a formally-based, model-driven approach to design and verify scenarios involving human-robot interactions. Some of the features of our approach are tailored to the healthcare domain, from which our case studies are derived. In our approach, the designer specifies the main parameters of the mission to generate the model of the application, which includes mobile robots, the humans to be served, including some of their physiological features, and the decision-maker that orchestrates the execution. All components are modeled through hybrid automata to capture variables with complex dynamics. The model is verified through Statistical Model Checking (SMC), using the Uppaal tool, to determine the probability of success of the mission. The results are examined by the developer, who iteratively refines the design until the probability of success is satisfactory.