The convergence of Artificial Intelligence (AI) and the Industrial Internet of Things (IIoT) is reshaping Cyber-Physical Systems (CPSs), enabling intelligent automation, real-time decision-making, and adaptive control across diverse industrial domains. A key enabler of this transformation is On-Device AI, where training and inference occur directly on edge devices. While deploying AI models in constrained environments presents challenges-such as limited computational resources and hardware heterogeneity-the benefits of reduced latency, improved energy efficiency, and enhanced data privacy make this approach essential for next-generation CPSs. However, scaling and managing AI-enabled CPSs introduces new complexities, including efficient coordination among sensing, computation, and actuation, as well as the need for dynamic model adaptation in resource-constrained settings. Addressing these challenges requires architectural solutions that support distributed intelligence while maintaining system responsiveness and robustness. This paper investigates the use of Digital Twins (DTs) as a cyber-physical abstraction layer that enhances the deployment and management of On-Device AI. By maintaining synchronized, high-level digital representations of physical assets, DTs facilitate local AI execution, optimize resource allocation, and support low-latency decision-making. We validate our approach through experimental evaluation in a microfactory testbed, demonstrating how DTs improve lifecycle management, operational efficiency, and system adaptability in constrained environments. The results highlight the potential of DTs as a foundational technology for scalable, secure, and efficient AI-driven CPSs, offering valuable insights into the deployment of intelligent systems in heterogeneous industrial ecosystems.

Digital Twins (DTs) have emerged as a powerful paradigm for modeling, analyzing, and optimizing cyber-physical systems across a wide range of domains. Despite their growing adoption, the application of DTs to real-time systems—characterized by stringent timing constraints—remains a significant and underexplored area of research. Gaining a deeper understanding of the interaction between DTs and real-time systems is essential for extending the applicability of DT technologies to time-critical environments. This paper aims to establish a foundation for systematically addressing the challenges inherent in developing DTs for real-time systems. Specifically, we identify the key characteristics that a real-time system must possess to enable accurate and reliable DT replication. To illustrate these concepts, we present a case study in which a DT is employed to monitor and interact with a simulated robotic arm controlled by a real-time system. Our experimental findings highlight the strategic importance of this research direction and demonstrate the advantages of implementing a structured synchronization mechanism between the DT and its physical counterpart.

This paper makes a compelling case for the adoption of the recently proposed Data Mesh architecture within IoT-Edge-Cloud Continuum scenarios, particularly in the context of Intelligent Transportation Systems and Data-driven Mobility Services. Unlike centralized cloud-based approaches, based on data warehouses/lakes connected with ETL (Extract, Transform, and Load) pipelines, Data Mesh promotes a decentralized data ownership model which brings several advantages in addressing open challenges in IoT-Edge-Cloud Continuum scenarios. First, we present an overview of the Data Mesh concepts, and how they advance the state of the art in data management architectures. Secondly, we discuss how their adoption might ease the development of IoT -Edge-Cloud applications in terms of: (i) hiding the heterogeneity of the IoT Layer, (ii) mitigating latency by enabling full domain migrations, and (iii) promoting the adoption of AI techniques, such as MLOps and Federated Learning at the edge of the net-work. Finally, we provide practical guidelines for implementing such an architecture to enhance the safety of pedestrians and vulnerable users, based on our experience with the Modena Automotive Smart Area.

R4.2 In the Industry 4.0 era, Digital Twins (DTs) serve as virtual representations of physical objects and intermediaries between the physical world and the digital realm. DTs require proper modeling, design, and development to ensure their seamless integration along the cloud-to-edge continuum. In particular, this work introduces a microservices-based and serverless-ready model for DTs, laying the foundation for cost-effective DT deployment and orchestration. The joint adoption of microservices and serverless computing offers significant potential to address various challenges, including accommodating variable application requirements, managing load imbalances, and mitigating network faults. The proposed DT model has been implemented in different flavors: two serverless implementations—one that relies on a serverless framework of a cloud provider and one running at the edge on-premises—and a microservices one. These implementations have been experimentally evaluated with particular emphasis on the quality of cyber–physical entanglement. This work not only discusses the advantages and drawbacks of different implementations from a qualitative perspective but also quantitatively evaluates them with the in-the-field collection of experimental performance results. Notably, we report that a serverless implementation typically performs an order of magnitude worse than a microservices one in terms of entanglement, i.e., hundreds vs. tens of milliseconds.

The development of the Digital Twin (DT) approach is tilting research from initial approaches that aim at promoting early adoption to sophisticated attempts to develop, deploy, and maintain applications based on DTs. In this context, we propose a highly dynamic and distributed ecosystem where containerized DTs co-evolve with an orchestration middleware. DTs provide digitalized representations of the targeted physical systems, while the orchestration middleware monitors and re-configures the deployed DTs in light of application constraints, available resources, and the quality of cyber-physical entanglement. First, we lay out the reference scenario. Then, we discuss the limitations of current approaches and identify a set of requirements that shape both DTs and the orchestration middleware. Subsequently, we describe a blueprint architecture that meets those requirements. Finally, we report empirical evidence on both the feasibility and the effectiveness of a proof-of-concept implementation of the proposed ecosystem.

Digital Twins (DTs) have recently emerged as a valuable approach for modeling, monitoring, and controlling physical objects in Industrial Internet of Things applications. Measuring the quality of entanglement between the digital and physical counterparts plays a crucial role in the adoption of DTs. In this paper, we propose a concise yet expressive metric for representing the quality of entanglement, namely Overall Digital Twin Entanglement (ODTE), based on two key factors: timeliness and completeness. Furthermore, the paper presents the development of our industrial testbed implemented on top of Kubernetes, where we show practical applications of the proposed ODTE metric by highlighting and discussing its benefits in realistic use cases.

Digital factories represent an abstraction of real factories, which is useful to manage at a high level the processes as well as the interactions inside the factories but also the interactions between factories. This abstraction can automatize several processes and can enable to dynamically adapt the factory production to unexpected situations. Software agents can meet the requirements of digital factories by means of their features of autonomy, reactivity, proactivity and sociality. In this paper, we survey the use of software agents in the context of digital factories, showing how they can be exploited. A discussion about the advantages brought by software agents and the limitation of agent-based approaches completes the paper.

Research on social and mobile technologies recently provided tools to collect and mine massive amounts of mobility data. Ride sharing is one of the most prominent applications in this area. While a number of research and commercial initiatives already proposed solutions for long-distance journeys, the opportunities provided by modern pervasive systems can be used to promote local, daily ride sharing within the city. We present a set of algorithms to analyze urban mobility traces and to recognize matching rides along similar routes. These rides are amenable for ride sharing recommendations. We validate the proposed methodology using data provided by a large Italian telecom operator. Assuming the full set of considered users are willing to accept 1-km detours, experimental results on two large cities show that more than 60% of trips could be saved. These results can be used to evaluate the potential of a ride sharing system before its actual deployment and to actually support an opportunistic ride sharing recommender system.

Despite prejudice cannot be directly observed, nonverbal behaviours provide profound hints on people inclinations. In this paper, we use recent sensing technologies and machine learning techniques to automatically infer the results of psychological questionnaires frequently used to assess implicit prejudice. In particular, we recorded 32 students discussing with both white and black collaborators. Then, we identified a set of features allowing automatic extraction and measured their degree of correlation with psychological scores. Results confirmed that automated analysis of nonverbal behaviour is actually possible thus paving the way for innovative clinical tools and eventually more secure societies.

Urban environments are increasingly pervaded by ICT devices. Soon, citizens and technologies could collaboratively constitute large-scale socio-technical organisms supporting both individual and collective awareness. This paper illustrates a modern awareness framework rooted in this scenario. The framework has been designed to collect and classify data streams in a modular way. It supports service oriented, reconfigurable components and provides a solid background to put at joint work specification- and data-driven approaches. Furthermore, we use the framework to experimentally evaluate an innovative meta-classification scheme based on state-automata for (i) improving energy efficiency, (ii) improving classification accuracy and (iii) improving software engineering of aware systems.

Pervasive services may have to rely on multimodal classification to implement situation-recognition. However, the effectiveness of current multimodal classifiers is often not satisfactory. In this paper, we describe a novel approach to multimodal classification based on integrating a vision sensor with a commonsense knowledge base. Specifically, our approach is based on extracting the individual objects perceived by a camera and classifying them individually with non-parametric algorithms; then, using a commonsense knowledge base, classifying the overall scene with high effectiveness. Such classification results can then be fused together with other sensors, again on a commonsense basis, for both improving classification accuracy and dealing with missing labels. Experimental results are presented to assess, under different configurations, the effectiveness of our vision sensor and its integration with other kinds of sensors, proving that the approach is effective and able to correctly recognize a number of situations in open-ended environments.

The future large-scale deployment of pervasive sensor network infrastructures calls for mechanisms enabling the extraction of general-purpose data at limited energy costs. The approach presented in this article relies on a simple algorithm to let a sensor network self-organize a virtual partitioning in correspondence to spatial regions characterized by similar sensing patterns, and to let distributed aggregation of sensorial data take place on a per-region basis. The result of this process is that a sensor network can be modeled as a collection of virtual macro sensors, each associated to a well-characterized region of the physical environment. Within each region, each physical sensor has the local availability of aggregated data about its region and is able to act as an access point to such data. This feature promises to be very suitable for a number of emerging usage scenarios. Our approach is described and evaluated in both a simulation environment and a real test bed, and quantitatively compared with related works in the area. Current limitations and areas of future development are also discussed.

A novel capability for modelling and simulating intra- and inter-organizational collaboration in an emergency-response domain is presented. This capability combines the prescriptive, top-down view of organizations, which describes how they work “on paper,” and the descriptive, bottom-up view, which describes how they actually work, by focusing on three components in the light of agent-based modelling and simulation tools-structural, functional, and normative. Our approach enables decision-makers to anticipate the evolution of an emerging crisis and evaluate the effectiveness of different configurations on the response. The initial results of our simulation, based on an experiment which investigates the impact of three separate parameters, are also presented and reveal the joint effectiveness of the organizations involved.

Pervasive computing services exploit information about the physical world both to adapt their own behavior in a context-aware way and to deliver to users enhanced means of interaction with their surrounding environment. The technology to acquire digital information about the physical world is becoming more available, making services at risk of being overwhelmed by such growing amounts of data. This calls for novel approaches to represent and automatically organize, aggregate, and prune such data before delivering them to services. In particular, individual data items should form a sort of self-organized ecology in which, by linking and combining with each other into sorts of “knowledge networks” (KNs), they are able to provide compact and easy-to-be-managed higher level knowledge about situations occurring in the environment. In this context, the contribution of this paper is twofold. First, with the help of a simple case study, we motivate the need to evolve from models of “context awareness” toward models of “situation awareness” via proper self-organized “KN” tools, and we introduce a general reference architecture for KNs. Second, we describe the design and implementation of a KN toolkit that we have developed, and we exemplify and evaluate algorithms for knowledge self-organization integrated within it. Open issues and future research directions are also discussed.

A problem in large and dynamic networks consists in making available at each node global information about the state of the network. Gossip-based aggregation schemes are a simple yet effective mechanism to solve the problem. However, they have to cope with the dynamics either of the network and the values being aggregated and thus have to integrate specific solutions to deal with them. The contribution of this paper is to analyze and compare three different solutions to handle network and values dynamics in gossip-based aggregation schemes: (i) an epoch-based approach based on periodic restarts, (ii) an optimized epoch-based approach based on concurrent aggregation threads and (iii) an original approach based on values evaporation that does not require periodic restarts. Experimental results show that our proposal is effective and often more accurate than epoch-based techniques.

Adapting to current context of usage is of fundamental importance for pervasive computing services. As the technology for acquiring contextual information is increasingly available and as it is producing growing amounts of data, there is the need for tools to organize such data before delivering it to services. This produces a sort of "knowledge networks " representing comprehensive knowledge related to a "situation " in an expressive yet manageable way. In this paper, also with the help of a simple case study, we motivate the need for situation-awareness and for knowledge networks, introduce a reference architecture for knowledge networks, and exemplify a prototype implementation thereof. Finally, current and future research directions are discussed.