XLII Cycle
Up to 20 students can enroll in the XLII Cycle of the PhD Program in Computer and Data Science for Technological and Social Innovation.
● 14 positions are funded with a scholarship:
– 4 scholarships are without a specific topic (the topic can be chosen among the ones available in the first list “a” below);
– 10 scholarships have a specific topic (the topics are specific to the scholarship, and are shown in the list “b” after the first one).
● 3 positions are not funded (senza borsa).
Themes for the 4 scholarships without specific topic (and for the 3 positions without scholarship)
| a.1 | Closing the Gap Between Cybersecurity and Runtime Verification |
| Keywords: Cybersecurity, Runtime Verification, Runtime Monitoring, Post-Quantum Cryptography, AI-Generated Code, Software Security, Autonomous Systems, Cyber-Physical Systems, Formal Methods, Artificial Intelligence Research Motivation Cybersecurity has become a critical challenge for modern software-intensive systems, especially as applications increasingly rely on distributed architectures, autonomous behaviour, AI-generated code, and cryptographic protocols that must remain secure in the post-quantum era. While runtime verification (RV) provides strong foundations for monitoring system behaviour against formal specifications, a significant gap remains between theoretical RV techniques and practical cybersecurity needs in real-world systems. Cybersecurity properties are often dynamic, context-dependent, and difficult to fully capture at design time, particularly when systems evolve, interact with uncertain environments, or include components generated or adapted by AI tools. At the same time, the transition towards post-quantum cryptography introduces new risks related to incorrect implementations, insecure protocol usage, migration errors, and runtime misconfigurations. In this context, runtime verification can also be extended beyond passive monitoring and used as an adaptive decision-support layer for autonomous cybersecurity agents. In particular, RV monitors can act as a reinforcement learning (RL) feedback mechanism for both blue-team and red-team agents operating within simulated environments. Closing the gap between cybersecurity and runtime verification therefore represents a timely and impactful research direction. By developing practical RV-based techniques for detecting, explaining, and mitigating cybersecurity violations at runtime, this research aims to make formal monitoring approaches more applicable to real-world security problems. Research Objectives ● Security-Oriented Runtime Monitoring: Developing expressive runtime monitoring techniques for specifying and detecting cybersecurity-relevant behaviours, including policy violations, insecure API usage, suspicious interaction patterns, and deviations from expected security protocols. ● Runtime Assurance for AI-Generated Code and Behaviour: Investigating how RV can be used to monitor, validate, and constrain code or behaviours produced by AI systems, reducing the risk of vulnerabilities, unsafe adaptations, and unintended security-critical actions. ● Post-Quantum Security as a Case Study: Applying the proposed techniques to the migration towards post-quantum cryptography, focusing on runtime detection of misconfigurations, incorrect protocol usage, insecure fallbacks, and deviations from expected cryptographic workflows. ● Practical Tooling and Real-World Validation: Designing and implementing usable RV-based tools that can be integrated into realistic software development and deployment pipelines, supported by theoretical foundations but evaluated through practical cybersecurity scenarios. Expected Outcomes and Impact ● Practical runtime verification methods for improving cybersecurity in modern software systems, with a focus on deployability and real-world applicability. ● New theoretical and methodological foundations for expressing and monitoring cybersecurity properties at runtime, including dynamic and context-sensitive security requirements. ● RV-based techniques for reducing security risks introduced by AI-generated code, autonomous behaviour, and adaptive software components. ● Demonstrated applicability to post-quantum cryptography migration and other security-critical settings, supporting safer adoption of emerging technologies. ● Prototype tools and experimental evaluations showing how runtime monitoring can complement traditional cybersecurity practices such as testing, static analysis, auditing, and intrusion detection. This research will contribute to closing the gap between cybersecurity and RV, transforming RV from a mainly formal assurance technique into a practical approach for detecting and mitigating real-world cybersecurity risks in AI-enabled, evolving, and post-quantum software systems. Supervisor: Angelo Ferrando | |
| a.2 | Expressive Runtime Monitoring for Safe and Reliable Reinforcement Learning |
| Keywords: Reinforcement Learning, Runtime Verification, Reward Machines, Non-Markovian Tasks, Safety-Critical Systems, Artificial Intelligence, Autonomous Systems Research Motivation Reinforcement Learning (RL) has demonstrated remarkable potential in enabling autonomous systems to learn complex tasks through interactions with their environment. However, the safety and reliability of RL systems in safety-critical domains – such as autonomous driving, robotics, and healthcare – remain significant concerns due to the risk of reward misspecification and unintended behaviors. Traditional RL reward structures are limited in expressiveness, often incapable of adequately capturing complex, temporal, or non-Markovian tasks, resulting in unintended and potentially harmful outcomes. Runtime monitoring languages and Reward Machine-based approaches have recently emerged as promising frameworks to overcome these limitations. By specifying sophisticated reward conditions at runtime, such approaches enable safer, more reliable, and interpretable RL systems capable of handling complex temporal dependencies and context-sensitive tasks. Research Objectives This PhD project aims to advance the theory and practice of runtime verification for reinforcement learning systems. The primary objectives include: ● Expressive Reward Specification: Developing expressive runtime verification techniques to specify and evaluate complex non-Markovian reward structures that are beyond regular-language specifications. ● Safety Assurance in RL: Integrating runtime monitors with RL algorithms to provide continuous assurance of safety-critical properties and to mitigate risks associated with reward misspecification. ● Formal and Empirical Validation: Validating the developed methods through rigorous theoretical analysis and extensive experimental studies in realistic and safety-critical environments. Expected Outcomes and Impact ● Enhanced safety and reliability for reinforcement learning systems through expressive, runtime-verifiable reward specifications. ● New theoretical foundations and practical tools for integrating runtime verification within RL, significantly mitigating the risks of reward misspecification. ● Demonstrated applicability in real-world scenarios, improving trust and enabling broader adoption of RL technologies in safety-critical domains. This research will advance safe and reliable reinforcement learning, bridging theoretical expressiveness and real-world applicability. Supervisor: Angelo Ferrando | |
| a.3 | Neuroscience-Inspired Cognitive Architectures for Socially Intelligent Multi-Agent Systems |
| Keywords: Multi-Agent Systems, Neuroscience, Cognitive Architectures, Theory of Mind, Emotions, Human-Agent Interaction, Belief-Desire-Intention (BDI), Social Robotics, Artificial Intelligence Research Motivation Integrating insights from neuroscience into cognitive multi-agent systems (MAS) offers significant potential to enhance human-agent interactions across various fields, including social robotics, healthcare, education, and virtual environments. Despite extensive work on cognitive architectures, especially those based on Belief-Desire-Intention (BDI) models, explicit incorporation of neuroscientific findings – such as Theory of Mind, emotional processing, and memory structures – remains limited. Addressing this gap can lead to more adaptive, empathetic, and socially intelligent artificial agents. Research Objectives This PhD project aims to design, develop, and validate neuroscience-inspired cognitive architectures for socially intelligent MAS by: ● Neuroscience-based Cognitive Modeling: Integrating neuroscientific concepts (e.g., Theory of Mind, emotional regulation, episodic/semantic memory) into computational agent models. ● Emotion-driven Decision Making: Developing cognitive architectures that simulate human-like emotions to enhance decision-making and interaction believability. ● Advanced Theory of Mind: Improving agents’ abilities to understand and respond to human intentions, emotions, and mental states. ● Adaptive Communication Models: Designing computational frameworks distinguishing casual (“small talk”) and meaningful (“deep talk”) communication to dynamically enhance interactions. ● Experimental Validation: Conducting human-agent studies to evaluate the effectiveness, usability, and acceptance of these architectures in realistic scenarios. Expected Outcomes and Impact ● Novel neuroscience-inspired cognitive models improving social intelligence in MAS. ● Enhanced naturalness and empathy in interactions, fostering greater human acceptance. ● Broader applicability in critical domains such as healthcare, education, social robotics, and interactive digital environments. This research will significantly advance understanding of how neuroscience can inform the development of socially intelligent and human-compatible artificial agents. Supervisor: Angelo Ferrando | |
| a.4 | Formal Verification of AI-Based Autonomous Systems |
| Keywords: Formal Methods, Formal Verification, Model Checking, Runtime Verification, Multi-Agent Systems, Robotics, Artificial Intelligence, Reliability, Safety. Research Motivation The rapid adoption of autonomous systems – particularly multi-agent systems (MAS) – across sectors such as robotics, transportation, and smart infrastructures has transformed modern technological landscapes. However, the increasing complexity, adaptive learning mechanisms, and inherent unpredictability of AI-based autonomous systems pose significant challenges to ensuring their reliability and safety, especially when these systems directly influence human lives and critical infrastructure. Formal methods offer rigorous, mathematically grounded approaches for system verification, yet their application to autonomous, AI-driven systems is still limited by scalability, adaptability, and real-time operational constraints. Addressing these limitations is critical for building confidence in autonomous systems deployed in safety-critical applications. Research Objectives This PhD project aims to advance formal verification techniques specifically tailored to autonomous AI-based systems, with a strong focus on multi-agent systems (MAS). Key objectives include: ● Systematic Evaluation: Conducting an extensive survey and critical analysis of existing formal verification methods for AI-driven systems, pinpointing gaps and challenges related to scalability, adaptability, and robustness. ● Development of Novel Methods: Creating innovative verification algorithms and techniques capable of addressing the dynamic, adaptive, and high-dimensional characteristics of MAS, particularly those incorporating machine learning and AI components. ● Practical Validation: Implementing and validating these new methods through case studies involving realistic scenarios in robotics, transportation, or smart infrastructures, demonstrating their effectiveness and robustness. ● Generalizability and Impact: Ensuring that the developed verification methods can be generalized beyond MAS, providing practical tools and methodologies applicable across diverse AI-driven autonomous systems. Expected Outcomes and Impact The outcomes of this research are anticipated to significantly enhance the reliability, safety, and trustworthiness of autonomous AI systems. By bridging the existing gap between formal verification theory and practical AI implementations, this research will contribute essential tools to the scientific community and industry practitioners, ultimately facilitating safer deployment of intelligent, autonomous technologies. Supervisor: Angelo Ferrando (Unimore) Co-Supervisor: Vadim Malvone (Telecom Paris) | |
| a.5 | Natural Language Interfaces for Intelligent Agent-Based Virtual Environments |
| Keywords: Natural Language Processing, Multi-Agent Systems, Virtual Reality, Human-AI Interaction, Simulation, Cognitive Agents, Digital Twins Research Motivation The integration of artificial intelligence and virtual reality is reshaping how humans interact with complex systems.Traditional simulation tools for environments like smart factories or digital twins often demand advanced programming skills, limiting accessibility for domain experts without a technical background. The VEsNA framework (Virtual Environments via Natural Language Agents) [https://github.com/VEsNA-ToolKit] addresses this gap by enabling users to construct and manage virtual environments through natural language commands. By combining agent-based reasoning, natural language processing, and immersive simulations, VEsNA offers a user-friendly platform for designing and interacting with intelligent virtual spaces. Research Objectives This PhD project aims to advance the VEsNA framework by enhancing its natural language interfaces and agent-based reasoning capabilities. The key objectives are: ● Enhance Natural Language Understanding: Develop advanced NLP models tailored for interpreting user commands within virtual environments, ensuring accurate and context-aware interactions. ● Integrate Cognitive Agents: Implement cognitive agents capable of reasoning about user intents, environmental constraints, and potential outcomes to facilitate dynamic and intelligent responses. ● Expand Virtual Environment Capabilities: Augment the virtual environments with richer simulations, including physics-based interactions and real-time feedback mechanisms, to provide more immersive experiences. ● Evaluate User Interaction: Conduct user studies to assess the effectiveness, usability, and accessibility of the enhanced VEsNA framework for individuals with varying technical backgrounds. Expected Outcomes and Impact The project is expected to deliver a robust, user-centric platform that democratizes the creation and management of intelligent virtual environments. By enabling natural language interactions, the enhanced VEsNA framework will empower a broader range of users to design, simulate, and analyze complex systems without requiring extensive programming knowledge. This advancement has the potential to accelerate innovation in fields such as manufacturing, education, and urban planning, where virtual simulations play a critical role. Supervisor: Angelo Ferrando | |
| a.6 | Formal Methods for Certification of Autonomous Vehicle Systems under ISO 26262 |
| Keywords: Formal Verification, Runtime Verification, ISO 26262, Automotive Safety Integrity Level (ASIL), Autonomous Vehicles, Certification, Model Checking, Safety-Critical Systems Research Motivation The automotive industry is rapidly advancing towards higher levels of vehicle autonomy, introducing complex software-driven functionalities. Ensuring the safety and reliability of these autonomous systems is paramount, especially as they operate in unpredictable real-world environments. The ISO 26262 standard provides a framework for functional safety in road vehicles, emphasizing the need for rigorous verification and validation processes to achieve Automotive Safety Integrity Levels (ASIL). Traditional testing methods often fall short in covering the vast state space and dynamic behaviors of autonomous systems. Formal methods, offering mathematically rigorous techniques, present a promising avenue to enhance the certification process by providing stronger guarantees of system correctness and safety. Research Objectives This PhD project aims to integrate formal methods into the certification process of autonomous vehicle systems, aligning with ISO 26262 requirements. The primary objectives include: ● Framework Development: Design a comprehensive framework that incorporates formal verification techniques, such as model checking and theorem proving, into the safety lifecycle defined by ISO 26262. ● Runtime Verification Integration: Explore the application of runtime verification to monitor system behaviors during operation, providing continuous assurance and facilitating adaptive responses to unforeseen scenarios. ● Toolchain Evaluation: Assess existing formal verification tools for their applicability in the automotive domain, identifying gaps and proposing enhancements to meet industry-specific requirements. ● Case Studies: Apply the developed methodologies to real-world autonomous vehicle subsystems, evaluating their effectiveness in achieving ASIL compliance and improving overall system safety. Expected Outcomes and Impact The research is expected to yield a validated methodology for integrating formal methods into the certification process of autonomous vehicles, enhancing the rigor and efficiency of achieving ISO 26262 compliance. By bridging the gap between formal verification techniques and industry certification standards, this work aims to contribute to the development of safer autonomous systems and foster greater trust in their deployment. Supervisor: Angelo Ferrando | |
| a.7 | Autonomic computing for collective self-adaptive systems |
| Keywords: Autonomic computing, distributed systems, adaptive systems, IoT, autonomous vehicles Research objectives: Collective adaptive systems are more and more spreading in our life. Situations like sets of mobile phones or IoT devices are becoming real and can be exploited to support human activities, but this requires appropriate approaches to be managed; in the future we envision sets of autonomous vehicles that must be coordinated. Autonomic computing is a very good candidate paradigm to address these scenarios. The objective of the research is to define a framework for the development of collective self-adaptive systems, based on autonomic computing. The framework will be composed of a methodology that guides the developers in the development addressing the self-* properties, and of tools enabling the support for the developers. Some case studies will be proposed to test the framework. Proposed research activity: • State of the art in autonomic computing • State of the art in collective adaptive systems • Definition of a framework for autonomic computing in collective adaptive systems • Definition of a methodology • Definition of case studies • Test of the framework • Participation to relevant international school Supporting research projects (and Department) H2020 – FIRST (FIM) Possible connections with research groups, companies, universities. Dr. Antonio Bucchiarone, FBK Trento (I) Dr. Lai Xu, Bournemouth University (UK) Prof. Emma Hart, Edinburgh Napier University (UK) Prof. Marco Aiello, Stuttgart University (D) Supervisor: Prof. Giacomo Cabri | |
| a.8 | Software engineering for autonomous vehicles |
| Keywords: Software engineering, autonomous vehicles, distributed systems, adaptive systems, IoT Research objectives: Autonomous vehicles are spreading and more and more research is needed to enable their engineering. In particular, both the single vehicle and the coordination of sets of vehicles rely on software components that must be designed, implemented and verified; the current methods and methodologies could not be suitable for this new scenario. Appropriate approaches are needed. The objective of the research is studying the (meta)requirements of the development of software components and systems for autonomous vehicles, in order to define one or more approaches that are suitable for this scenario. Proposed research activity: • State of the art in software engineering • State of the art in autonomous vehicles • Definition of approaches to engineer the development of software components and systems for autonomous vehicles • Definition of a methodology • Definition of case studies • Test of the proposed approaches • Participation to relevant international school Supporting research projects (and Department): WASABI 2023 (FIM) Possible connections with research groups, companies, universities: Dr. Antonio Bucchiarone, FBK Trento (I) Prof. Emma Hart, Edinburgh Napier University (UK) Prof. Marco Aiello, Stuttgart University (D) Supervisor: Prof. Giacomo Cabri | |
| a.9 | Real-time Collaborative 3D Editing |
| Keywords: Research objectives: Collaborative editing ala Google Docs is still not widespread in the 3D world. The goal of this thesis is to explore real-time collaborative models for 3D. We obtain encouraging first results by extending distributed version control, ala git and GitHub, to 3D content. In this thesis, we plan to explore the design space of real-time collaborative 3D editing, focusing on local-first models such as CRDTs and differential synchronization. The thesis work be fully focused on developing new algorithm and data structure for graphics, disregarding all other aspects of real-world collaborative systems, such as security, networking, authentication, etc. No prior knowledge of distributed systems is required.Our prior work on version control for 3D includes: • MeshGit • MeshHisto • cSculpt • SceneGit • others under review Supervisor: Prof. Fabio Pellacini | |
| a.10 | AI-assisted Computer Graphics: AI-Assisted Editing of Procedural Programs |
| Keywords: Computer Graphics, Artificial Intelligence, Neural Networks, Program Synthesis, Automatic Differentiation Research objectives: Assets used to describe 3D scenes are either measured, hand-painted, or synthesized by programs. The latter category, sometimes called procedural models, is the most scalable in the production of large amounts of content. But procedural models are hard to author. The goal of this project is to explore new algorithms for authoring procedural models using recent results in artificial intelligence. In particular, we will explore the idea of writing procedural programs using LLMs, using recent result in LLMs for code generation. The main novelty of this work lies in linking final appearance with program synthesis using multimodal models for the synthesis. Supervisor: Prof. Fabio Pellacini | |
| a.11 | Non-Euclidean Generative Models for 3D Assets |
| Keywords: Computer Graphics, Artificial Intelligence, Generative Models Research objectives: Textures and environment maps are fundamental components in the representation of visual appearance in computer graphics. Traditional generative models for visual content are typically defined in Euclidean domains, making them difficult to apply directly to signals defined on curved surfaces or spherical domains. The goal of this project is to adapt current generative models for image generation to the synthesis and editing of textures on manifolds and environment maps on the sphere. Supervisor: Prof. Fabio Pellacini | |
| a.12 | Multi-level organization of complex systems |
| Keywords: Information theory, evolutionary computation, nonlinear dynamics, criticality, adaptation Research objectives: The detection of emerging intermediate structures in complex systems is not always a trivial task, while in contrast their characterization can lead to a meaningful description of the overall properties of the system, and in this way to its understanding. A large part of these structures is characterized by groups of variables (genes, chemical species, individuals, agents, robots…) that appear to be well coordinated among themselves and have a relatively weaker interaction with the remainder of the system. Notable examples are functional neuronal regions in the brain, autocatalytic systems in chemistry, or communities in socio-technological systems. We are therefore interested in identifying and studying such structures, proposing algorithms for their understanding and analyzing data from the world of biology, of socio-technological systems and of artificial systems. Supervisor: Prof. Marco Villani | |
| a.13 | Evolving artificial systems |
| Keywords: Information theory, evolutionary computation, nonlinear dynamics, criticality, adaptation Research objectives: Finding general properties of evolving systems has proven extremely difficult. A particularly interesting proposal is that evolution (either natural or artificial) drives complex systems towards “dynamically critical” states, which may have relevant advantages with respect to systems whose dynamics is ordered or disordered. According to this hypothesis, these critical systems can provide an optimal balance between stability and responsiveness. In this thesis we aim at determining under which conditions this turns out to be the case, by using abstract models. The systems described by these models will interact in an abstract “environment”, and the conditions under which critical systems have an edge will be analyzed. To achieve this ambitious goal, we will exploit the synergy among dynamical systems methods, information theory and evolutionary computation. Supervisor: Prof. Marco Villani | |
| a.14 | Innovation ecosystems, industrial districts, and global value chains: a network approach |
| Keywords: social network analysis, industrial districts, global value chain, sustainability, innovation Research objectives: Globalization has increased the speed of competition, continuously generating new opportunities and threats, where flexibility and innovation play a fundamental role. Industrial districts and global value chains are particularly affected by international dynamics and processes, and innovation is key for companies – located in industrial districts and embedded in global value chains – aiming to be successful in global markets. As part of their strategy, these companies rely on business networks to become more innovative and improve their performance. Social Network Analysis (SNA) is a powerful tool to assess the importance of (local and global) business networks and their impact on companies’ performance, and the proposed research will be focusing on the analysis of: • inter-organizational networks; • intra-organizational networks; • dynamic business networks; • novel approaches for mapping business networks. Supervisor: Prof. Stefano Ghinoi Possible connections with other international universities: • Prof. Bodo Steiner, University of Helsinki (FI) • Prof. Riccardo De Vita, Manchester Metropolitan University (UK) • Prof. Guido Conaldi, University of Greenwich (UK) | |
| a.15 | Latency Sensitive and Safety Critical GPU-accelerated real-time computing |
| Keywords: GP-GPU, Massively parallel computing, Real-Time, Compute Architecture, Programming Models. Research Objectives: Nowadays cyber physical systems are characterized by data hungry algorithms within a wide variety of applications. This implies facing notable challenges for reaching the desired performance, hence the hardware deployed in domain such as Automotive, Robotics, Telecommunication and industrial automation are implemented as heterogeneous systems in which multi-core CPU hosts work in concert with massively parallel accelerators. In this context, a widely known accelerator is the Graphic Processing Unit (GPU), a hardware device designed to maximize compute throughput for general purpose computations (GP-GPU). It is not trivial, however, to exploit the full potential of the GPU processing power due to the notable architectural differences between GPUs and more traditional multi-core CPUs. Significant effort is therefore required, for instance, to exploit the recently released architectural features of modern GPUs, such as specialized cores for tensor processing and traversal of bounding volume hierarchies. Moreover, GPUs are designed to maximize throughput, hence inherently sacrificing latencies. This research aims at understanding how programming models, APIs and compilers could be enhanced in order to facilitate the work of the system engineer when implementing GPU accelerated applications, but also for accounting for stringent latency and safety requirements imposed by modern applications in the autonomous systems domain. Proposed research activity: ● State of the art on GP-GPU computing: from programming models to applications. ● Design and implementation of mechanisms that act at the level of APIs/programming models to enable real-time/safety critical GPU computing. ● Enhancing current compilers/source-to-source translators for simplyfing the programmer access to the GPU’s specilized cores (e.g. for tensor operations). ● Participation to relevant international schools and conferences. Supervisor: Nicola Capodieci Co-supervisor: Andrea Marongiu | |
| a.16 | Enhancing Performance and Efficiency of Deep Learning Models for Human-machine Interaction Applications |
| Keywords: neural machine translation, image captioning, language detection, performance, deep learning Research Objectives: Human-machine interaction domain is composed of important applications such as neural machine translation, image captioning, and language detection. However, these models often suffer from issues such as poor performance, high computational complexity, and limited scalability. The proposed research project aims to investigate and develop novel techniques to improve the performance of these models. The research project also aims to evaluate the effectiveness of these techniques on various benchmark datasets and compare them with existing state-of-the-art techniques. Proposed Research Activity: ● Literature review and state-of-the-art analysis of neural machine translation, image captioning, and language detection models ● Development of novel techniques to improve the performance of these models, such as attention mechanisms, transfer learning, and model compression ● Evaluation of the proposed techniques on various benchmark datasets, such as WMT, COCO, and LDC ● Comparison of the proposed techniques with existing state-of-the-art techniques ● Participation in relevant international conferences and workshops Supervisor: Roberto Cavicchioli Co-supervisor: Alessandro Capotondi | |
| a.17 | Digital Intelligent Assistants for Industry 5.0 |
| Keywords: voice assistant, human-in-the-loop, digital factory, data analytics Research objectives: Voice assistants, alternatively mentioned as conversational agents or Digital Intelligent Assistants (DIA), allow users to interact intuitively by using their natural language. In the industrial sector, the adoption of conversational agents has the potential to drive the digital transformation of organizations, improve both customer and user experience, and make their internal processes more efficient. This PhD project proposal aims to delve into the main research challenges that emerge in the design and development of DIAs in the industrial context where shopfloor operators need to interact with the available physical assets and data sources to solve data analytics requests. Topics of interest are, for instance, the integration of LLMs, approaches for DIA evaluations in realistic contexts and their continuous improvements. Proposed research activity: ● Investigate platforms and technology stacks for DIA development ● Address DIA benchmarking and evaluation issues in Industry 5.0 ● Explore Tool-augmented LLMs solutions ● Explore continual learning solutions in the context of DIA models Supporting research projects (and Department): The PhD student will be hosted at the Department of Physics Informatics and Math where she/he will be a member of the ISGroup (www.isgroup.unimore.it) led by Prof. Federica Mandreoli. The group has been working in different projects on digital factories and it is currently involved in the Horizon Europe project WASABI https://wasabiproject.eu/ . Possible connections with research groups, companies, universities: On the topics of the proposal, the ISGroup has connections with BIBA – Bremer Institut für Produktion und Logistik (Germany), ICCS (Greece), and Sapienza University of Rome (Italy). Moreover, connections with the companies involved in the WASABI project are currently active. Supervisor: Prof. Federica Mandreoli | |
| a.18 | Data-centric machine learning |
| Keywords: Model fairness, model robustness, data drift, real-world data Research objectives: Artificial Intelligence (AI) has traditionally relied on two key components: data and algorithms. However, the dominant model-centric paradigm has primarily focused on refining algorithms, often treating data as static and secondary. This has led to increasingly complex and opaque models, which tend to be unreliable and unfair when applied to real-world scenarios. A new paradigm — data-centric AI — has recently emerged, placing data at the center of the AI development process and extending its role beyond the pre-processing phase. This PhD project aims to investigate the main research challenges within this paradigm, including topics such as data drift, learning from real-world data, model interpretability, robustness, and fairness. Proposed research activity Investigate key principles and challenges in data-centric machine learning Propose novel approaches to address these challenges Develop prototypes of the proposed solutions Evaluate the prototypes using international benchmarks Compare the results with state-of-the-art methods Supporting research projects (and Department) The PhD student will be hosted at the Department of Physics Informatics and Math where she/he will be a member of the ISGroup (www.isgroup.unimore.it) led by Prof. Federica Mandreoli. Possible connections with research groups, companies, universities. On the topics of the proposal, the ISGroup has connections with Prof. Paolo Missier (University of Birmingham UK) and Prof. Paolo Ciaccia (Univ. Of Bologna). Supervisor: Prof. Federica Mandreoli | |
| a.19 | Environmental information and communication: fake news, bias and distortion, and data analysis. Possible impacts on policy and public opinion. |
| Keywords: Environmental information, Information and communication theory, social media and conversation analysis, Cognitive and perceptual biases and climate change, Qualitative and quantitative methods in data analysis. Research objectives: The purpose of this research area is to analyze the processes of dissemination and production of fake news, cognitive bias related to information on the issues of environment, climate crisis and climate change. The intention is to work on textual corpora and conversations regarding environmental communication and information issues, both online and social media, and offline, by: (a) acquiring skills regarding the relationship between quantitative and qualitative data analysis; and (b) using tools such as NVivo and Altas and qualitative methodologies related to thematic, narrative analysis and different forms of “netnography” of data. Supervisor: Prof. Federico Montanari, DCE | |
| a.20 | AI, social discourses, metaphors and agency. |
| Keywords: AI, responsibility, social discourse, sociosemiotics, agency, images. Research objectives: This research project aims to explore and critically analyze the social and sociosemiotic implications of artificial intelligence technologies, with a particular focus on two main lines of inquiry. Firstly, we intend to examine the metaphors, rhetoric, and public discourse that develop around AI. This aspect concerns the analysis of the languages, narratives, and interpretative frameworks through which AI is represented, discussed, and made socially understandable in the media, popular culture, political and institutional discourse, in images, as well as in scientific and technological contexts. The goal is to understand how these discursive constructions contribute to the production of meaning, influence collective perceptions, and shape social expectations and fears regarding artificial intelligence. Second, the project aims to investigate how it is possible to conceptualize, analyze, and—from a design perspective—imagine and develop forms of non-human agency and social agents. This involves theoretical and critical work on the very idea of agency, questioning established categories and proposing alternatives that take into account the specificity of artificial intelligence. This area of research questions how AI can be understood not only as a technical tool, but as a social subject (or quasi-subject) endowed with a form of presence, influence, and participation in communicative, cultural, and relational dynamics. Overall, the project aims to contribute to an in-depth and interdisciplinary reflection on the growing role of artificial intelligence in the contemporary social fabric, with the goal of proposing interpretative models and critical perspectives that help to understand—and perhaps rethink—the relationships between humans and non-humans in the age of cognitive automation. Supervisor: Prof. Federico Montanari, DCE, Unimore. Possible connections with research groups, companies, universities: University of California, San Diego, Univ. Liège, Univ. Paris Cité. | |
| a.21 | Sustainable Edge Computing Architectures for Green IoT Applications |
| Keywords: green computing, edge computing, sustainable IoT, energy-aware systems, load balancing, network simulators, microservices, performance evaluation Research Objectives: In the last few years Edge computing has emerged as a novel approach to support modern IoT applications based on microservices. These applications typically involve sensors located in multiple geographic locations producing big amounts of data. While edge computing addresses the limitations of cloud-only models by bringing computation closer to data sources, it also introduces new challenges in terms of energy consumption, carbon footprint, and resource efficiency. The main goal of this research is to design and develop sustainable, energy-efficient distributed computing systems—spanning both edge and cloud infrastructures—that support modern IoT applications while minimizing environmental impact. Central to this goal is the development of an adaptive decision-making model for workload distribution that incorporates carbon footprint as a core weighting factor, alongside traditional metrics. The model will estimate the optimal placement of microservice-based workloads across heterogeneous computing nodes (edge and cloud), based on real-time conditions and environmental metrics. The study will leverage realistic smart mobility and IoT application traces, evaluate the trade-offs between energy efficiency and performance, and validate the effectiveness of the proposed model through simulation and experimentation. Proposed Research Activities: ● State-of-the-art energy-aware and carbon-aware computing models in edge and cloud environments, with a focus on IoT applications ● Collection and analysis of real-world IoT and smart mobility traces ● Study and modeling of carbon emissions associated with various edge and cloud deployment strategies ● Design of a carbon-aware workload distribution model ● Development of simulation and/or emulation tools to test the proposed model in realistic edge-cloud IoT scenarios ● Test and validation of the proposed solutions platform ● Participation to relevant international schools and conferences Supervisor: Prof. Claudia Canali | |
| a.22 | Virtualized collaborative platforms for STEM and computer science education |
| Keywords: collaborative platforms, data science, virtualization, performance evaluation Research objectives: The European job market is suffering an increased mismatch due to a severe shortage of workers with STEM and computer science skills. Several factors contribute to this issue, such as the need to train and support educators to include computer science in teaching activities and the lack of tools allowing trainers to collaborate. The main goal of this research is to design and develop a collaborative platform that supports educators in sharing projects, methodologies and experiences related to computer and data science education, including evaluation methodologies focused on continuous monitoring of students’ self-efficacy. The aim of the platform is twofold: a) provide educators with the possibility to share projects and innovative teaching approaches; b) allow educators to find resources and communities helpful to improve their experience. The underlying technology will allow remote execution of Notebooks combined with storage support and containerization environments, to offer advantages such as storage and performance scalability, collaboration, privacy, and security thanks to self-hosted solutions. Proposed research activity: ● Analysis on state of the art on educational collaborative platforms and tools ● Design and development of a collaborative platform based on Python Notebooks and online storage tool (e.g. NextCloud) ● Test and validation of the developed platform, with active collaboration with school educators ● Participation to relevant international schools and conferences Supervisor: Prof. Claudia Canali | |
| a.23 | Efficient CNN and LLM deployment on edge devices |
| Keywords: Efficient AI; TinyML; Embedded systems; Edge devices; HW/SW co-design; Research Objectives: The rapid advancements in AI and machine learning have led to the development of state-of-the-art models, such as Convolutional Neural Networks (CNNs) and Large Language Models (LLMs). However, deploying these models on resource-constrained edge devices remains a significant challenge due to their computational and memory requirements. This research aims to explore efficient deployment strategies and on-device training mechanisms for AI models on emerging edge devices (e.g., FPGA, ASIC accelerators, like Axelera and Hailo). The project will focus on optimizing inference, enabling continual learning and domain adaptation by exploiting hardware accelerators for edge computing. The outcomes will contribute to the development of robust, adaptive, and efficient AI systems for real-world edge applications. Supporting research projects (and Department): dAIEDGE (https://daiedge.eu/). Possible connections with research groups, companies, universities: SUPSI Lugano (Svizzera) ETH Zurich (Svizzera) KU Leuven (Belgio) Northeastern University (USA) Supervisor: Prof. Andrea Marongiu Co-supervisor: Dott. Alessandro Capotondi, UNIMORE Co-supervisor: Prof. Francesco Restuccia, Northeastern University (USA) | |
| a.24 | Compiler-aided parallel programming model for next generation high performance predictable heterogeneous platforms |
| Keywords: Compilers; Parallel Programming Models; Runtime; Heterogenous Systems; Research Objectives: The primary focus of this project is to address the programming challenges associated with emerging high-performance heterogeneous systems. The project aims to develop compiler and runtime support for heterogeneous and parallel programming models, explicitly targeting the cyber-physical systems domain (robotics, automation, manufacturing). The objective is to enhance the adoption of these systems by improving performance and timing predictability, while maintaining a simple programming interface. Proposed Research Activity: ● In-depth study of the challenges involved ● Design and development of compiler and runtime system extensions specifically tailored for Commercial-off-the-Shelf (COTS) platforms and open-source hardware architectures like RISC-V. ● Validate the proposed solutions on real-life problems from the targeted application domains ● Participation in relevant international conferences and workshops Prior work on the topic: ● HEPREM: https://ieeexplore.ieee.org/document/9035630 ● HERO: https://dl.acm.org/doi/10.1145/3295816.3295821 ● PULP: https://pulp-platform.org/ Possible connections with research groups, companies, universities: ● ETH Zurich, Switzerland ● Barcelona Supercomputing Center, Spain ● RI.SE., Sweden Supervisor: Prof. Andrea Marongiu Co-supervisor: Dott. Alessandro Capotondi | |
| a.25 | Hardware-software Co-Design for Resilient Adaptive AI/ML in FPGA Hardware |
| Keywords: Efficient AI; FPGA; Embedded systems; HW/SW co-design; Research objectives: The overall objective is to design and evaluate new hardware-software co-design strategies that will enable real-time reconfiguration of artificial intelligence (AI) algorithms implemented in field-programmable gate arrays (FPGA) to ensure resilience against intentional and unintentional perturbations and to support fast and effective AI reconfigurability in case of change of operational objectives and/or systems constraints. The ideal outcome of the project would be reconfigurable AI software/hardware architectures that will withstand by design extremely dynamic scenarios. The research will be prototyped on testbeds composed of System-on-Chip (SoC) platforms connected to wireless transceivers and sensors for multi-modal data acquisition. Applications of interested include spectrum sensing and computer vision-based classification and control. Possible connections with research groups, companies, universities: Northeastern University (USA) Supervisor: Prof. Andrea Marongiu Co-supervisor: Dott. Alessandro Capotondi, UNIMORE Co-supervisor: Prof. Francesco Restuccia, Northeastern University (USA) | |
| a.26 | Exploring adaptive and context-aware edge computing for next generation autonomous systems |
| Keywords: intelligent systems, adaptive computing; embedded systems; autonomous systems Modern computing systems expose a high degree of automation and tightly interact with the surrounding environment, acting as a bridge between the real world and the cybernetic one. Autonomous vehicles and robots are a prominent example of state-of-the-art in the field. Such systems must process a massive amount of information from heterogeneous data sources that can be on-board sensors and environmental data transmitted via wireless and wired connectivity (such as V2X in the vehicular case), but at the same time the shall react to dynamically changing working conditions in real-time, depending on the operational context. The edge computing paradigm is currently based on high-performance embedded computers that provide energy-efficient computation. However, they still miss this flexibility of re-adapting themself to the operational context, and to do it in real-time. Reseach Objectives Study of state-of-the-art design methodologies for edge intelligent systems. Design novel adaptive strategies (e.g., time adaptation, continual learning) to enable operational adaptivity directly on the edge systems. Study and evaluation on realistic use-case, such as autonomous vehicles or smart-city infrastructures. Possible connections with research groups, companies, universities: KU Leuven (Belgium) TU Graz + AVL GmbH (Austria) Supervisor: Dr. Paolo Burgio Co-supervisor: Dr. Alessandro Capotondi | |
| a.27 | Automated Cyber Operations |
| Keywords: cyber security, graph theory, planning, automation, artificial intelligence Research objectives: The volume and complexity of modern attacks and defenses (as testified by the size of the CVE, CWE, CAPEC, MITRE ATTACK catalogues) is making manual exploitation and hardening of systems increasingly unfeasible. In the next few years we foresee an adoption of even more automated offensive and defensive tools and decision makers. This research thesis is motivated by the need for frameworks, algorithms, tools that help the human operator to carry on offensive operations (vulnerability assessments, penetration testing) and defensive operations (systems hardening, source code auditing) in a semi-automated fashion. The main goal is to explore the potential for automation and artificial intelligence in the activities involved in classical security assessments (representation of security-related knowledge, decision making, task planning and execution, creating digital twins, monitoring of progress, reporting) in order to improve efficiency and efficacy. Supervisor: Mauro Andreolini | |
| a.28 | Scoring Systems for Cyber Security |
| Keywords: cyber security, graph theory, algorithms, cyber ranges, monitoring To increase the resilience of their infrastructures, both military and civilian organizations have started to train security personnel on cyber ranges, pre-arranged virtual environments through which it is possible to effectively simulate realistic security scenarios on a system architecture closely resembling the original one. Training goals are many and diverse in nature: to discover vulnerabilities in existing systems, to harden existing systems, to evaluate the security of a soon-to-be deployed component, to teach secure programming practices, to perform incident response on a compromised system. Unfortunately, current evaluation strategies of student performance in an exercise share a severe limitation: they are focused exclusively on goal achievement (yes or no), and not on the specific path followed by the student. Therefore, giving a more precise assessment and, more generally, understanding the reasons behind success or failure, is impossible. This research is motivated by the need for techniques, algorithms and tools to model user activities in an exercise, compare the path carried out by the student with an “optimal” path devised by an instructor and suggest avenues for improvement. The goal is to propose novel cyber scores that can be used to capture the abilities of a student (speed, precision, ability to discover new vulnerabilities), highlight potential weaknesses, compare different students in the same scenario. Supervisor: Mauro Andreolini | |
| a.29 | Attacking and defending cryptographic protocols implementations |
| Keywords: Cyber security, applied cryptography, secure development, side channel attacks Research objectives: Implementing cryptographic schemes and protocols is a hard task, related to having interdisciplinary knowledge on theoretical and applied cryptography, secure code development, and real-world attacks to code and system architectures. Moreover, many dedicated attack and defense techniques specifically related to cryptographic implementations have been designed throughout years of research, and novel techniques are still emerging, in particular related to implementation of post-quantum cryptography and to defense against advanced attack surfaces encompassing gray and white -box security. This research thesis involves the study of such techniques, on studying their applicability to existing and emerging protocols and systems, and on designing novel tools to detect non-secure implementations. Proposed research activity: State of the art in cryptographic engineering attacks and defenses Analyzing emerging security threats due to implementation of post-quantum cryptography Designing and implementation of known and novel strategies for attacking and defending cryptographic implementations Advisor: Luca Ferretti Co-advisor: Mauro Andreolini | |
| a.30 | Efficient confidential and verifiable data management strategies |
| Keywords: Cyber security, Applied cryptography, Databases, Data structures Research objectives: Emerging security solutions try to minimize systems attack surface and mitigate information leakage even in presence of partially compromised systems. Among those, promising advanced cryptographic strategies are being studied to enable efficient query execution on encrypted databases and to verify correct execution of queries on data. These security solutions help mitigating data breaches perpetrated by external adversaries or even worse by legitimate insiders, and enable strong auditing strategies for outsourced data managed by external parties. The research activity focuses on studying state-of-the-art for database security and encryption, analyzing practical techniques for achieving practical performance and acceptable security, designing engineering strategies to embed these techniques within real-world database systems. Proposed research activity: • State of the art on database encryption and verification • Analyzing trade-offs in terms of security and performance • Designing novel strategies for improving security and performance of encrypted data • Evaluating security of existing and novel techniques • Studying solutions for proper integration with existing database management systems • Implementation of proof of concepts for performance evaluation Supervisor: Luca Ferretti | |
| a.31 | Privacy, Security and Resiliency of Authentication and Authorization |
| Keywords: Cyber security, Applied cryptography, Authentication, Authorization, Distributed systems, Anonymity, Tracking Research objectives: The emergence of novel computing systems is showing limitations of existing security solutions for authentication and authorization procedures, either due to limited capabilities of constrained devices and networks, limited usability and scalability within systems consisting of a huge amount of devices, attack surfaces including physical access to devices. Moreover, novel security paradigms require designing augmented security guarantees including increased privacy of users identities and reduced trust in identity providers. This research involves students in studying state-of-the-art authentication and authorization protocols, applied cryptography, and network security, acquiring expertise in analyzing and designing threat models and for novel secure computer systems. Proposed research activity: • State of the art on Web Authentication and Authorization, on privacy preserving protocols and distributed architectures • Analyzing and measuring trade-offs of privacy-preserving authentication • Analyzing deployment in real-world systems • Designing secure and practical authentication systems • Implementation of proof of concepts for heterogeneous platforms Supervisor: Luca Ferretti | |
| a.32 | Multi-user activity recognition |
| Keywords: activity recognition, inertial sensor, machine learning Research objectives: Human Actitivy Recognition (HAR) is a set of techniques which identifies the activity a user is performing, through the analysis of a set of sensors which describe actions and movements from users. This allows to configure a computing system based on the current scenario of the user, and provide a more tailored experience. More recently classic HAR system have also been proposed to detect and classify group actions, in which it is not only important to classify signals obtained from a single person, but also to share data among a set of devices carried by different people, which is the main topic of this thesis work. Proposed research activity: • State of the art on HAR for groups of people • Design and development of a data gathering platform • Analysis and study of techniques to realize HAR for groups of people • Participation to relevant international schools and conferences Possible connections with research groups, companies, universities: University of Stuttgart (Germany), University of Bamberg (Germany), University of New South Wales (Australia) Supervisor: Prof. Luca Bedogni | |
| a.33 | Digital Twins on Mobile devices |
| Keywords: digital twins, mobile development, simulation Research objectives: In recent years, the concept of a Digital Twin— a digital replica of physical systems used for monitoring, simulation, and optimization—has gained significant traction across various industries. This thesis investigates the technological advancements enabling this transition, including the integration of mobile sensing, real-time data processing, and edge computing. It focus specifically on mobile devices, owned by user, to raise the privacy of the data produced by such mobile devices, without the need to offload it completely to an edge or cloud server. Several case studies are examined to demonstrate the practical applications of mobile-based Digital Twins across different sectors such as healthcare, manufacturing, and smart cities. In healthcare, for instance, the implementation of Digital Twins on smartphones can enhance personalized health monitoring and predictive analytics. In manufacturing, mobile Digital Twins enable real-time monitoring and maintenance of equipment even in remote locations. For smart cities, they provide dynamic management and optimization of urban infrastructure. Proposed research activity: • State of the art on DT and mobile devices • design and development of mobile DT • implementation of shadowing • user acceptance Possible connections with research groups, companies, universities: UCI, UNSW Supervisor: Prof. Luca Bedogni | |
| a.34 | Efficient, scalable and flexible device to edge offloading |
| Keywords: edge computing, machine learning, smart systems, context-aware-computing Research objectives: Edge computing refers to computing units which are placed at the edge of the network, close to the end devices. They typically serve as a lower latency computing unit to process information which devices cannot process on their own, due to battery or computational constraints. In IoT systems, tasks are often monolithic and defined with a strict software and hardware dependency. In this thesis, the PhD candidate will explore how to decompose monolithic tasks in atomic functions, that can be deployed on systems leveraging a fluid computing architecture. Proposed research activity: • State of the art on Edge computing offloading • Monolithic task decomposition • Design and implementation of a general purpose testbed for Device-to-edge offloading • Design, implementation and evaluation of Device-to-edge offloading protocols and techniques • Participation to relevant international schools and conferences Possible connections with research groups, companies, universities: University of California, Irvine (USA), CNR Italy, University of Bologna, Sapienza University Supervisor: Prof. Luca Bedogni | |
| a.35 | Explainable Artificial Intelligence for Trustworthy Human-Centred Decision Systems |
| Keywords: Explainable AI, Trustworthy AI, Human-AI Interaction, Machine Learning, Decision Support Systems, Medical AI, Industrial AI, AI Education Research Motivation: Artificial Intelligence systems are increasingly adopted in high-impact domains such as healthcare, industrial automation, predictive maintenance, education, and decision support. However, many state-of-the-art AI models, especially deep learning and large language models, often operate as “black boxes”, producing outputs that are difficult to interpret, justify, and critically assess. This lack of transparency limits their adoption in contexts where reliability, accountability, safety, and human trust are essential. Explainable Artificial Intelligence (XAI) aims to address this challenge by developing models, methods, and interfaces capable of making AI decisions understandable to human users. In industrial settings, explainability is crucial to support operators, engineers, and managers in understanding system recommendations, detecting anomalies, preventing failures, and improving process optimization. In the medical domain, explainability is even more critical, as AI-based predictions may influence diagnosis, triage, treatment planning, and patient monitoring. Clinicians need systems that do not simply provide predictions, but also offer understandable evidence, uncertainty estimates, and reasons behind their outputs. At the same time, explainability is not only a technical requirement but also an educational and social need. AI is rapidly entering schools, universities, and professional training environments both as a tool and as a subject of study. Therefore, future citizens, students, teachers, and professionals must be equipped with the conceptual and practical skills needed to understand how AI systems work, what their limitations are, and how to use them responsibly. In this sense, explainable AI can become a bridge between advanced technological development and AI literacy, supporting a more conscious, critical, and inclusive use of intelligent systems. Research Objectives: This PhD project aims to investigate and develop explainable AI methods and tools for trustworthy decision-making in industrial, medical, and educational contexts. The main objectives include: ● Explainable AI Models and Methods: Study and develop machine learning and deep learning approaches that provide transparent, interpretable, and human-understandable explanations of their predictions and recommendations. ● Industrial Decision Support: Apply XAI techniques to industrial scenarios such as predictive maintenance, anomaly detection, quality control, production optimization, and human-machine collaboration, with the goal of improving trust, safety, and operational efficiency. ● AI Literacy and Education: Develop educational frameworks, tools, and demonstrators that introduce explainable AI in schools and universities, treating AI both as a learning tool and as a core topic of digital and scientific education. ● Evaluation of Trust and Usability: Assess the effectiveness of explainable AI systems through quantitative and qualitative evaluation, including user studies with domain experts, students, teachers, and non-technical users. Proposed Research Activities: ● Analysis of industrial and medical use cases where lack of interpretability limits AI adoption. ● Development of XAI methods for selected machine learning and deep learning models. ● Implementation of prototypes for industrial, medical, and educational scenarios. ● Evaluation of the proposed systems through benchmark datasets and user-centred studies. ● Development of teaching modules and educational activities on explainable AI for schools, universities, and professional training. Expected Outcomes and Impact: The project is expected to deliver: ● Novel explainable AI methods and prototypes for high-impact decision-making contexts. ● Human-centred tools that improve transparency, trust, and accountability in AI-assisted industrial and medical applications. ● Educational resources and demonstrators to support AI literacy in schools and universities. ● A methodological framework for evaluating explainability not only from a technical perspective, but also from the viewpoint of human understanding, usability, and responsible adoption. Supervisor: Dr. Giorgia Franchini Co-supervisor: Prof. Marko Bertogna |
Themes for the 10 scholarships with specific topic
Scholarships funded by Companies or University Departments
| Theme | Funding entity | |
| b.1 | Machine-learning-based object detection, tracking, and motion forecasting for extreme autonomous driving | HipeRT S.r.L |
| Keywords: machine learning, autonomous driving, perception, detection, motion forecasting Research Motivation Object detection-related tasks in autonomous driving are usually solved by fusing different approaches, varying from classical computer vision to machine-learning algorithms. Neural methods are commonly used and effective in nominal scenarios for object detection, while for motion forecasting are usually monitored and evaluated against limited but robust classical algorithms. Inherently robust, accurate, and complete perception and motion forecasting pipelines through machine learning techniques are still an open topic that can increase the reliability of autonomous systems, particularly in edge case scenarios. Research Objectives The PhD project aims to build a motion forecasting module able to robustly and accurately predict the detection’s future actions. In particular: ● Investigate the state of the art of machine learning-based object detection and motion forecasting ● Design, develop, and benchmark different motion forecasting algorithms in edge scenarios such as autonomous driving in bad weather, off-road, and racing. ● Test and validate the solution with synthetic and real data. ● Apply the solution in a physical system. Expected Outcomes and Impact ● Novel motion forecasting algorithms for autonomous vehicles ● Demonstration of the capabilities on real-world applications Supervisor: Prof. Marko Bertogna Co-supervisor: Dott. Ayoub Raji (HipeRT S.r.L.) | ||
| b.2 | Motion planning and control at and beyond the defined Operative Design Domain (ODD) in extreme autonomous driving | FIM Department |
| Keywords autonomous driving, operative design domain, motion planning, control, mining Research Motivation The definition of the Operational Design Domain (ODD) is crucial to implement a complete and safe autonomous system, limiting its operation to specific areas and conditions. Building an autonomous driving system capable of driving at the limit of the ODD while detecting and correctly reacting to conditions out of the ODD is not straightforward. At the state of the art, common solutions include an on-off switch between nominal and out-of-ODD driving. Novel solutions are needed to improve the overall safety and reliability of such systems in a specific industry environment, such as heavy-duty mining trucks. Research Objectives This PhD project aims to build a planning and control module capable of detecting and correctly adapting to different conditions, self-optimizing its operation. In particular, the objectives are: ● Investigate state-of-the-art motion planning and control algorithms capable of self-optimization and adaptability, including model-based and machine-learning-based methods. ● Develop an ODD-aware motion planning and control autonomous system inherently capable of adapting the actions and methodology relative to the actual level of proximity to the bounds of the ODD. ● Prototype and test the solution on a high-fidelity simulation and a real-world application. Expected Outcomes and Impact ● Novel planning and control algorithms for autonomous driving at and beyond the ODD ● Demonstration of the capabilities on a high-fidelity simulation and a real-world application ● The implemented solution should improve the productivity of the vehicle’s operation Supervisor: Prof. Marko Bertogna Co-supervisor: Dott. Ayoub Raji (HipeRT S.r.L.) | ||
| b.3 | FIM Department | |
| Keywords: vehicle swarming, motion planning, autonomous driving, control, cooperative planning, mining Research Motivation: Planning and Control for autonomous driving pose critical challenges when the scenarios involve multiple agents operating simultaneously in the same environment, with a strict dependence on others’ actions. This is particularly noticeable in edge-case applications such as autonomous heavy-duty mining trucks and industrial vehicles. Algorithms that consider both cooperative and non-cooperative actions are essential for safe and reliable autonomous capabilities. Research Objectives: The PhD project aims to search, design, implement, and validate algorithms for cooperative and non-cooperative motion planning and control for heavy-duty autonomous vehicles. In particular, the objectives can be summarized in: ● Search and benchmark actual state-of-the-art planning and control algorithms. Model-based, machine-learning-based, and more classical methods should be investigated. ● Investigate and benchmark vehicle swarming algorithms. ● Develop a motion-planning and control software module capable of driving heavy-duty mining trucks cooperatively. ● Test the algorithms in a digital twin and hardware-in-the-loop platform. ● Validate the algorithms in real vehicles and environments. Expected Outcomes and Impact ● Novel planning and control algorithms for vehicle swarming ● Demonstration of the capabilities on a high-fidelity simulation and a real-world application ● The implemented solution should improve the productivity of the vehicle’s operation Supervisor: Prof. Marko Bertogna Co-supervisor: Dott. Ayoub Raji (HipeRT S.r.L.) | ||
| b.4 | Co-Scheduling of CPU, Memory, and Hardware Accelerators for Deterministic Real-Time Guarantees in Modern Computing Systems | FIM Department |
| Keywords: Heterogeneous Computing, Hardware Accelerators (GPU/NPU/FPGA), Real-Time Systems, Resource Contention, Memory Orchestration, CPU-GPU Co-scheduling, Determinism, Embedded AI, HW/SW Co-design. Research Motivation: The evolution of intelligent autonomous systems has led to the adoption of heterogeneous architectures, where general-purpose CPUs are paired with specialized hardware accelerators (GPUs, NPUs, or FPGAs) to handle computationally intensive workloads like deep learning and signal processing. However, these accelerators make challenges even harder for temporal predictability, when it comes to time-sensitive applications. The CPU and accelerators often share a single path to the main memory. This creates a high-pressure bottleneck where high-bandwidth accelerator traffic can “starve” the CPU, in addition intra-CPU memory interference, leading to even more unpredictable execution delays for critical control tasks. Existing scheduling theories typically treat accelerators as black-box offloading engines, failing to account for the complex interplay between heterogeneous processing units and the shared memory interconnect. This research aims to evolve traditional scheduling into a multi-dimensional orchestration problem. The goal is to manage not just the execution flow across heterogeneous cores, but also the concurrent access to the shared memory fabric to ensure end-to-end real-time guarantees. Research Objectives: • Characterize cross-domain interference between CPUs and hardware accelerators in shared-memory heterogeneous SoCs. • Design holistic co-scheduling policies that synchronize task execution on accelerators with CPU-side resource management. • Develop memory-aware arbitration mechanisms to prioritize safety-critical memory transactions over non-critical high-throughput accelerator data. • Investigate software-defined resource partitioning (e.g., memory bandwidth throttling, cache locking) tailored for heterogeneous pipelines. • Optimize the offloading process to minimize synchronization overhead and data transfer latency between processing domains. • Propose a unified formal model to verify the schedulability of real-time tasks spanning both CPUs and accelerators. • Validate the integrated framework using AI-driven real-time workloads (e.g., autonomous driving perception) on heterogeneous platforms Expected Outcomes and Impact: • An integrated orchestration framework for the simultaneous management of CPU, memory, and accelerator resources. • Deterministic execution models for heterogeneous pipelines, allowing for reduced safety margins and higher hardware efficiency. • Driver-level or OS-level implementations of priority-aware memory and accelerator controllers. • New performance metrics and benchmarks specifically designed for real-time heterogeneous computing. The expected impact is to provide a foundation for the next generation of safe, AI-enabled autonomous systems. By addressing the contention between general-purpose logic and high-performance accelerators, this research will enable predictable, low-latency execution for complex multimodal applications in safety-critical environments. Supervisor: Marko Bertogna Co-supervisor: Paolo Valente | ||
| b.5 | Low-Latency Integration of Intra-Vehicle and V2X Communication Protocols for Cooperative and Autonomous Driving | FIM Department |
| Keywords: V2X, IEEE 802.11p, CAN Bus, Intra-Vehicle Networks, Cooperative Driving, Low Latency Communications, Edge Computing, Vehicular Telemetry, Real-Time Systems Research Motivation: The transition toward cooperative and autonomous driving requires a tight integration between intra-vehicle communication systems (e.g., CAN Bus, sensors, Electronic Control Units) and inter-vehicle communication technologies (V2X). Currently, these two domains are often designed and managed independently, leading to inefficiencies in the propagation of critical information and increased end-to-end latency. In safety-critical scenarios such as platooning or cooperative intersection management, even millisecond-level delays can significantly impact system effectiveness. At the same time, modern vehicles generate high-frequency telemetry data that can be exploited to improve situational awareness, provided that efficient mechanisms for real-time data selection, aggregation, and dissemination are available. This research aims to bridge the gap between intra- and inter-vehicle communication layers by developing integrated, latency-aware architectures for next-generation vehicular systems. Research Objectives: Analyze latency and throughput characteristics of intra-vehicle protocols (e.g., CAN, CAN-FD) and V2X technologies (IEEE 802.11p). Design an integrated architecture for real-time acquisition and processing of vehicular telemetry data. Develop data selection and prioritization mechanisms for V2X transmission. Investigate edge computing techniques to reduce end-to-end latency. Implement and validate prototypes in real-world environments (e.g., motorcycles/vehicles within the MASA LivingLab). Evaluate system performance in cooperative scenarios (collision avoidance, cooperative awareness, platooning). Expected Outcomes and Impact Novel architectures for latency-aware intra/inter-vehicle integration. Optimization models for information flow across sensors, ECUs, and V2X networks. Experimental prototypes validated on real-world platforms (MASA LivingLab). High-impact scientific publications in vehicular networking and intelligent transportation systems. Contributions toward best practices for cooperative driving systems. Expected Outcomes and Impact: The expected impact includes improved road safety, enablement of advanced cooperative applications, and advancement of the state of the art in low-latency V2X communications. Supervisor: Marko Bertogna Co-supervisor: Carlo Augusto Grazia | ||
| b.6 | Latency-Aware Orchestration of Multimodal Applications in Intelligent Living Labs for Smart Mobility | FIM Department |
| Keywords: Living Lab, Smart Mobility, Low Latency Systems, Edge AI, Computer Vision, V2X, ETSI Standards, Data Fusion, Wireless Communications, Real-Time Orchestration Research Motivation: Modern smart mobility infrastructures require the coexistence of heterogeneous applications operating in both the visual domain (e.g., detection, classification, tracking) and the electromagnetic domain (wireless communication of metadata). However, these systems are typically developed independently, without coordinated management of computational and networking resources. In a LivingLab environment such as MASA, where real-world data is continuously acquired, processed, and disseminated, there is a strong need for intelligent orchestration mechanisms capable of minimizing end-to-end latency while ensuring scalability and robustness. An additional challenge lies in data standardization: the lack of shared formats limits interoperability across platforms and reduces the effectiveness of large-scale urban information dissemination. This research aims to design and implement an integrated framework for latency-aware orchestration of multimodal applications, with a particular focus on ETSI-compliant data standardization. Research Objectives: Design an intelligent LivingLab architecture for managing multimodal applications. Develop orchestration strategies for computer vision pipelines (detection, tracking, classification). Integrate wireless communication systems for efficient dissemination of metadata. Investigate data fusion techniques across visual and electromagnetic domains. Define and implement ETSI-compliant message formats for urban state dissemination. Optimize end-to-end latency through edge computing and dynamic resource allocation. Validate the proposed solutions within the MASA LivingLab using real-world scenarios. Expected Outcomes and Impact A framework for integrated management of low-latency multimodal applications. Orchestration algorithms for distributed edge-cloud environments. ETSI-aligned data models and message formats for smart mobility applications. Experimental datasets and validation platforms based on the MASA LivingLab. Scientific publications and potential collaborations with industry and standardization bodies. Expected Outcomes and Impact: The expected impact includes improved efficiency of smart urban infrastructures, reduced latency in critical applications, and enhanced interoperability across heterogeneous systems. Supervisor: Marko Bertogna Co-supervisor: Carlo Augusto Grazia | ||
| b.7 | Multimodal Generative Digital Twins for Personalized Education: Heterogeneous Encoders and Adaptive Content Creation | DISMI Department |
| Keywords: Educational AI, multimodal transformers, personalized learning, digital twins, heterogeneous encoders, generative models, knowledge tracing Research Objectives: This PhD project aims to develop the first multimodal generative digital twin for personalized education. While digital twins have been proposed as learning models to represent student knowledge and adapt educational content, current approaches are limited to unimodal (text-only) representations and lack the ability to dynamically generate diverse content formats. This project will create a framework that extends the student digital twin with multimodal generative capabilities, capable of producing personalized explanations (text), illustrative diagrams (images), and explanatory audio, dynamically adapted to each learner’s profile, knowledge state, and contextual constraints. The research integrates recent advances in heterogeneous encoders for NLP, AI-based image captioning, and modern multimodal transformer architectures, combined with efficient GPU-accelerated inference on edge platforms. The key challenge is to achieve high-quality, personalized multimodal content generation within the latency constraints of interactive educational applications, enabling deployment on resource-constrained devices such as tablets, smart classrooms, and edge servers. Proposed Research Activities: ● State-of-the-art survey on digital twins for education, multimodal generative models, and heterogeneous encoder architectures ● Design of a multimodal knowledge representation for the student digital twin, integrating text, image, and audio modalities, extending heterogeneous encoder techniques to educational contexts ● Development of a multimodal content generation pipeline that dynamically selects and synthesizes the most appropriate content format based on the learner’s profile and knowledge state ● Design and implementation of GPU-optimized inference for real-time multimodal content generation, including model compression and efficient decoding strategies ● Implementation of a dynamic modality-switching mechanism that adapts content format based on learner profile, contextual constraints, and pedagogical objectives ● Deployment and validation of a prototype system in real educational contexts with active collaboration with school educators and assessment of pedagogical effectiveness through controlled studies ● Participation in relevant international schools and conferences Supervisor: Prof. Roberto Cavicchioli Co-supervisor: Prof. Marco Furini | ||
| b.8 | DHisGram – Syntax out of Africa: Deep History through Human Grammars Subproject 1 – Modeling syntactic diversity through the Parametric Comparison Method | DCE Department |
| Keywords: Language History, Syntactic Modeling, Historical Comparison Overview This research theme is part of the DHisGram project, which applies the Parametric Comparison Method (PCM) to investigate the deep linguistic history of two highly controversial macro-areas: South-East Asia and the Pacific, and the Americas. This research theme can be developed into several subprojects, defined according to the candidate’s expertise. This subproject is situated in the field of comparative and historical linguistics, with a specific interest in cross-linguistic comparison based on formal grammatical properties. It is intended for applicants who wish to investigate linguistic diversity and historical relations among languages through a theoretically informed and methodologically explicit comparative framework. Proposals may focus on one or more language groups located in any of the two macro-areas under analysis. Selection and justification of the languages to be investigated must be consistent with the research goals of the project. Here follow some suggestions: the lists are indicative, not exhaustive. Southeast Asia:Sino-Tibetan (Sinitic: e.g. Mandarin, Cantonese; Tibeto-Burman); Austroasiatic (e.g. Vietnamese, Khmer, Mon, Wa) Pacific and Oceania: Austronesian (Oceanic: e.g. Tongan, Samoan, Māori, Hawaiian, Äiwoo; Formosan; Malayo-Polynesian: e.g. Malay, Tagalog, Javanese, Malagasy; West New Guinea; Papuan) Australia: Pama-Nyungan, Non-Pama-Nyungan The Americas: Eskimo-Aleut (Inuktitut), Na-Dene (e.g. Navajo), Salishan (e.g. Straits Salish), Muskogean (e.g. Chickasaw), Quechuan (e.g. Quechua/Quichua), Uto-Aztecan (e.g. Pima, Papago, Nahuatl), Mayan (e.g. Tzotzil), Arawakan (e.g. Garifuna), Tupian / Tupi-Guaraní (e.g. Munduruku, Paraguayan Guaraní), Waikuruan (e.g. Kadiweu), Macro-Jê (e.g. Kaingang), Cariban (e.g. Kuikuro) Research focus Analysis of the internal structure of the nominal domain in a selected set of languages, compared on the basis of a system of nominal parameters, which will be further enriched and expanded to accommodate to typologically diverse languages Analysis of the distribution of syntactic diversity across the selected languages, through the adoption of statistical methodologies and computer assisted taxonomic techniques Analysis of the phylogenetic structure of the group under investigation, to test existing phylogenetic hypotheses, especially for languages or areas whose historical interpretation remains controversial or insufficiently explored Research tasks 1) Data collection and documentation. Gathering of the relevant data from elicitation sessions with language experts and native speakers, with the possibility of consulting grammars, corpora, databases, depending on the available resources. Data collection/elicitation is guided by the PCM-questionnaire 2) Syntactic analysis. Comparative analysis of selected structures within the nominal domain. Formulation of novel theoretical hypotheses, if needed, and refinement of the parameter system 3) Comparative evaluation. Measurement and interpretation of similarities and differences across the selected languages, including formulation and testing of historical or phylogenetic hypotheses Critical dialogue with the literature. Comparison of the results with existing classifications, reconstructions, or competing accounts in the relevant literature Proposed activities ● Literature review on the theoretical framework, the languages under investigation, and the relevant comparative and historical debates ● Training in comparative methodology, formal analysis, and, where relevant, quantitative or computational tools ● Data collection and organization, including the construction of a structured comparative dataset within the PCM database ● Analytical work on selected nominal subdomain(s), with discussion of the emerging results in relation to the broader research question ● Systematic collaboration and interaction with specialists on the languages/areas involved Supervisor: Cristina Guardiano Co-supervisor: one or more members of the International Teaching Staff with expertise in the relevant fields. | ||
| b.9 | DHisGram – Syntax out of Africa: Deep History through Human Grammars Subproject 2 – Scaling the Parametric Comparison Method via Large Language Models and Graph Analytics: A Computational Framework for Syntactic Inference | DCE Department |
| Keywords: Neuro-Symbolic Computing, Low-resource NLP, Graph Neural Networks, Latent Space Inference. Overview This research theme is part of the DHisGram project, which applies the Parametric Comparison Method (PCM) to investigate the deep linguistic history of two highly controversial macro-areas: South-East Asia and the Pacific, and the Americas. This research theme can be developed into several subprojects, defined according to the candidate’s expertise. Background The PCM framework has so far relied on a system of rules developed within classical formal linguistics, without the support of automated techniques. While effective for medium-scale comparison, this approach has not yet been tested on a systematic, global scale. The project addresses this gap by integrating formal and comparative linguistic theory with State-of-the-Art (SotA) Machine Learning (ML) models, specifically designed to handle structured symbolic knowledge and sparse data. Research focus The core objective is to design and implement a model for automatic data extraction and analysis that is fully compatible with the PCM framework. The model will support linguists in identifying the underlying syntactic structures that generate observable patterns across languages.The work will be based on targeted datasets: rather than large corpora, the PCM relies on carefully selected data from controlled syntactic configurations, interpreted through speaker judgments. This poses specific challenges for Informed ML: developing algorithms that can learn from “Small Data” by incorporating linguistic vincles as inductive biases, especially for under-documented languages. Research tasks The project will focus on three tightly connected directions: 1) Data construction and representation: Multi-Task NLP (Natural Language Processing ) & LLMs (Large Language Models ) Develop methodologies to collect and structure data that reliably encode the observable syntactic patterns (e.g. word orders and their interpretations) generated from the underlying rules to be compared. Implementation of Zero-Shot and Few-Shot Information Extraction (IE) pipelines to map descriptive grammatical texts into PCM-compliant binary vectors. 2) Pattern detection and correlation: Graph-based Dependency Discovery Identify co-occurrence patterns among surface structures, with the goal of uncovering non-obvious dependencies linked to shared underlying rules. Modeling the PCM parameters as a Knowledge Graph. We will apply Link Prediction and Relation Extraction to uncover non-obvious dependencies (hidden rules) and perform Matrix Completion to infer missing parameters in languages where data is incomplete or noisy. 3) Cross-linguistic modeling Analyze how abstract syntactic rules are distributed across languages and mapping them at a macro-area scale. Development of Generative Latent Variable Models (such as Variational Autoencoders – VAE) to project syntactic structures into a continuous manifold. This allows for the application of Unsupervised Clustering and Manifold Learning to reconstruct phylogenetic trajectories and measure syntactic divergence in a mathematically rigorous latent space. Proposed activities ● Critical review of relevant literature (computational methods implemented in formal linguistic, historical comparison and SotA data extraction (Transformers, GNNs-Graph Neural Networks) ● Evaluation and adaptation of existing analytical and extraction techniques. Implementation of a hybrid architecture that combines the logical consistency of PCM (Symbolic) with the generalization power of Neural Networks (Deep Learning) ● Design and implementation of novel models tailored to PCM requirements. Integration of Bayesian Neural Networks or Conformal Prediction to provide confidence intervals for the inferred syntactic rules, crucial for historical reconstructions Supervisor: Marko Bertogna Co-supervisor: Giorgia Franchini Co-supervisor: Cristina Guardiano |
Previous Cycles (PhD Course on Computer and Data Science for Technological and Social Innovation)
Research Theses of the XLI Cycle
Research Theses of the XL Cycle