Computer Science/Engineering
Title: Art-inspired multimodal interactive systems for health and cultural welfare
Proposer: Prof. Antonio Camurri (DIBRIS)
Keywords: Interaction design/UX design, real-time multimodal analysis of expressive full-body movement, affective computing, computational models of non-verbal social emotions
Cultural welfare is a new integrated model for promoting the well-being and health of individuals and communities, through practices based on the visual and performing arts and cultural heritage. This model focuses on a concept of health understood globally as a restorative harmony in the relationship between a person's physical, intellectual, emotional, and social functions. Cultural welfare is based on the recognition of the effectiveness of specific cultural, artistic, and creative activities as factors that promote health, well-being, and social cohesion, foster active aging, combat depression and psychophysical decline, promote the inclusion of people with disabilities, marginalized, or disadvantaged individuals, support relationships between caregivers and patients, and mitigate degenerative diseases. Technologies represent a formidable catalyst for cultural well-being, providing both enabling platforms and interaction mechanisms for new paradigms in the experience of art and cultural heritage. Beyond providing content for novel theories and conceptual frameworks, and technologies that support cultural well-being experiences, art can suggest ways and means to humanize these technologies and to discover novel scientific approaches integrating human computer interaction, affective computing, embodied, soma-based interactive systems. The idea of art as a source of inspiration for science and technologies that mediate between humans and art opens the door to a new generation of arts-based theories and technologies for cultural well-being. However, the use of technologies often presents shortcomings that impede their adoption or even have the opposite effect, worsening human well-being. This proposal aims to address these issues through a transdisciplinary approach to science, technology and art. The activities are based on an in-depth analysis of the state of the art on innovative interactive systems inspired by visual and performing arts, characterized by the integration of human-computer interaction, interactive sonification of expressive movement, affective computing, artificial intelligence, with particular reference to the treatment protocols and the interactive systems DanzArTe [1] and RespirArTe [2] developed starting from the results obtained in the European project EnTimeMent (entimement.dibris.unige.it) by the Casa Paganini – InfoMus research center of Unige. The activities require basic skills in interaction design (in particular UX design and evaluation), in techniques for the real-time, multimodal, automated analysis of expressive full-body movement, affective computing, and computational models of non-verbal social emotions.
References
[1] A.Camurri, E.Seminerio, W.Morganti, C.Canepa, N.Ferrari, S.Ghisio, A.Cera, P.Coletta, M.Barbagelata, G.Puleo, A.Pilotto (2024) Development and validation of an art-inspired multimodal interactive technology system for a multi-component intervention for older people: a pilot study. Frontiers in Computer Science, 5.
[2] N.Corbellini, N.Ferrari, C.Gasparotti, G.Romano, G.Sicari, I.Vottero, N.Sukumaran, V.Prefumo, A.Finizio, A.Cera, A.Camurri (2026) Slow Mood, Aesthetic Resonance, and Embodied Interaction: Design Principles for Art-Aided Rehabilitation. Intl. Workshop on Movement and Computing MOCO 2026, Montpellier, ACM.
Geometric frameworks for human motion representations
Prof. Nicoletta Noceti (DIBRIS)
Keywords: Computer Vision; Machine Learning; Optimisation
Human motion representation and generation remain central challenges in Computer Vision and Machine Learning. Although human poses and movements have been historically modelled in Euclidean spaces, in recent years there has been a widespread diffusion of parametric models (e.g. SMPLs) in which body representations are based on joint rotations. The latter naturally belong to non-Euclidean manifolds such as SO(3) and their Cartesian products. This observation motivates the adoption of differential geometry, and in particular Riemannian geometry, as a principled framework for modelling the intrinsic structure of the human body and motion. This PhD project aims to develop a geometric framework for the representation, analysis, and generation of human motion. The research will investigate how geometric information can be incorporated into modern Machine Learning models through two complementary strategies [1,2]: embedding geometry directly into latent representations and integrating it into learning objectives via manifold-aware (e.g. geodesic) loss functions. Building on concepts from motor control theory, differential geometry, and geometric deep learning, the project will explore the benefit of interpreting natural human movements as geodesic trajectories on learned motion manifolds [3], extending existing geometric approaches for static pose modelling. This perspective will establish a connection between classical optimality principles in human motor behaviour, most notably the minimum-jerk model and its later Riemannian reformulation, and modern generative models. Downstream applications of interest may include human-robot interaction, rehabilitation and motor control, and sports.
The activity will be carried out at the MaLGa research center (https://ml.unige.it).
References
[1] H. Beik-Mohammadi, S. Hauberg, G. Arvanitidis, G. Neumann, and L. Rozo. Learning Rie- mannian manifolds for geodesic motion skills. In Robotics: Science and Systems (RSS), 2021
[2] H. Lan et al. End-to-end motion capture from rigid body markers with geodesic loss. arXiv:2511.16418, 2025.
[3] Y. He, G. Tiwari, T. Birdal, J. E. Lenssen, and G. Pons-Moll. NRDF: Neural Riemannian distance fields for learning articulated pose priors. In CVPR, pp. 1661–1671, 2024.
AI-driven GPU-accelerated bioinformatic applications for big data analysis
Prof. Daniele D'Agostino (DIBRIS)
Keywords: AI-driven GPU-accelerated bioinformatic applications for big data analysis
Modern bioinformatics has to deal with very large and heterogeneous datasets, produced by two complementary families of high-throughput technologies. On one side, single-cell technologies — including single-cell RNA sequencing (scRNA-seq), single-cell ATAC sequencing (scATAC-seq), and their spatial counterparts — describe biological samples at the resolution of individual cells, capturing the heterogeneity of tissues and the spatial organisation of cell populations: in essence, they tell us which cells are transcribing and where. On the other side, bulk technologies, in particular for epigenetics — such as Hi-C and other chromosome conformation capture techniques, methylation arrays and bisulfite sequencing, ChIP-seq and related profiling assays, together with bulk RNA-seq used as a reference layer — provide a complementary view: although they average over many cells, they aim at explaining why transcription occurs, by measuring chromatin accessibility, three-dimensional genome organisation, DNA methylation, and the binding of transcription factors and chromatin regulators. Both classes of data are inherently large and complex, and their joint interpretation is one of the main open problems of contemporary functional genomics. Traditional CPU-based pipelines are no longer sufficient to handle them in reasonable times, while Artificial Intelligence — and in particular deep learning — has proven especially effective in extracting patterns from such data. Training and applying these models at scale, however, requires GPU acceleration and modern High-Performance Computing infrastructures.
This PhD project aims to develop AI-driven, GPU-accelerated bioinformatic applications for the analysis of large omics datasets, combining single-cell technologies (scRNA-seq, scATAC-seq, and their spatial variants) with bulk technologies, in particular for epigenetics (Hi-C and other 3C assays, methylation arrays and bisulfite sequencing, ChIP-seq, complemented by bulk RNA-seq as a reference layer). The student will design deep learning models and GPU-accelerated pipelines for tasks such as cell-type identification, integration of transcriptomic and chromatin layers, trajectory inference, reconstruction of three-dimensional chromatin organisation, characterisation of the methylation landscape, and inference of the regulatory mechanisms underlying transcriptional activity. A central goal will be the integration of these complementary perspectives, connecting the cellular heterogeneity observed at single-cell level with the regulatory programs revealed by bulk epigenetic data. The project could also explore the construction of reference atlases and structured multi-omics databases, designed to be efficiently queried through Large Language Models (LLMs) and conversational interfaces. All methods will be designed to fully exploit modern hybrid CPU/GPU HPC infrastructures, leveraging containerised workflows and GPU-native libraries (e.g. RAPIDS, cuML, NVIDIA Parabricks, PyTorch distributed). Applications will focus on diseases involving the immune system — autoimmune disorders, chronic inflammation, immuno-oncology, and infectious diseases — in collaboration with biological and clinical partners, contributing to open-source, FAIR-compliant tools for the bioinformatics community.
Computational Methods for Modeling Immune System Dynamics
Prof. Giorgio Delzanno (DIBRIS), Prof. Raffaele De Palma (DIMI), IperMedImmune Research Infrastructure
The increasing availability of high-dimensional biological and clinical data has created new opportunities for computational approaches to model complex physiological systems. The immune system, characterized by highly dynamic and interconnected molecular and cellular networks, represents an ideal application domain for advanced artificial intelligence, machine learning, and digital twin technologies. This PhD project aims to develop novel computational methods for the integration, analysis, and modeling of large-scale multi-modal biomedical data, including genomics, transcriptomics, proteomics, metabolomics, single-cell, spatial omics, and clinical information. The research will focus on designing machine learning and explainable AI algorithms capable of identifying latent patterns, predicting disease progression, and characterizing individual variability in immune responses. A central objective will be the development of patient-specific digital twins of the immune system. These computational models will integrate mechanistic and data-driven approaches to simulate immune dynamics under physiological and pathological conditions and to predict responses to therapeutic interventions. Particular attention will be devoted to multimodal data fusion, graph-based representations of biological networks, temporal modeling of immune processes, uncertainty quantification, and interpretable AI techniques. The candidate will work within the interdisciplinary IperMedImmune infrastructure, collaborating with experts in immunology, bioengineering, and computational sciences. The project will leverage state-of-the-art high-performance computing resources and large-scale biomedical datasets to develop scalable computational frameworks for precision medicine. Expected outcomes include novel AI methodologies for systems immunology, predictive models of disease and treatment response, and digital twin platforms supporting personalized healthcare, clinical decision support, and in silico testing of therapeutic strategies. This research lies at the intersection of artificial intelligence, computational biology, complex systems modeling, and digital health.
At the core of the infrastructure is the integration of experimental (“wet-lab”) and computational (“digital”) approaches, with the goal of developing digital twins specifically designed for immune system disorders. These digital twins aim to decode the key mechanisms regulating immune responses in both pathological and physiological conditions, including responses to environmental and biological stressors, while accounting for individual variability. Such an approach supports the identification of personalized treatments, the discovery of novel molecular targets for therapeutic intervention, the in silico evaluation of treatment strategies, and the development of diagnostic and predictive algorithms for patient management.
The IperMedImmune Research Infrastructure combines the complementary expertise of two research groups at the Department of Internal Medicine and Medical Specialties (DIMI), and at the Department of Informatics, Bioengineering, Robotics and Systems Engineering (DIBRIS). In addition, it maintains established collaborations and formal agreements with several national and international institutions that actively contribute to its research activities. These include the Institute for Applied Computing (IAC-CNR), the Institute of Biomolecular Chemistry (IBC-CNR), and the Division of Immunology at Harvard University in Boston, USA.
Image reconstruction for biomedical imaging modelling uncertainty via flow matching
Prof. Luca Calatroni (CIL unit, Machine Learning Genoa Centre, DIBRIS, Università degli Studi di Genova, Italy)
Keywords: Image reconstruction for biomedical imaging modelling uncertainty via flow matching
This PhD project aims to develop novel methodologies for image reconstruction in biomedical imaging under uncertain and partially known forward models, leveraging recent advances in flow matching and generative modeling. The research will focus on two representative and highly relevant imaging modalities: fluorescence microscopy and X-ray computed tomography (CT). In many practical settings, image formation is governed by complex physical processes that are only approximately characterized, owing to unknown optical aberrations, imperfect calibration, scattering effects, detector nonlinearities, or variability in acquisition protocols. Consequently, the forward operator relating the underlying biological structure to the measured data is often unknown, partially known, or affected by modelling errors, leading to significant degradation in reconstruction quality. The proposed research will investigate flow matching as a principled framework for learning probability flows that map measurement distributions to high-quality image reconstructions while explicitly accounting for uncertainty in the imaging model. Particular emphasis will be placed on jointly learning image priors and corrections to the forward operator, developing robust inverse problem formulations capable of handling model mismatch, and quantifying reconstruction uncertainty. By combining physics-informed constraints with data-driven generative models, the project seeks to establish a new generation of reconstruction algorithms that remain reliable when the imaging system is imperfectly characterized. Applications will include super-resolution and denoising in fluorescence microscopy, as well as sparse-view and low-dose reconstruction in CT imaging, with the ultimate goal of improving the accuracy, interpretability, and robustness of biomedical image analysis.
Title: Human Computer Interaction in eXtended Reality: the role of visual realism, natural interaction, self- and companion avatars
Proposers: Manuela Chessa and Fabio Solari
Curriculum : Computer Science
Short Description Interaction in Virtual, Augmented or eXtended environments is influenced by several perceptual and behavioral properties, such as the visual realism of the environments and of the agents acting inside them [1]. In the literature, there are many works trying to understand the role of visual realism, natural interaction and behaviour of avatars and environments on interaction tasks, the sense of presence and perceptual judgments [2,3]. Moreover, the graphic properties of the avatars could lead to the observation of the uncanny valley problem [4]. Finally, the adoption of advanced interaction techniques, and manipulation of the behaviour and appearance of avatars can be the software support for both innovative entertainment applications and for understanding self-consciousness, with potential application in the context of neuro-rehabilitation, pain treatments, and to contribute to the understanding of neurological and psychiatric disease.
The aim of this research theme is to develop novel and efficient solutions to create and manipulate self- and companion avatars (in general virtual agents) inside virtual and extended reality environments, addressing the alignment and the spatial co-localization of both virtual and real elements, but also considering nonrealistic and “impossible” situations. In particular, the realism and the graphics properties of the virtual agents should be considered by addressing computer graphics techniques. Moreover, the virtual agents might be characterized by conversational abilities, showing the ability to adapt to different contexts [5]. The effect of the presence of the virtual agents, also considering their visual and behavioral realism and degree of complexity, will be examined with respect to efficacy of the interaction, embodiment, sense of presence and acceptance of the developed systems.
References
[1] Lee, Y. J., & Ji, Y. G. (2025). Effects of visual realism on avatar perception in immersive and non-immersive virtual environments. International Journal of Human–Computer Interaction, 41(7), 4362-4375.
[2] Pan Y, Steed A (2017) The impact of self-avatars on trust and collaboration in shared virtual environments. PLoS ONE 12(12): e0189078. https://doi.org/10.1371/journal.pone.0189078
[3] Bruno, F., Hussain, R., Chessa, M., Sacco, G., Manera, V., Addoum, M., & Solari, F. (2026). Are Digital Characters Suitable for Emotion Recognition Tasks? An Evaluation Study Using MetaHumans. In Proceedings of the 1st International Conference on Human-Computer Interaction in the Alps (pp. 65-71).
[4] Schwind, V., Wolf, K., & Henze, N. (2018). Avoiding the uncanny valley in virtual character design. interactions, 25(5), 45-49.
[5] Rad, S., Hussain, R., Chessa, M., & Solari, F. (2026). Designing Emotionally Intelligent Embodied Agents for Immersive Virtual Reality Experiences. In 2026 IEEE International Conference on Artificial Intelligence and eXtended and Virtual Reality (AIxVR) (pp. 128-136). IEEE.
Link to the group/personal webpage: pilab.unige.it
Title: Visual Perception and Semantic Knowledge for Human-Object Interaction Understanding
Proposer: Francesca Odone
Curriculum : Computer Science
Keywords: Computer Vision
Human-Object Interaction (HOI) understanding is a key capability for intelligent systems operating in complex environments, enabling the recognition of actions, activities, and semantic relationships between humans and surrounding objects. Recent advances have shown complementary strengths between purely visual approaches and Vision-Language Models (VLMs), yet the role of linguistic knowledge, visual reasoning, and data supervision in HOI detection remains only partially understood. This project aims to investigate the interplay between visual representations and vision-language priors for robust and data-efficient HOI understanding. Particular attention will be devoted to the design of lightweight and interpretable architectures that can effectively exploit both visual cues and semantic knowledge without relying on increasingly complex model designs. The research will explore methods for learning from limited annotations, handling ambiguous or incomplete interaction taxonomies, and generalising to novel objects and interactions in open-set scenarios. A second objective concerns the analysis and enhancement of existing HOI datasets, studying how annotation quality, label granularity, and dataset biases affect model performance and generalization. Novel strategies for weakly-supervised, semi-supervised, or synthetic-data-driven learning will also be investigated.
Title: Multimodal Machine Learaning for Multiple Sclerosis Progression: Linking MRI Biomarkers and Video-Based Motor Function
Proposers: Matteo Moro and Francesca Odone
Curriculum : Computer Science
Keywords: Machine Learning, Medical Image Analysis
The project would study data-efficient machine learning methods to integrate complementary MS biomarkers from medical imaging and markerless motion analysis. On the imaging side, it would build on lesion-level and patient-level MRI analysis, including advanced markers such as paramagnetic rim lesions, where recent work proposes QSM–FLAIR multimodal modelling under limited data and class imbalance. On the functional side, it would exploit RGB/video-based gait and movement descriptors, already shown to be attractive for non-invasive clinical assessment in MS. The project will complement the two approaches on which we already carried out initial research (see for instance [1] and [2,3]), and will open to clinically relevant directions including which visual biomarkers are actually predictive of disease progression or treatment response, and how can they be combined under scarce, heterogeneous, partially annotated clinical data.
References
[1] V Pignedoli, G Boffa, N Noceti, M Inglese, F Odone, M Moro "3D Classification of Paramagnetic Rim Lesions in Multiple Sclerosis via Asymmetric QSM-FLAIR Modeling " arXiv preprint arXiv:2606.16756, 2026
[2] L Turolla, M Moro, G Boffa, F Odone, M Inglese, LG Brayda, M Casadio "Video-based 2D markerless gait analysis in people with multiple sclerosis" Multiple Sclerosis and Related Disorders, 107285, 2026
[3] M Moro, G Marchesi, F Hesse, F Odone, M Casadio "Markerless vs. marker-based gait analysis: A proof of concept study" Sensors 22 (5), 2011, 2022
Systems Engineering
Title: Bio-Inspired Swarm Intelligence: Cooperative Flapper Drone Networks for Autonomous Greenhouse Monitoring and Microclimate Control
Proposer: Prof. Roberto Sacile (DIBRIS)
Keywords: Systems, Control, Optimization
Traditional greenhouse monitoring relies on static sensor networks that often miss localized microclimate anomalies, leading to undetected crop stress and inefficient resource use. This proposal introduces a novel, biomimetic approach to precision agriculture by deploying a cooperative team of bio-inspired "flapper" drones (ornithopters) designed to mimic the agile, low-impact flight of birds or insects. Unlike conventional quadcopters, these flapper drones offer enhanced safety around delicate crops, reduced downwash turbulence, and high maneuverability in confined spaces.The core of this research focuses on the development of decentralized control algorithms that enable the drones to operate as a cohesive swarm. By leveraging cooperative task allocation and distributed sensing, the team will dynamically partition the greenhouse space to map temperature, humidity, and CO2 gradients in real time. Furthermore, the swarm will not just monitor, but actively interface with the greenhouse environmental control systems—acting as mobile actuators to trigger localized misting, ventilation, or shading. This seamless integration of bio-inspired hardware and collective intelligence promises to maximize crop yield, minimize energy consumption, and pave the way for fully autonomous, resilient agricultural ecosystems.
Positions restricted to employees of the AITRUST company
Position 1
Title: Advanced Computer Vision and Artificial Intelligence Methodologies for Industrial Applications
Application restricted to AITRUST (https://www.aitrust.it/) employees
This project is part of the technological innovation program promoted by AITRUST S.R.L. and aims to investigate and develop advanced Computer Vision and Artificial Intelligence methodologies applicable to the various industrial contexts relevant to the company. The research activities will focus on the study of innovative techniques for the acquisition, processing, and interpretation of complex data, as well as on the design of efficient and scalable software architectures for managing the information generated by intelligent systems. Particular attention will be devoted to the integration of modern Computer Vision and Artificial Intelligence technologies into application processes, fostering the development of innovative solutions, technology transfer, and the enhancement of research outcomes within industrial environments.
Position 2
Title: Large Language Model-Based Methodologies and RAG Architectures for Knowledge Base Management and Optimization of Corporate Document Processes
Application restricted to AITRUST (https://www.aitrust.it/) employees
This project is part of the technological innovation program promoted by AITRUST S.R.L. and addresses the challenges of Generative Artificial Intelligence applied to knowledge management and the optimization of operational workflows. The research will primarily focus on the study of advanced methodologies based on Large Language Models (LLMs) and the engineering of Retrieval-Augmented Generation (RAG) architectures to enable secure access to corporate information. In addition to, or as a functional alternative to, information retrieval activities, the project will extend its research toward the development of intelligent systems capable of supporting and optimizing the production of complex documentation, such as technical proposals, project deliverables, structured reports, and data tables. The work will investigate model optimization criteria, vector database management, generated-text coherence, and controlled formatting techniques, while ensuring data protection and the scalability of solutions in industrial environments.