Recent advancements in generative AI have enabled the integration of Large Language Models (LLMs) into Virtual Environments (VEs). This integration has been particularly useful to create Embodied Conversational Agents (ECAs) capable of engaging in meaningful interactions. However, such ECAs often lack environmental awareness, which may limit the quality of interactions and conversations. A potential solution could be to use Vision Language Models (VLMs), which could empower ECAs with structured environmental knowledge. However, the use of VLMs for such applications is largely underexplored. This paper explores how VLMs can be employed to integrate environmental awareness in ECAs and investigates their impact on player perception. To this end, an architecture leveraging multi-image inference and scene graph generation was designed. A within-subject user study was then conducted in a virtual reality game, comparing ECAs with and without environment-aware capabilities. The evaluation considered several dimensions, including perceived knowledge, intelligence, factuality, willingness of future interaction and conversational abilities.
Cognitive robotic systems require low-latency, reliable perception to support safe and context-aware interaction with humans. Multimodal biometrics (e.g., face and gait) improve robustness under real-world variations but the resulting pipelines are often too compute-intensive and memory-heavy for edge deployment. This paper explores the possibility of exploiting multimodal redundancy to enable aggressive optimization of person recognition stacks without excessive system-level degradation. We analyze a modular face+gait pipeline and apply architectural exploration together with post-training compression, namely quantization and pruning, targeting the dominant bottleneck (face recognition). Experiments on CASIA-B under a Rank-1 protocol over near-frontal views show that while face-only accuracy degrades significantly under compression (0.971 to 0.825), score-level fusion remains comparatively stable (0.986 to 0.966) while improving throughput from 49.8 to 75.2 FPS and reducing the deployed footprint from 295.2 MB to 111.9 MB. These results indicate that multimodal fusion can serve as a resilience mechanism, enabling lightweight and optimized perception modules for real-time cognitive agents.
Smart and secure access to vehicles is a crucial aspect of the evolving automotive industry. This article focuses on the development of an end-to-end multimodal biometric recognition framework that identifies people walking toward a vehicle from an RGB video feed. The framework is based on a deep learning pipeline for person detection and tracking, face and gait feature extraction, and fusion of the two modalities at the score and feature level. Traditional face recognition (FR) systems can suffer from variations in lighting and occlusions. In order to deal with these issues, the proposed framework integrates face and gait features with the aim to enhance accuracy. The pipeline is modular, enabling seamless integration of new models for each step of person identification without the need for additional training. Baseline face and gait recognition (GR) models, as well as score-and feature-level fusion (FLF) techniques, are evaluated on subsets of the CASIA-A and CASIA-B datasets. Experimental results show that weighted mean score-level fusion (SLF) significantly improves both Rank-1 accuracy and verification accuracy (TAR@FAR = 10(-5)) over unimodal baselines. Overall, the reported work provides insights into current limitations and suggests directions for future research about secure identity verification in vehicles.
Human social interactions are undergoing profound transformations, driven by the rapid evolution of communication technologies. In particular, Extended Reality (XR) technologies have reshaped how people connect, collaborate, and communicate across distances. Beyond simply reproducing real-world interactions, XR enriches communication by enabling the exchange of social and emotional cues, clarifying intent, enhancing emotional expression, and supporting collaboration. The recent advancement of artificial intelligence further amplifies this potential by allowing adaptive, context-sensitive augmentation. Research on social augmentation contributes both theoretically, by deepening our understanding of human social and emotional interaction, and practically, by informing the design of XR systems that foster meaningful, inclusive, and ethically grounded interactions. However, research often overlooks long-term engagement, inclusivity, ethics, and the subtle dynamics of social-emotional exchange. This workshop brings together researchers, designers, and practitioners to explore the challenges, opportunities, and methodologies of Social XR, fostering interdisciplinary dialogue and laying the foundation for sustainable, inclusive, and ethically responsible research and practice.
Computer vision has evolved dramatically from traditional handcrafted image processing methods to advanced deep learning models. However, despite achieving notable results, these purely statistical methods often suffer from limitations in interpretability, data efficiency, generalization, and reasoning capabilities. Neurosymbolic (NeSy) artificial intelligence has emerged as a promising paradigm that integrates the powerful pattern recognition of neural networks with the structured, logical reasoning of symbolic systems. This article provides a comprehensive introduction to NeSy applications in computer vision, covering tasks such as image classification, object detection, scene understanding, and action recognition. We explore key NeSy frameworks, including logic tensor networks, highlighting their ability to improve interpretability, robustness, and reasoning. Finally, we discuss the challenges and future directions this promising hybrid approach poses toward explainable and trustworthy computer vision solutions.
Veterinary sciences are undergoing a paradigm shift in the way students are trained for practical skills since classical educational approaches, based on apprenticeships and wet labs, are no longer applicable due to the increasingly specialized nature of veterinary teaching hospitals, rising costs, and ethical issues. This paper presents and evaluates the effectiveness of a Virtual Reality Training System (VRTS) designed to allow students to train on equine upper airway endoscopy procedures without the risks and pressure associated with manipulating delicate instruments and approaching live animals, as well as avoiding stress for ill horses. Several pathological conditions have been implemented in the VRTS, whose design was carried out in collaboration with domain experts, aiming to improve the students’ diagnostic skills. The VRTS was evaluated against the established, traditional learning approach (based on a classroom lecture with image- and video-based instructional material) on a sample from the population of veterinary students by following a between-subjects design. Training effectiveness was assessed by evaluating the participants performing the endoscopy procedure on alive ill horses. Results suggests that students trained with the devised VRTS showed better control over the endoscopic tool and exhibited higher diagnostic accuracy during the in-vivo evaluation than those trained with the traditional learning approach.
Navigating complex indoor environments can be challenging, particularly for individuals with impairments. This paper presents a navigation system using Virtual Reality (VR) and haptic feedback to support wheelchair driving training. Directional cues are provided through a custom multi-actuator vibrotactile handle, designed to replace a standard wheelchair joystick knob. It combines hard and soft materials to deliver precise and discernable vibrations at the palm. The system employs real-time pathfinding within a gamified scenario designed to maintain user engagement. A user study involving 18 participants evaluated navigation accuracy, task completion time, and feedback preferences comparing three conditions: Visual, Haptic, and Combined feedback. Gamification was used to increase engagement and ecological validity of the VR tasks; the study investigates how feedback modality performs within these gamified scenarios. Results indicate that although visual feedback produced objective performance comparable to the Combined feedback modality, the majority of participants (72
Today, an increasing number of applications in domains such as cultural heritage, healthcare, education, entertainment, and fashion require high-fidelity 3D avatars. However, generating avatars that faithfully reproduce users’ bodies through modeling or acquisition techniques remains challenging and time-consuming, particularly in applications where the accurate quantitative reproduction of body shape and precise anthropometric measurements is required. Thus, attention is shifting towards machine learning-based approaches, in particular those able to fit a parametric model representing the avatar to the intended body shape. Among these models, the Sparse Unified Part-Based Human Representation (SUPR) has been proven to offer superior performance compared to other representations. However, its adoption is primarily hindered by the lack of datasets built upon it. This paper addresses this gap by proposing BOdy shape parameter and 3D meshes of Individuals basEd on SUPR (BODIES), a dataset containing 84,000 synthetic-generated subjects described using the SUPR model with different numbers of parameters. The paper also presents the results of three experimental studies aimed at assessing the improvements brought by the SUPR model over the state-of-the-art when used to feed an existing framework for generating 3D avatar meshes.
Zero-shot learning (ZSL) aims to train deep neural networks to recognize objects from unseen classes, starting from a semantic description of the concepts. Neuro-symbolic (NeSy) integration refers to a class of techniques that incorporate symbolic knowledge representation and reasoning with the learning capabilities of deep neural networks. However, to date, few studies have explored how to leverage NeSy techniques to inject prior knowledge during the training process to boost ZSL capabilities. Here, we present Fuzzy Logic Prototypical Network (FLPN) that formulates the classification task as prototype matching in a visual-semantic embedding space, which is trained by optimizing a NeSy loss. Specifically, FLPN exploits the Logic Tensor Network (LTN) framework to incorporate background knowledge in the form of logical axioms by grounding a first-order logic language as differentiable operations between real tensors. This prior knowledge includes class hierarchies (classes and macroclasses) along with robust high-level inductive biases. The latter allow, for instance, to handle exceptions in class-level attributes and to enforce similarity between images of the same class, preventing premature overfitting to seen classes and improving overall performance. Both class-level and attribute-level prototypes through an attention mechanism specialized for either convolutional-or transformer-based backbones. FLPN achieves state-of-the-art performance on the GZSL benchmarks AWA2 and SUN, matching or exceeding the performance of competing algorithms with minimal computational overhead. The code is available at https://github.com/FrancescoManigrass/FLPN.
In recent years, the digitization of Cultural Heritage (CH) has gained momentum, with Historical Building Information Modeling (HBIM) playing a vital role in developing services that enhance the understanding of built heritage. However, current HBIM-to-VR workflows, crucial for Virtual Heritage Experiences (VHEs), often rely on proprietary tools that may limit interoperability. To address this issue, some studies have used the Industry Foundation Classes (IFC) file format which, unfortunately, lacks functionalities for VHE creation, complicating the process for non-IT users. This paper introduces a semi-automated HBIM-to-VR workflow using an enriched IFC file. The process includes a manual phase for preparing the IFC file with metadata and a subsequent automated phase where a VR platform generates a virtual visit. A use case involving the “Hall of Seasons” in Palazzo Carignano, Turin, demonstrates the workflow’s potential to improve accessibility and scalability in creating VHEs.
Gait recognition identifies individuals based on their unique walking patterns and has traditionally relied on silhouette-based methods for their higher accuracy. However, advancements in 3D Human Pose Estimation (HPE) algorithms have enabled skeleton-based methods to emerge as a promising alternative. In this work, we quantitatively evaluate the performance of state-of-the-art 3D HPE algorithms in a modular skeleton-based gait recognition pipeline. Our system combines 3D pose extraction from RGB video with a contrastive attention-based gait encoder, enabling direct comparisons of HPE models based on their impact on gait recognition accuracy. Experiments on the CASIA-B dataset demonstrate how improvements in pose estimation accuracy translate into recognition performance, providing new insights into the suitability of 3D HPE models for this task. Our findings establish a benchmark for skeleton-based methods and highlight their growing viability for real-world gait recognition applications.
Technological advancements are prompting the digitization of many industries, including fashion. Many brands are exploring ways to enhance customers' experience [e.g., offering new shopping-oriented services like virtual fitting rooms (VFRs)]. However, there are still challenges that prevent customers from effectively using these tools for trying on digital garments. Challenges are associated with difficulties in obtaining high-fidelity reconstructions of body shapes and providing realistic visualizations of animated clothes following real-time customers' movements. This article tackles such lacks by proposing a semiautomated pipeline supporting the creation of VFR experiences by exploiting state-of-the-art techniques for the accurate description and reconstruction of customers' 3-D avatars, motion capture-based animation, as well as realistic garment design and simulation. A user study in which the resulting VFR experience was compared with those created with two existing tools showed the benefits of the devised solution in terms of usability, embodiment, model accuracy, perceived value, adoption, and purchase intention.
The advent of self-driving cars promises to enable occupants to repurpose commuting time. However, although in conditional automation (SAE Level 3) drivers can engage in non-driving related tasks (NDRTs), they must be ready to intervene when prompted by the system with a take-over request (TOR). The vehicle may also need to warn the driver about critical events without notifying a TOR (as in sudden hard braking due to a jaywalker). Clearly and effectively communicating these events and their urgency is crucial for the successful adoption of autonomous vehicles. This work analyzes the impact of multimodal visual and audio cues in conveying this information. It considers an augmented reality (AR) windshield display (WSD) combining screen-fixed elements and world-registered AR overlays, alongside an auditory interface providing explanations and alerts through speech and abstract sounds. The effectiveness of these combined stimuli was evaluated through a user study conducted in a VR-based driving simulator.
The growing availability of affordable Virtual Reality (VR) hardware and the increasing interest in the Metaverse are driving the expansion of Social VR (SVR) platforms. These platforms allow users to embody avatars in immersive social virtual environments, enabling real-time interactions using consumer devices. Beyond merely replicating real-life social dynamics, SVR platforms offer opportunities to surpass real-world constraints by augmenting these interactions. One example of such augmentation is Artificial Facial Mimicry (AFM), which holds significant potential to enhance social experiences. Mimicry, the unconscious imitation of verbal and non-verbal behaviors, has been shown to positively affect human-agent interactions, yet its role in avatar-mediated human-to-human communication remains under-explored. AFM presents various possibilities, such as amplifying emotional expressions, or substituting one emotion for another to better align with the context. Furthermore, AFM can address the limitations of current facial tracking technologies in fully capturing users' emotions. To investigate the potential benefits of AFM in SVR, an automated AM system was developed. This system provides AFM, along with other kinds of head mimicry (nodding and eye contact), and it is compatible with consumer VR devices equipped with facial tracking. This system was deployed within a test-bench immersive SVR application. A between-dyads user study was conducted to assess the potential benefits of AFM for interpersonal communication while maintaining avatar behavioral naturalness, comparing the experiences of pairs of participants communicating with AFM enabled against a baseline condition. Subjective measures revealed that AFM improved interpersonal closeness, aspects of social attraction, interpersonal trust, social presence, and naturalness compared to the baseline condition. These findings demonstrate AFM's positive impact on key aspects of social interaction and highlight its potential applications across various SVR domains.
Accessibility in cultural heritage is essential to foster inclusivity, particularly for individuals with Autism Spectrum Disorder (ASD) who may face sensory and social challenges in museums, such as bright lights, loud noises, and crowded spaces. Virtual Reality (VR) offers a promising solution by providing controlled and customizable experiences that can be tailored to their needs. However, a consolidated methodology for designing VR-based museum experiences for ASD users is still lacking. This paper proposes a structured approach based on three key design principles: modularity, predictability, and attention. These principles serve as the foundation for creating virtual visits specifically suited to the needs of individuals with ASD, aiming to provide accessible, engaging, and enriching experiences inspired by real-world visits. The results of a user study comparing a VR experience based on these principles with a traditional museum visit demonstrated significant advantages in terms of museum experience, with variations depending on ASD levels.