Understanding and ascribing to others’ intentions and beliefs based on the observed behavior is a key aspect of people’s everyday social lives. This is crucial also in Human-Robot Interaction because both humans and robots need to make sense of each others’ behavior in collaborative settings. Such a complex mechanism is known as the Theory of Mind (ToM), and it still holds secrets, although it has been investigated in HRI for several years.This study focuses on the second-order ToM attributions using an inverted Sally-Anne paradigm, a well-established False Belief task. The humanoid robot Pepper, equipped with vision algorithms, assumes the role of Anne and predicts where Sally (a human researcher) will search for a ball, contingent on Sally’s presence or absence during its relocation by a neutral experimenter. Two scenarios are tested: Sally exits the room (false belief) or observes the relocation (true belief). We had two experimental conditions, where the Pepper robot exhibited passive (monotonic voice, rigid gestures) and active (dynamic voice, fluid gestures) behaviors, respectively. Participants, as external observers, watch video recordings of the interactions and answer structured questions to assess how behavioral cues influence robots’ ToM attributions.Results showed that people tended to ascribe high-level ToM skills to the active robot rather than to the passive one, highlighting the importance of designing robots with appropriate expressive behaviors. By examining how humans interpret a robot’s capacity for second-order ToM, this work advances our understanding of the cognitive assumptions people make about artificial agents and offers a foundation for developing socially intelligent systems that can seamlessly integrate into collaborative environments.

The application of artificial intelligence (AI) in medical imaging has significantly improved diagnostic accuracy. However, the reliance on black-box models remains a barrier to its widespread adoption in clinical settings. This article compares the main techniques of explainable AI (XAI), such as Grad-CAM, LIME, and SHAP, evaluating their effectiveness in interpreting deep learning models used in medical imaging. Two case studies are analyzed, comparing the three methods and highlighting their strengths and weaknesses. The results of this analysis show that Grad-CAM provides intuitive visualizations; LIME offers excellent flexibility in application; and SHAP delivers complete and accurate explanations, despite its high computational load.

In recent years, numerous studies on emotion recognition have employed non-invasive electroencephalography (EEG) to measure neuronal activity. Various Machine Learning and Deep Learning techniques have been applied to classify emotions based on EEG signals. However, Graph Neural Networks (GNNs) remain relatively unexplored in this domain. Given that GNNs can accommodate dynamic and variablesized data as input, we hypothesize their potential for superior performance compared to traditional models like Support Vector Machines (SVMs) or Convolutional Neural Networks (CNNs), which rigidly fix the distance among electrodes in a pixel-like matrix. Furthermore, GNNs are designed to leverage the biological topology between different brain regions, capturing both local and global relationships among EEG channels. In this study, two GNN models are experimented with: a Dynamical Graph Convolutional Neural Network (DGCNN) and a Regularized Graph Neural Network (RGNN). Results indicate that the DGCNN model proposed in this work outperforms state-of-the-art models and the RGNN, achieving an average accuracy of 95%, compared to 53% for the latter. Additionally, the DGCNN exhibits significantly faster training times, opening up new avenues for research in this field.

In this work, we are interested in investigating if a distinct personality of the robot may impact the emotional state of the users, which we propose to detect using neuroscience theories that allow us to classify emotions based on valence and arousal metrics derived from brain wave activity analysis. We devised an experimental research study in which EEG data was gathered while individuals interacted with a robot with different personalities. Support Vector Machine, Decision Tree, Random Forest, K-Nearest Neighbors, and Multi-Layer Perceptrons have all been trained using EEG-signal, valence, and arousal data. All proposed classifiers were subjected to a Global optimization Model (GOM) that used feature selection and hyper-parameter optimization techniques to improve classification results and address common issues that affect classifier accuracy when attempting to solve a supervised learning problem, such as bias-variance trade-off, dimensionality of the input space, and noise in the input data space. The findings of the experiments will be presented and debated.

Deep learning techniques have led to a vast improve-ment in various fields of computer vision, mainly using large-scale labelled datasets. Obtaining a large dataset of medical images for diagnostics is still an open challenge; the most significant obstacles are data imbalances and privacy issues related to sensitive patient information. Furthermore, the limited size of datasets for the training of a neural network can affect the performance of supervised learning and cause model overfitting problems. For this reason, data augmentation techniques are used to expand existing datasets. Generative Adversarial Networks (GANs) represent an innovative solution for acquiring additional information from a dataset since they can generate synthetic samples indistinguishable from real sample images. This work explores the use of GAN networks on Digital Breast Tomosynthe-sis (DBT) images, which is, in our knowledge, a completely new approach in this domain. In particular, we apply an optimization approach to the learning process of GAN networks based on evolutionary techniques.

During the last years, many solutions have been proposed to achieve a natural Human-Robot Interaction (HRI) and Communication paving the way to new paradigms of under-standing and adaptation based on mutual affective perception. Especially in human-robot social interaction, it is helpful not only that people can understand the robot’s behavioral state, but also robots possess the ability to detect, interpret and adaptively react to human affective responses. Typical approaches are able to assess humans’ affective responses from the observation of overt behavior. However, there are cases in which the overt observable behaviors could not match with the internal states (e.g., people with diseases compromising normal emotional responses). In such cases, having an objective measure of the users’ state from ‘inside’ is of paramount importance. This work presents an affect detection model able to provide a measure of the human affective state, with particular focus on the stress state, from the analysis of EEG users’ activity during the interaction with a social humanoid robot endowed with diverse affective elicitation behaviors. We argue that monitoring the stress state of a human during HRI is necessary to adapt the robot behavior in a way to avoid possible counterproductive effects of its use.

Several studies focused on single human activity recognition, while the classification of group activities is still under-investigated. In this paper, we present an approach for classifying the activity performed by a group of people during daily life tasks at work. We address the problem in a hierarchical way by first examining individual person actions, reconstructed from data coming from wearable and ambient sensors. We then observe if common temporal/spatial dynamics exist at the level of group activity. We deployed a Multimodal Deep Learning Network, where the term multimodal is not intended to separately elaborate the considered different input modalities, but refers to the possibility of extracting activity-related features for each group member, and then merge them through shared levels. We evaluated the proposed approach in a laboratory environment, where the employees are monitored during their normal activities. The experimental results demonstrate the effectiveness of the proposed model with respect to an SVM benchmark.

Robotic psychological assessment is a novel field of research that explores social robots as psychometric tools for providing quick and reliable screening exams. In this study, we involved elderly participants to compare the prototype of a robotic cognitive test with a traditional paper-and-pencil psychometric tool. Moreover, we explored the influence of personality factors and technology acceptance on the testing. Results demonstrate the validity of the robotic assessment conducted under professional supervision. Additionally, results show the positive influence of Openness to experience on the interaction with robot's interfaces, and that some factors influencing technology acceptance, such as Anxiety, Trust, and Intention to use, correlate with the performance in the psychometric tests. Technical feasibility and user acceptance of the robotic platform are also discussed.

Automatic movement-prothesis control aims to increase the quality of life for patients with diseases causing temporary or permanent paralysis or, in the worst case, the lost of limbs. This technology requires the interaction between the user and the device through a control interface that detects the user's movement intention. Basing on the Motor-Imagery theory, many researchers have explored a wide variety of Classifiers to identify patients' physiological signals from many different sources in order to detect patients' moves intentions. We here propose a novel approach relying on the use of a Weightless Neural Network-based classifier, whose design lends itself to an easy hardware implementation. Additionally, we employ a non-invasive light weight and easy donning EEG-helmet in order to provide a portable controller interface. The developed interface is connected to a robotic hand for controlling open/close actions. We compared the proposed classifier with state of the art classifiers by showing that the proposed method achieves similar performance and contemporaneously represents a viable and practicable solution due to its portability on hardware devices, which will permit its direct implementation on the helmet board.

In recent years, humanoid robots have become quite ubiquitous finding wide applicability in many different fields, spanning from education to entertainment and assistance. They can be considered as more complex cyber-physical systems (CPS) and, as such, they are exposed to the same vulnerabilities. This can be very dangerous for people acting that close with these robots, since attackers by exploiting their vulnerabilities, can not only violate people's privacy, but, more importantly, they can command the robot behavior causing them bodily harm, thus leading to devastating consequences. In this paper, we propose a solution not yet investigated in this field, which relies on the use of secure enclaves, which in our opinion could represent a valuable solution for coping with most of the possible attacks, while suggesting developers to adopt such a precaution during the robot design phase.

During the last decade, we witnessed a constantly increasing digitalization in the health-care domain that, while from the one hand, has increased the average life expectancy representing one of the crowning achievements of the last years, from the other hand, has introduced extra challenges due to the simultaneous increasing of the proliferation of cyber-crime and the creation of malicious applications which try to access health sensitive data. This created the need for increased security implementations, leading to improved user acceptance of such applications and thus to large-scale adoption of these technologies and to full exploitation of their advantages. We here propose the use of a SIEM-based framework specifically tailored for a healthcare portal developed within the context of the Italian National Project eHealthNet, which allows real time monitoring of portal accesses with the aim of detecting potential threats and anomalies that could cause major security issues.

A robotic system that is designed to coexist with humans has to adapt its behavioral and social interaction parameters not only with respect to the task it is supposed to accomplish, but also with respect to the human being it is interacting with by profiling her habits, preferences, and personality. This is particularly relevant in the domain of assistive robotics where the behavioral adaptability has been shown to enhance the users' acceptability of a robot. In this work, we propose a neuro-fuzzy-Bayesian system able to adapt the robot proxemics behavior with respect to the human users' personality and the action she is currently performing. The user's personality is evaluated according to the Big-Five factors model and the activity recognition is obtained by classifying data from a wearable device through the use of a Bayesian Network classifier. As shown by a statistical study, the proposed framework is capable of computing the most appropriate robot proxemics behavior in order to improve human feeling in interacting with artificial agents, such as robots.

In many real life scenarios, which span from domestic interactions to industrial manufacturing processes, the objects to be manipulated are non-rigid and deformable, hence, both the location of the object and its deformation have to be tracked. Different methodologies have been applied in literature, using different sensors and techniques for addressing this problem. The main contribution of this paper is to propose a Weightless Neural Network approach for non-rigid deformable object tracking. The proposed approach allows deploying an on-line training on the shape features of the object, to adapt in real-time to changes, and to partially cope with occlusions. Moreover, the use of parallel classifiers trained on the same set of images allows tracking the movements of the objects. In this work, we evaluate the filtering/segmentation performance that is a fundamental step for the correct operation of our approach, in the scenario of pizza making.

People detection and tracking are essential capabilities in human-robot interaction (HRI). Typically, a tracker performance is evaluated by measuring objective data, such as the tracking error. However, in HRI applications, human- tracking performance does not have to be evaluated by considering it as a passive sensing behavior, but as an active sensing process, where both the robot and the human are involved within-the-loop. In this context, we foresee that the robotic non-verbal feedback, such as the head movement, plays an important role in improving the system tracking performance, as well as in reducing the human effort in the interactive tracking process. In order to verify this assumption, we evaluate a tracker performance in a joint task between a human and a robot, modeled as a game, and in three different settings. We adopt common HRI performance measures, such as the robot attention demand or the human effort, to evaluate the HRI human tracking performance scaling up with respect to the used robot feedback channels.

The aim of the proposed investigation is to provide the users with the capability of creating robot teams “on- the-fly” using grouping strategies expressed through speech. Our working hypothesis is that people are inclined to assemble objects into macro-entities, or groups, according to perceptual principles. We observed the real linguistic utterances used by individuals in a testing environment, showing that the type of robots and their mutual arrangements can affect both the choice of elements to form a team, and the way such choice is made. Moreover, we provide an initial insight for the capabilities needed by a robot for reasoning about its membership in a team.

We propose a framework where the human-robot interaction is modeled as a multimodal dialogue which is regulated by an attentional system that guides the system towards the execution of structured tasks. We introduce a simple case study to illustrate the system at work in different conditions considering top-down regulations and dialogue flows in synergic and conflicting situations.

In this work, we present a reliable and continuous gesture recognition method that supports a natural and flexible interaction between the human and the robot. The aim is to provide a system that can be trained online with few samples and can cope with intra user variability during the gesture execution. The proposed approach relies on the generation of an ad-hoc Hidden Markov Model (HMM) for each gesture exploiting a direct estimation of the parameters. Each model represents the best prototype candidate from the associated gesture training set. The generated models are then employed within a continuous recognition process that provides the probability of each gesture at each step. The proposed method is evaluated in two case studies: a hand-performed letters recognizer and a natural gesture recognizer. Finally, we show the overall system at work in a simple human-robot interaction scenario.

We propose a framework where the human-robot interaction is modeled as a multimodal dialogue which is regulated by an attentional system that guides the robot towards the execution of structured tasks. Specifically, we propose an approach where the dialogue between the human and the robot is represented as a Partially Obervable Markov Decision Process (POMDP), while the associated dialogue policy is enhanced by top-down attentional mechanisms that provide contextual and task-related contents. We introduce simple case studies that illustrate the system at work in different conditions considering top-down regulations and dialogue flows in synergistic and conflicting situations.

We present a robotic control system endowed with attentional mechanisms suitable for balancing the trade off between safe human-robot interaction and effective task execution. These mechanisms allow the robot to increase or decrease the degree of attention toward relevant activities modulating the frequency of the monitoring rate and the speed associated to the robot movements. In this framework, we consider pick-and-place and give-and-receive attentional behaviors.

In this paper, we propose a human-robot interaction system that exploits emotion and attention to regulate and adapt the robotic interactive behavior. In particular, we will focus on the relation between arousal, predictability, and attentional allocation considering as a case study a robotic manipulator interacting with a human operator. We rely on a frequency based model of attention allocation and a 4-dimensional model of emotion. The experiment reported in this paper explores the effectiveness of an attentional regulation mechanisms modulated by arousal and predictability values extracted from the human voice. The collected results show that the attentional modulation, mediated by basic emotional speech features, provides a natural and computationally light regulation mechanism for coordinating the robotic behaviors.