The automatic monitoring and assessment of the engagement level of learners in distance education may help in understanding problems and providing personalized support during the learning process. This article presents a research aiming to investigate how student engagement level can be assessed from facial behavior and proposes a model based on Long Short-Term Memory (LSTM) networks to predict the level of engagement from facial action units, gaze, and head poses. The dataset used to learn the model is the one of the EmotiW 2019 challenge datasets. In order to test its performance in learning contexts, an experiment, involving students attending an online lecture, was performed. The aim of the study was to compare the self-evaluation of the engagement perceived by the students with the one assessed by the model. During the experiment we collected videos of students behavior and, at the end of each session, we asked students to answer a questionnaire for assessing their perceived engagement. Then, the collected videos were analyzed automatically with a software that implements the model and provides an interface for the visual analysis of the model outcome. Results show that, globally, engagement prediction from students’ facial behavior was weakly correlated to their subjective answers. However, when considering only the emotional dimension of engagement, this correlation is stronger and the analysis of facial action units and head pose (facial movements) are positively correlated with it, while there is an inverse correlation with the gaze, meaning that the more the student’s feels engaged the less are the gaze movements.

In this paper, we investigate the use of a social robot as an engaging interface of a serious game intended to make children more aware and well disposed towards waste recycle. The game has been designed as a competition between the robot Pepper and a child. During the game, the robot simultaneously challenges and teaches the child how to recycle waste materials. To endow the robot with the capability to play as a game opponent in a real-world context, it is equipped with an image recognition module based on a Convolutional Neural Network to detect and classify the waste material as a child would do, i.e. by simply looking at it. A formal experiment involving 51 primary school students is carried out to evaluate the effectiveness of the game in terms of different factors such as the interaction with the robot, the users’ cognitive and affective dimensions towards ecological sustainability, and the propensity to recycle. The obtained results are encouraging and draw promising scenarios for educational robotics in changing children’s attitudes toward recycling. Indeed Pepper turns out to be positively evaluated by children as a trustful and believable companion and this allows children to be concentrated on the “memorization” task during the game. Moreover, the use of real objects as waste items during the game turns out to be a successful approach not only for perceived learning effectiveness but also for the children’s engagement.

Smart home environments should proactively support users in their activities, anticipating their needs according to their preferences. Understanding what the user is doing in the environment is important for adapting the environment's behavior, as well as for identifying situations that could be problematic for the user. Enabling the environment to exploit models of the user's most common behaviors is an important step toward this objective. In particular, models of the daily routines of a user can be exploited not only for predicting his/her needs, but also for comparing the actual situation at a given moment with the expected one, in order to detect anomalies in his/her behavior. While manually setting up process models in business and factory environments may be cost-effective, building models of the processes involved in people's everyday life is infeasible. This fact fully justifies the interest of the Ambient Intelligence community in automatically learning such models from examples of actual behavior. Incremental adaptation of the models and the ability to express/learn complex conditions on the involved tasks are also desirable. This article describes how process mining can be used for learning users’ daily routines from a dataset of annotated sensor data. The solution that we propose relies on a First-Order Logic learning approach. Indeed, First-Order Logic provides a single, comprehensive and powerful framework for supporting all the previously mentioned features. Our experiments, performed both on a proprietary toy dataset and on publicly available real-world ones, indicate that this approach is efficient and effective for learning and modeling daily routines in Smart Home Environments.