Study and Developing a Deep-Learning model to customize Pepper's behaviors based on Facial Age and Gender estimation

The ability to personalize behaviors is essential when interacting with Socially Assistive Robots (SARs). SARs are robots designed to assist humans, and their interactions must be person-oriented. In this thesis, the behavior of the Socially Assistive Robot (SAR) Pepper is customized according to a person’s age group and gender. These two variables are not known and are predicted during a Human-Robot Interaction (HRI) experiment with the SAR Pepper. Our hypothesis is that age and gender predictions have an impact on the robot’s perception and trust if they are communicated to the participant. This thesis has a twofold aim. The first purpose is to develop a Deep-Learning model to predict age and gender based on low-resolution images in real-time interaction. In this context, real-time means predicting during the HRI. The second purpose is to validate the robot’s perception and trust, and the model accuracy, in an experiment of HRI with a personalized dialogue based on age and gender estimations. The dialogue is offered in two modes: explicit and implicit. In the explicit, the prediction of age group and gender is directly communicated to the person, while in the implicit, a custom dialogue is proposed according to the age group and gender. Regarding the first aim, the final Deep-Learning model is a Convolutional Neural Network (CNN) model trained from a dataset available in the literature. The dataset has been manipulated to be consistent with the style of the Pepper camera and Stylebased Age Manipulation (SAM) proposed by Alaluf et al. and JoJoGAN proposed by Chong et al. approaches are used. The best prediction model achieves an accuracy on the age group of at maximum 67.74%, on the gender of 95.16%, and combining age group and gender of 66.13%. Mean Absolute Error (MAE) on age prediction of the best model is of 4.6 years. Regarding the second aim, when the age and gender are correctly estimated, Pepper’s social behavior becomes more trustworthy and likeable. Otherwise, in the explicit mode, if the age was overestimated or the gender was wrong, trust and general perception have a decreasing trend.