Modeling of Gaze Behavior in Children Using Deep Neural Network and Robot Implementation

Observing eye movements can unveil people's focus and intentions during social interactions. To ensure appropriate engagement, social robots must direct their gaze accurately and anticipate actions. Our initial study aimed to train deep neural networks to replicate children's gaze patterns in social scenarios. Additionally, we aimed to assess the model's effectiveness when implemented on a Nao robot in real-world settings. We developed two video clips (one animated and one live-action) portraying social situations and employed eye-tracking technology to gather data from 12 children. Subsequently, we used the Transformer Network to analyze and model gaze patterns. Our models achieved 62-70% accuracy in predicting people's locations in the next frame on the first attempt, with a 20% improvement on the second attempt, suggesting predictable patterns in gaze behavior. Feedback was collected from 30 new participants who watched videos of the robot in action and completed a detailed questionnaire to assess the robot's performance. The results indicated participants' satisfaction with the robot's interaction, including its attentiveness, intelligence, and responsiveness. However, participants did not perceive the robot as an equivalent to a human social companion. This exploratory research aims to demonstrate the potential for enhancing robots' social acceptance by incorporating human nonverbal cues, paving the way for future studies in this domain.