Personalization of Child-Robot Interaction Through Reinforcement Learning and User Classification

Social robots offer promising avenues for personalized interactions, particularly in aiding children undergoing minimally invasive surgery who often experience pain, fear, and anxiety. While distraction methods like cartoons have shown an effect, they are not adaptive and lack personalization to each child’s needs. We propose an approach that combines reinforcement learning ( for learning a set of baseline policies for different types of users) with user modeling and classification to create personalized and adaptive interactions for social robots with the aim to provide higher engagement and adequate distraction from pain
in children. In the proposed approach, first a fixed policy is employed
during an assessment phase, collecting data on child-robot interactions
for a new user. Next, this data is compared to a set of user models,
in order to classify the new user into one of these models and its corresponding policy. The selected baseline policy is used during the next interaction which should take place post-surgery. We conducted experiments to test this approach with simulated user models and our results show that baseline policies perform best with their corresponding user model but also achieve good results for unseen models of users who will interact similarly within the interaction framework. Finally, we discuss how these results can inform future research and how they can be used for real-world implementations