Teachable Robots Learn What to Say: Improving Child Engagement During Teaching Interaction

Teachable robots are a promising technology to promote engagement in the classroom for young students. They are capable of displaying and adapting different social behaviours and characteristics, such as speech, gaze, and vocal pitch, to personalise the teaching interaction, and sustain interest for students over time. Research in this field shows a growing use of reinforcement learning to achieve this adaptation, however there is limited research on adaptive dialog behaviours for teachable robots. Our work proposes an adaptive dialog selection algorithm, implemented using Q-learning, which aims to personalise the dialog choices of a teachable robot in order to optimise for task engagement, measured by the time taken per teaching input, and the amount paraphrasing in the user’s response. We investigate the effect of this approach in a case study with children aged 9–10 years old. The results show that this demographic responds positively to the teaching interaction, and provide useful insights into their preferences and abilities.