Challenging math problems provide rich opportunities to cultivate perseverance to forge ahead in face of ambiguity or difficulty. An ideal teacher closely monitors the problem solving process and provides personalized cognitive, emotional or social supports. Given the potential high cognitive loads on the teacher, an affect sensitive social robot has the potential to assist. In this paper, we will describe a multi-modal dataset we collected from multiple sessions of a young child solving math problems coached by his parent tutor. We report initial findings and their implications in the interaction design of a robotic companion that responds dynamically to the child’s non-verbal cues. We also describe an ongoing study involving multiple parent-child pairs with additional data elements.