Education is an art close to theater. A teacher is taking a role; he works his speeches and his gestures and he plays with the attention of his audience. But it is harder: more than entertaining, a teacher must shape the skills, the knowledge and the motivation of his students. This requires, more than just understanding the learning dynamic of students, the talent to control the way he is understood so he can manipulate this learning dynamic. We call it mutual understanding, formalized by the accuracy of the prediction of others and of the prediction of oneself by the others.

Robots for education, a field that emerges from novel approaches involving new technologies, opens a large horizon of unexplored pedagogical activities. Indeed, robots can take roles that were not doable by humans. For example, CoWriter is a robot that personifies a very unskilled beginner so even a child with strong difficulties can teach it handwriting: involving an adult would not be convincing and calling another child would be unethical for this role. However, a strong limitation lies in the fact that robots have a restricted perception to understand humans and are hardly understandable by humans. By consequence, robots for education suffer the poor -- even nonexistent -- level of mutual understanding required by educational interactions.

The first part of this thesis highlights the importance of the human-robot mutual understanding in pedagogical collaborative activities like CoWriter and is based on real-world experimentation. The next two parts form a suggestion to implement such an ability in a robot aiming to interact with humans by focusing on the modelling of motivations. One part regards the external orchestration of the different models built by the robot to make predictions and to be predictable. The other part focuses on the internal mechanisms of these models, based on the computational framework of reinforcement learning.