Teachable Robots for Education: Bridging Children Learning and Cognitive Modelling

This thesis investigates how social robots can serve as peers within the learning-byteaching paradigm. Taking inspiration from developmental psychology, it validates a cognitive peer-model architecture grounded in Interactive Reinforcement Learning, where children teach a robot through evaluative feedback and observe its progress. The model was evaluated in a large-scale study with primary school pupils. To maximise ecological validity, the study preserved the classroom environment and enabled multiple one-on-one child–robot interactions to occur in parallel. Findings showed that children engaged more with the robot than with independent practice, which in some cases translated into higher retention of the taught material. Because human–robot interaction extends beyond cognitive outcomes, the thesis also examines psychological and social dimensions of teachable robots. A controlled study explored how varying levels of involvement in teaching influenced perceptions of trust, ownership, and anthropomorphism. Results indicated that while deeper involvement increased self-investment, positive perceptions depended strongly on the robot’s transparency and behaviour. Finally, the thesis proposes two new frameworks that align robot learning more closely with natural human teaching. ECLAIR enables robots to interpret the diverse verbal feedback children naturally provide, while INFORM, inspired by memory consolidation, allows robots to replay experiences offline to recover structure and generalise beyond immediate tasks. Taken together, these contributions present a path toward robots that learn as peers: responsive, imperfect, and capable of growing alongside children. The thesis argues that making robots adaptive learners also improves the learning experience of the children who teach them.