Learning object affordances by imitation Research Progress Report 3 Marek

The report explores the role of imitation in learning object affordances. Since I intend to experiment with a robotic arm enabling only pushing objects, properties of objects will be the main factor affecting the overall complexity of this “toy world”. Learning in the toy world is a process of discovering objects’ affordances, as well as the kinds of actions and goals related to them. I intend to create an algorithm, where imitation can be thought of as a teacher-guided search, which will help a robot to constrain its own exploration to relevant “behavioural patterns”, accentuating either action plans or action goals. In effect, a robot which begins with very little knowledge about actions and objects, will be able not only to recognise and understand novel object affordances and goals, but also to learn how to generalise the acquired knowledge. Thus, a robot will demonstrate a high level of manipulation competence playing with entirely new objects “it will master the toy world”.