Strategic and interactive learning of a hierarchical set of tasks by the Poppy humanoid robot

We present an active learning architecture that allows a robot to actively learn which data collection strategy is most efficient for acquiring motor skills to achieve multiple outcomes, and generalise over its experience to achieve new outcomes for cumulative learning. In the present work, we consider the learning of tasks that are hierarchically organised, interrelated and more and more difficult. This paper proposes an algorithmic architecture, called Socially Guided Intrinsic Motivation with Active Choice of Task and Strategy for Cumulative Learning (SGIM-ACTSCL). It relies on hierarchical active decisions of what and how to learn, driven by empirical evaluation of learning progress for each learning strategy. Our learning agent uses both interactive learning and autonomous goal-babbling. It actively decides at the same time, which tasks to focus on, when to explore autonomously, and when and what to request for social guidance. We present experimental results on the physical humanoid robot Poppy that learns different types of motor skills, encoded by Dynamic Movement Primitives, in order to use a tablet (Fig. 1). We show that SGIM-ACTSCL learns significantly more efficiently than other algorithms. Moreover, it automatically organises its learning process focusing on easy tasks first, and difficult tasks afterwards. It coherently selects the best strategy with respect to the chosen outcome, manages to learn to associate the teacher with his competence domain in order to actively request social guidance for the appropriate tasks.