Open-ended learning of skills in robots: insights from looking at the brain

The capacity to learn an increasing number of ?exible sensorimotor skills in an open-ended autonomous fashion is one of the most remarkable features of human intelligence. How can we design robot controllers that exhibit the same capability? We start from the idea that to accomplish this objective one needs to design a complex architecture formed by multiple components, rather than single algorithms. The decisions at the high-level are critical for the speci?cation of the single components and their interplay, and hence for the overall success of the system. Here we look at the brain architecture and functioning possibly underlying open-ended development in humans and claim that it suggests two possible principles that can be exploited to build open-ended learning robots: (a) The sensorimotor hierarchy underlying motor behaviour should have three speci?c levels of organisation rather than a continuously graded granularity; (b) Intrinsic motivations should guide the formation of speci?c goals, and regulate the skill learning processes pivoting on them, rather than guiding learning processes at ?ne spatial and temporal scales. Here we draw from the biology to support these principles and present our past and current work implementing and testing their utility for open-ended learning robots