Autonomous selection of the “what” and the “how” of learning: An intrinsically motivated system tested with a two armed robot

In our previous research we focused on the role of Intrinsically motivated learning signals in driving the selection and learning of different skills. This work makes a further step towards more autonomous and versatile robots, implementing a 3-level hierarchical architecture with the mechanisms necessary to both select goals to pursue and search for the best way to achieve them. In particular, we focus on the important problem of providing artificial agents with a decoupled architecture that separates the selection of goals from the selection of resources. To verify our solution, we use the architecture to control the two redundant arms of a simulated iCub robotic platform tested in a reaching task within a 3D environment. We compare its performance to a previous model having a coupled architecture where the different goals are associated at design-time to different modules pursuing them.