Intrinsically Motivated Discovered Outcomes Boost User's Goals Achievement in a Humanoid Robot

Intrinsic motivations have been successfully employed in machine learning and robotics to improve the autonomous acquisition of knowledge and skills. While forming an ample repertoire of skills is considered advantageous for future tasks accomplishment, few works have focused on how to do this in particular. Here we present a system that first discovers new outcomes and new motor skills with intrinsic motivations, and then exploits goal-based mechanisms to accomplish human assigned extrinsic goals. The approach is tested with an iCub robot learning to displace a ball on a table with a tool.