Open-Ended Learning of Grasp Strategies using Intrinsically Motivated Self-Supervision∗