Model-Based Reinforcement Learning for Humanoids: A Study on Forming Rewards with the iCub platform