Invited talk: Finding the right Balance of Modelling and Learning for Vision-based Robot Manipulation