Learning to Manipulate Unknown Objects in Clutter by Reinforcement