In this thesis we focus on designing the planner for MIT s entry in the Amazon Picking Challenge, a robotic competition aiming at pushing the frontiers of manipulation until robots can substitute human pickers in warehouses. Given a set of manipulation primitives (such as grasping, suction, scooping, placing or pushing) we designed a system capable of learning a planner from a set of manipulation experiments. After learning, given any configuration of objects, the planner can come up with the optimal sequence of primitives applied to any object on the scene so as to maximize the probability of successfully picking the goal object. In doing this research we have analyzed Reinforcement Learning, Deep Learning and Planning approaches. For each one, we first describe the background theory, characterizing it for our application to robotics. Then we describe a prototype done in the area and the lessons learned from it. Finally, we combine the strengths of all the areas to create the final design of our system.