Learning Deep Policies for Robot Bin Picking by Simulating Robust Grasping Sequences