Abstract: Learning deep networks from large simulation datasets is a promising approach for robot grasping, but previous work has so far been limited to the simplified problem of overhead, parallel-jaw grasps. This paper considers learning grasps in the full 6D position and orientation pose space for non-parallel-jaw grippers. We generate a database of millions of simulated successful and unsuccessful grasps for a three-fingered underactuated gripper and thousands of objects, and then learn a modified convolutional neural network (CNN) to predict grasp quality from overhead depth images of novel objects. To generate a valid grasp from the 6D pose space, we introduce a novel optimization-based method that optimizes current suboptimal grasps using the learned grasp quality function.