Learning 6-DOF Grasping Interaction via Deep Geometry-aware 3D Representations