Robotic grasping, once considered to be a solely geometrical problem, has seen a major change with the emergence of deep learning and trial-and-error methods. Combining exhaustive simulation and real-world experiments, this new approach is more efficient at teaching robots how to grasp.