Images used during training of simulated grasping experience with procedurally generated objects (left) and of real-world experience with a varied collection of everyday physical objects (right). In both cases, we see pairs of image inputs with and without the robot arm present. |