this paper uses the common 2-step procedure to first eliminate most of unlikely detection windows (high recall), then use a network with higher capacity for better discrimination (high precision). Deep learning (in the unsupervised sense) helps having features optimized for each of these 2 different tasks, adapt them for different situations (different robotics grippers) and beat hand-designed features for detection of graspable areas, using a mixture of inputs (depth + rgb + xyz).