Abstract: In warehouse logistics, the bin-picking of scattered objects is a key procedure that determines system efficiency, whereby robotic grasping is widely used due to its high efficiency and reaching ability. To increase the success rate of picking and the label acquisition efficiency, we propose a two-stage object-suction grasp affordance detection network and its self-supervised learning method. A region estimation network fuses the depth and RGB(red, green, blue) features to detect regions of interest independently of the types and the geometrical features of objects. In the selected regions of interest, a suction-grasp-point affordance network also fuses the depth and RGB features to determine pickable points. Using the proposed self-supervised learning method, data and labels are sampled and generated automatically, so that the two-stage object-suction affordance detection method achieves a high picking accuracy in a short period of time. The experimental results validate the effectiveness of the proposed two-stage detection method and its self-learning method.