RGB-D Instance Segmentation-based Suction Point Detection for Grasping.

Suction is a common way for robots to pick objects in industry, which has higher stability and reliability. However, it is necessary to evaluate the suction position on the object's surface when dealing with objects of different shapes. The previous method decouples it into a two-stage suction point detection and objects classification process, making the task more complex and time-consuming. On the other hand, there are significant challenges in acquiring training data. To address that, we propose an object instance segmentation-based suction grasping point detection method. Instance segmentation, as an object pixel-level perception method, can classify objects and find the orientation of objects in the plane, which is more efficient than simple classification. Specifically, our method generate a suction heatmap and object instance segmentation from RGB-D inputs and find the suction grasping position of each object after candidate points screening. To generate the training data, a heatmap mask based on 2D Gaussian-encoded is used to annotate the suction grasping ground truth. We evaluate each object's predicted suction grasping point results and conduct grasping experiments in real environments. The suction grasping success rate of 94% in dense scenes demonstrates the remarkable performance of the proposed method.