Robot Grasp in Cluttered Scene Using a Multi-Stage Deep Learning Model

Object grasping in cluttered scene is a practical robotic skill which has a wide range of applications. In this paper, we propose a novel maximum graspness metric which can help extract high-quality scene grasp points effectively. The graspness scores of a single-view point cloud are generated using the proposed interpolation approach. The graspness model is implemented using a compact encoder-decoder model which takes a depth image as input. On the other hand, the grasp point features are extracted. They are further grouped and sampled to predict approaching vectors and in-plane rotations of the grasp poses using residual point blocks. The proposed model is evaluated using a large scale benchmark GraspNet-1Billion dataset and can outperform prior state-of-the-art method by a margin (+4.91 AP) on all camera types. Through real-world cluttered scenario testing, our approach achieves grasping successful rate of 89.60% using a UR-5 robotic arm and a RealSense camera.