SUGrasping: a semantic grasping framework based on multi-head 3D U-Net

Object grasping is an important skill for robots to interact with the real world, especially in unstructured environments where occlusions and different shapes of target objects are present. In this work, we introduce a robot grasping pipeline called SUGrasping, which can obtain the grasping poses more precisely for target objects. The grasping pipeline treats the Truncated Signed Distance Function (TSDF) and point clouds of the grasping scene as input simultaneously. The proposed multi-head 3D U-Net accepts reconstructed TSDF representation and outputs the grasping configurations, including predicted grasp quality, orientation and width of the gripper. The point cloud is fed into PointNet to obtain the semantic segmentation results for all objects in the grasping workspace. With the help of point cloud inside the gripper, the relationship between the gripper and semantic information can be established. It makes robots know which object they are grasping, rather than just removing objects in the workspace like previous works. Experimental results show that the proposed method has an improvement in grasping success rate and percent cleared of target objects, which outperforms state-of-the-art methods compared in this paper.