Taylor Neural Network for Unseen Object Instance Segmentation in Hierarchical Grasping

Grasping unseen cluttered objects is a challenging task for robots due to the uncertainty in the shape, size, and pose of objects to be grasped. In this article, we propose a novel method for instance segmentation of unseen objects, which is used to generate effective grasps for parallel grippers in cluttered scenes. Focusing on the grasp generation part, we propose a two-stage grasp generator to generate pixelwise grasps. In the first stage, a Taylor neural network (TNN) is used to design an instance segmentation model for segmenting unseen objects that explicitly learns second-order and above feature information from the input. In the second stage, a hierarchical grasp generation algorithm is employed for deducing the occlusion relationships between different instance masks. To reduce potential interference from nearby objects, grasps are estimated under different occlusion situations, and the optimal grasp is selected by combining depth information. We evaluate our method and achieve good segmentation results, outperforming state-of-the-art methods for unseen object instance segmentation. We also demonstrate a grasp success rate of 97.0% in cluttered object scenes.