SS-ARGNet: A Novel Cascaded Schema for Robots’ 7-DoF Grasping in Adjacent and Stacked Object Scenarios

Abstract

Multiple objects in tightly adjacent and stacked configurations pose significant challenges for robotic systems in achieving reliable grasping, as existing algorithms often struggle to distinguish stacked objects and are prone to causing disruptions to the original scene due to improper grasping postures and collisions. In order to address these challenges, we developed a category-agnostic segmentation and cascaded 7-degrees of freedom (DoF) pose prediction approach for adjacent and stacked objects grasping, using a single vision image. Specifically, a stacked segmentation network (SS-Net) was tailored based on transformer and region proposal modules to achieve robust mask prediction, thereby accurately localizing candidates within the scene. Simultaneously, the attention residual grasping network (ARG-Net) was proposed to estimate the 7-DoF pose of individual targets, employing a new collision-free strategy to avoid interference between the gripper and the candidates. The integrated SS-Net and ARG-Net (SS-ARGNet) schema significantly enhances robotic performance in practical applications, achieving grasp completion rates of 92.8% and 89.1% for adjacent scenarios, and 87.4% and 84.9% for stacked scenarios, for similar and unknown objects, respectively, with a grasp response time of less than 0.9 s.