Dex-Net MM: Deep Grasping for Surface Decluttering with a Low-Precision Mobile Manipulator

摘要

Surface decluttering in homes and machine shops can be performed with a mobile manipulator that recognizes and grasps objects in the environment to place them into corresponding bins. In contrast to fixed industrial manipulators, mobile robots have low-precision sensors and actuators. In this paper, we modify the Dex-Net 4.0 grasp planner to adapt to the parameters of the mobile manipulator. Experiments on grasping objects with varying shape complexity suggest that the resulting policy, Dex-Net MM, significantly outperforms both Dex-Net 4.0 and a baseline that aligns the grasp axis orthogonally to the principal axis of the object. In a surface decluttering experiment where the objects are randomly selected from 40 common machine shop objects, the robot is able to recognize, grasp and place them into the appropriate class bins 117 out of 135 trials (86.67% including 15 detected grasp failures and recovery on retry).