A visual imitation learning algorithm for the selection of robots’ grasping points

Abstract

The short-run production and customization are increasingly common in the manufacturing industry, which results in the frequent adjustments of production lines. Industrial robots in these production lines are also required to quickly learn to perform new grasping tasks. However, the traditional approaches in grasping points selection, which rely on the geometric envelope models of objects in preset task space, can no longer meet the demands of fast-shifting production. Therefore, the present paper tackled the problem of grasping points selection with a novel CNN-based imitating learning framework. Our imitating model learns the correct grasping posture for objects from human grasping operations captured on camera. The experiments showed that this imitating learning algorithm can help a dual-arm robot master the correct grasping posture of an object within only 20 min. Compared to traditional geometric modeling-based methods, such as the pick-and-place module available in the Robot Operating System (ROS), this new approach can increase grasping planning efficiency by 26.1%.