Learning efficient push and grasp policy in a totebox from simulation. (2nd July 2020)

• Record Type: Journal Article
• Authors: Ni, Peiyuan; Zhang, Wenguang; Zhang, Haoruo; Cao, Qixin
• Is Part Of: Advanced robotics. Volume 34:Number 13(2020)
• Page Start: 873
• Page End: 887
• Publication Date: 2020-07-02
• Subjects: Robot grasping -- deep reinforcement learning -- robot manipulation -- robot learning; robot vision and monitoring -- manipulation and grasping; Robotics -- Periodicals; Robotics -- Japan -- Periodicals; Robotics; Japan; Periodicals; 629.89205
• DOI: 10.1080/01691864.2020.1757504
• Languages: English
• ISSNs: 0169-1864

Abstract

Abstract : Usually, grasping in a totebox always encounters bottlenecks when the object is at the edge or even at the corner of the totebox. Meanwhile, if the objects are stacked in a pile, there may be no grasps to be selected. In this paper, an algorithm based on deep reinforcement learning is applied to combine grasping with pushing to deal with these cases. In order to make sure that a push must increase grasp access, we propose to apply the changes of grasp's quality function Q g combined with forgetting mechanism to promote a pushing action. Moreover, a double experience replay is set up to increase the search on the boundaries. To make a balance between efficiency and robustness, the traditional policy π ( s ) = argmax { Q p, Q g } is improved using acceptance thresholds Q g ∗ and Q p ∗ with 99% precision. Our algorithm is trained in a simulation environment using YCB object dataset and finally is transferred into a real-world environment. In our experiment, our algorithm achieves the best results both in simulation and real world (with 86.67% completion for YCB objects and 83.37% completion for novel objects) compared to other famous works. GRAPHICAL ABSTRACT: UF0001