Learning synergetic push-shift-grasp policies for efficient robotic grasping in clutter

Purpose The purpose of this paper is to focus on the task of object-agnostic grasping in clutter with the aid of pushing or shifting, where pushing is more suitable for objects with enough operation space and shifting is more effective for restricted objects. Design/methodology/approach A deep reinforcement learning framework is proposed to learn synergetic push-shift-grasp (SPSG) policies from visual observations. The action-values (Q) are modeled by Q maps outputted by three separate branches on the top of a shared backbone and a multi-scale feature fusion module, resulting in finer Q predictions and better computation efficiency. The target network is integrated to reduce overestimation and instability in learning. Moreover, to improve the training sample efficiency, a rewarding strategy and a filter-mask function, both constructed upon the convex hull of objects, are proposed. The former makes sure that only the meaningful non-prehensile actions that disperse cluttered objects are rewarded, whereas the latter allows SPSG to focus action attempts on the object area with smoothed Q values. Findings Experiments in simulation and the real world are both performed. SPSG outperforms baselines in task completion rate and grasp success rate, at challenging test scenarios with tightly packed objects and randomly arranged novel objects, especially when objects located at the corners of a bin where preferential shift actions are required. Originality/value This paper presents a novel SPSG policy learning framework to efficiently learn effective SPSG policies, which outperforms existing methods in task completion rate and grasp success rate.