Self-Supervised Antipodal Grasp Learning With Fine-Grained Grasp Quality Feedback in Clutter

Abstract

It is a challenging goal in robotics to make a robot grasp like a human being in a cluttered environment. Self-supervised grasp learning is one of the most promising approaches to human-like robotic grasp. However, due to inadequate feedback on grasp quality, almost existing self-supervised grasp learning methods are coarse-grained. This paper presents a fine-grained antipodal grasp learning (FAGL) method with augmented learning feedback. First, an indicator called antipodal degree of a grasp (ADG) is designed as a non-increasing monotonous function for the fine-grained evaluation of grasp quality according to the disturbance incurred by a grasp to the surroundings in the image space. Next, we design a restorative sampling strategy to collect antipodal grasp samples and propose a refined affordance network to generate grasp affordance maps for FAGL's decision of grasp policies. Finally, in grasping actual metal workpieces, FAGL outperforms its peers in grasp success rate and ADG in cluttered scenarios by reducing the grasp effects on the surroundings.