Optimizing Correlated Graspability Score and Grasp Regression for Better Grasp Prediction

Abstract

Grasping objects is one of the most important abilities that a robot needs to\nmaster in order to interact with its environment. Current state-of-the-art\nmethods rely on deep neural networks trained to jointly predict a graspability\nscore together with a regression of an offset with respect to grasp reference\nparameters. However, these two predictions are performed independently, which\ncan lead to a decrease in the actual graspability score when applying the\npredicted offset. Therefore, in this paper, we extend a state-of-the-art neural\nnetwork with a scorer that evaluates the graspability of a given position, and\nintroduce a novel loss function which correlates regression of grasp parameters\nwith graspability score. We show that this novel architecture improves\nperformance from 82.13% for a state-of-the-art grasp detection network to\n85.74% on Jacquard dataset. When the learned model is transferred onto a real\nrobot, the proposed method correlating graspability and grasp regression\nachieves a 92.4% rate compared to 88.1% for the baseline trained without the\ncorrelation.\n