A 6-DoF Grasping Network Using Feature Augmentation for Novel Domain Generalization

Liran Zhang, Yunhan Lin, Yan Xu, Huasong Min

Abstract: In this paper, we propose a grasping network using feature augmentation to improve the problem of low generalization in novel object grasping. Three proposed modules, including Gaussian Noise Mix (GNM), Resblock, and Local Features Interpolation (LFI), use GSNet as the baseline. GNM is used for feature augmentation of backbone features during training to reduce the empirical risk of the model when dealing with novel samples. Resblock is designed to expand the prior weights of the features to improve the accuracy and importance of point-wise features for global context. Moreover, LFI is introduced to enhance the local geometric representation of seed point features and reduce the model’s dependence on sample data by inserting hybrid features into seed point features. The effectiveness of our method is verified on the GraspNet-1Billion dataset through a range of experiments, including simulation experiments, ablation experiments, and generalization experiments. Our model is state-of-the-art and achieves a maximum improvement of 13.07% relative to the baseline on the similar set, particularly on the most challenging novel set with a maximum improvement of 4.06%. In practical experiments, similar and novel objects are selected to verify that the novel generalization ability of our network has been significantly improved.