Simulating Complete Points Representations for Single-View 6-DoF Grasp Detection
Zhixuan Liu, Zibo Chen, Wei‐Shi Zheng
Abstract
The utilization of single-view depth cameras for image capture is prevalent in the context of robotic grasping. However, the incomplete point cloud structure generated from a single-view depth image poses challenges in accurately predicting grasping configurations, due to the lack of some important shapes or contours of objects. In this work, we present a novel approach that enhances the performance of the single-view point cloud 6-DoF grasp detector by leveraging the representations from another complete point cloud pre-trained grasp detector. In contrast to the previous two-stage approaches involving shape completion or multi-view capturing and subsequent grasp detection, our proposed method only utilizes complete point cloud data during the training phase and directly detects grasps with single-view point cloud data for inference. Our approach introduces an effective knowledge distillation framework for grasp detection that transfers the 3D complete representations from three levels: grasp point features, geometric structural relations, and grasp responses. Experimental results on the GraspNet-1Billion dataset demonstrate that our approach outperforms other state-of-the-art single-view grasp detection methods. Furthermore, the results from physical experiments validate the efficacy of our method on real-world unseen objects.