Prior-Backbone-Prediction cascade depth completion framework for robotic grasp oriented to transparent objects

Abstract

Transparent objects are frequent manipulation objects for industrial production and manufacturing systems. The texture-less characteristics of transparent objects lead to insignificant RGB image features and incomplete and inaccurate Depth images, which significantly challenge robotic grasping detection and potentially dangerous for independent and safe manipulation in situations such as industrial production and manufacturing system. Reconstruction of depth data by fusing RGB-Depth (RGB-D) data through depth completion is the mainstream solution as a pre-processing robot-acquired data. Therefore, the quality, as well as the efficiency of depth completion, determines the effectiveness of the robot grasping transparent objects. However, the no-prior processing used in existing studies leads to an unimpressive depth completion. In addition, existing studies employ one-way prediction results and do not address the issue of image distortion. In general, current studies are not ideal in balancing accuracy and efficiency. To this end, we propose a Prior-Backbone-Prediction (PBP) cascade lightweight depth completion framework, named PBP framework. Prior: the proposed Reverse Texture Prior Guidance strategy guides the backbone network to focus on texture-less regions better. Backbone: the proposed Pyramid Interactive Residual Backbone network is used for in-depth feature learning. Prediction: the proposed consistency-difference Joint Prediction strategy maximizes the accuracy of prediction results and reduces distortion. The dual validation is performed on both public datasets and robotic platforms. PBP achieves state-of-the-art performance on the well-known TransCG dataset with a model size of only 17.2 M. The success rate of the Aubo i5 robot platform is improved by 44 % based on GG-CNN to grasp transparent objects.