TCG: Transparent Object Grasping Method Based on Depth Completion in Dense Clutter

We present a formidable challenge of ‘grasping tightly stacked transparent objects’. The methods relying on visual information face exceptional difficulties in generating effective grasping actions in this scenario. We introduce a comprehensive model, TCG, designed to tackle this obstacle, comprising a depth completion module, a semantic segmentation module, and a policy generation module. The depth completion module fills in missing depth information, the semantic segmentation module extracts target information, and the policy generation module selects the optimal actions for execution. For the training and evaluation of our model, we generated a basic dataset that simulates the biomedical manufacturing scenario. Significant trials substantiate the superior performance of our approach over cutting-edge grasping techniques in situations featuring closely stacked transparent objects, with a task completion rate exceeding by 9.7% (86.3% vs. 76.6%), a grasping success rate exceeding by 8.7% (80.5% vs. 71.8%), and an action execution efficiency exceeding by 25.9% (62.5% vs. 36.6%). Supplementary material can be accessed on: https://github.com/Gchaojian/TCG.