Confidence-Aware Depth Completion for Robust Grasping of Transparent Objects

Robustly grasping transparent objects remains challenging due to the noise and incompleteness of raw depth measurements caused by light refraction and reflection, often leading to failures in point-cloud-based 6-DoF grasp detection. Although depth completion can restore missing shapes, deterministic models lack the ability to self-assess the reliability of their predictions, which often leads to grasping failures caused by inaccurate reconstruction results. In this paper, we propose a confidence-aware framework that explicitly models perceptual uncertainty to bridge the gap between depth completion and robust robotic manipulation. We develop CAC-DepthNet, a multi-head network that concurrently yields dense depth maps and pixel-wise confidence maps. To enhance executive robustness, a dual-stage guidance strategy is introduced: firstly, unreliable points are filtered via confidence-based filtering during cloud generation; subsequently, grasp candidates are reweighted to prioritize those grounded on reliable surface geometries. Extensive evaluations on the TransCG dataset and real-world robot experiments demonstrate that our method achieves competitive completion accuracy and significantly improves the Grasp Success Rate in complex scenes. Our results validate that incorporating confidence information effectively addresses the depth inaccuracies, providing a robust solution for the automated handling of transparent objects.