UA-Depth: Uncertainty-Aware Depth Restoration for Material-Agnostic Object Manipulation

This research presents UA-Depth, a framework that iteratively co-refines depth maps and pixel-wise uncertainty to address failures of commodity RGB-D cameras on non-Lambertian objects such as transparent, reflective, and translucent materials. By leveraging a frozen depth foundation model as a prior and using uncertainty-gated features to suppress unreliable cues, the method achieves superior performance on an eight-dataset benchmark dominated by challenging materials. The authors report downstream improvements in category-level 6D pose estimation and 6D grasp synthesis. Code, models, and data are scheduled for release upon publication.