Depth-Noise-Robust 5-DOF Robotic Grasping of Metallic Objects Using Synthetic Depth Data

Metallic objects with specular surfaces induce severe optical noise in depth sensors, degrading the reliability of vision-based robotic grasping. This paper proposes a five-degree-of-freedom (5-DOF) grasping framework that is robust to such optical noise, using synthetic depth data generated by a physical sensor noise model. Using the SAPIEN simulator, depth images are rendered through an active stereo sensor model combined with the Landau et al. noise model to reproduce reflection noise on metallic surfaces. A grasp-quality convolutional neural network is trained on this dataset, and the conventional top-down grasp is extended to 5-DOF by adding tilt about the grasp axis based on wrench space invariance. Real-world experiments on ten objects at three tilt angles demonstrate that the proposed method achieves an overall grasp success rate of 90.4%. Additionally, it reliably grasps wall-adjacent objects inaccessible to vertical approaches.