Robust Grasp Pose Estimation Based on Point Cloud Uncertainty Modeling

Robust and accurate grasp pose estimation is crucial for robot manipulation. Although current point-cloud-based grasp poses estimation methods demonstrate excellent performance on various datasets, most of the current methods sometimes inevitably fail in real robot grasping tasks due to the unstable point cloud. In this paper, we propose a novel two-stage grasp pose estimation method, which solves this problem by embedding the uncertainty of the point cloud into the grasp pose estimation pipeline. Firstly, our method generates a high-quality point cloud and corresponding probability distributions, which characterize the uncertainty of the point cloud, based on a pair of binocular images. Secondly, the proposed method effectively utilizes the probability distribution of the point cloud to assist the grasp estimation process and thereby generate high-quality grasp poses. Experiments on simulated and real robots prove that our method outperforms other popular methods and verify the effectiveness of each stage of the proposed method.