Sampling-based Grasp Pose Estimation for Robotic Doorknob Grasping

In this paper, we propose a sampling-based grasp pose estimation algorithm. Our method generates multiple candidate grasps by sampling local surface normal vectors as approach vectors on the point cloud. It employs a grasp score prediction network that fuses point cloud features with grasp pose features via cross-attention. The grasp pose with the highest predicted score is selected for execution. The dataset includes both successful and unsuccessful grasp samples across various doorknobs, with each sample represented by a grasp pose on the point cloud. Experimental results demonstrate superior performance compared to an existing point cloud-based classification model for rigid body grasping. Additionally, real-world grasping evaluations confirm the ability to determine appropriate positions and orientations for doorknob grasping.