Category-Level 6-DOF Grasp Pose Estimation by Transferring Grasps from Priors

Recently, grasp transfer has gained popularity for 6-DOF grasp pose estimation. However, most methods are limited to training separately for each category, and network parameters cannot be shared among different categories. To address this issue, we propose a novel 6-DOF grasp pose transfer method that inputs the prior point cloud of each category into the network, so that the network can be trained for multiple categories. We use only one network to achieve grasp transfer for multiple categories while labeling only one prior point cloud per category. Specifically, we design a two-stage deformation and correspondence estimation network to reconstruct the complete point cloud of an instance. We transfer the grasp poses from the prior point cloud to the reconstructed instance point cloud, using the learned dense point-to-point correspondences. Finally, we design a grasp pose refinement and evaluation pipeline to fine-tune the grasps transferred to the instance. Sufficient experiments demonstrate that our method can transfer high-quality grasp poses with higher inference speed.