6-DoF Grasp Planning on Point Cloud for Human-to-Robot Handover Task

Human-to-robot handover is a key task in the human-robot collaboration. However, the diversity in object shapes and poses makes it extremely challenging for robots to ensure safe and successful grasping. To overcome these challenges, we present a method for generating 6-Degree-of-Freedom (6-DoF) grasp poses for human-to-robot handover tasks. The method performs grasp planning on 3D point cloud, which mainly consists of two stages: scene understanding and 6-DoF grasp pose prediction. In the first stage, for safe interaction, the Faster R-CNN model is utilized for accurate detection of hand and object positions. Then, by combining with segmentation and noise filtering, the safe handover points on objects can be extracted. In the second stage, a 6-DoF grasp planning model is built based on GraspNet for handover tasks. The experiments evaluate the developed model by performing grasp planning on several objects with different poses in the handover scenes. The results show that the proposed method not only predicts secure 6-DoF grasp postures for handover tasks, but also be capability to generalize to novel objects with diverse shapes and sizes.