Learning Grasping for Robot with Parallel Gripper from Human Demonstration via Contact Analysis

Recent studies in the field of robotic grasping have predominantly concentrated on generating valid grasps using geometric features of target objects, employing either analytical or deep learning methods. Although such approaches have proved successful in simple picking tasks, they often fall short in task-oriented robotic grasping that demands the grasp pose to be limited to specific parts of the object. In human hand-object interaction, individuals tend to grasp particular parts of an object to facilitate subsequent tasks based on their learned knowledge from daily life experiences. Some previous research has explored the mapping of the human hand pose to a dexterous gripper with similar degrees of freedom (DoF). Nevertheless, the majority of robotic grippers are still parallel jaw grippers, which presents a challenge in mapping high DoF human hand poses to the low DoF grasp pose of a parallel gripper. In this paper, we propose three schemes that map human hand poses to the grasp pose of a parallel gripper. Our quantitative results demonstrate that the optimal mapping scheme can achieve an impressive overall success rate of 87% in robust robotic grasping. Video: https://youtu.be/PXSF6HI5u6k.