Semantic Grasping for Unseen Objects with Post-grasp Stability Consideration

Task-oriented grasping (semantic grasping) is essential for robot arms to use tools in a human-like way. However, even if the estimated grasp position is semantically appropriate, it does not necessarily guarantee sufficient mechanical stability to withstand post-grasp operations. As a result, the subsequent task may fail even when the object has been grasped at a seemingly correct region. In this study, we propose a semantic grasping method that explicitly considers post-grasp stability without requiring a pre-existing 3D model of the target object. The proposed method determines a grasp pose such that the moment of the centroid acting at the grasp point is minimized, thereby maintaining a stable posture of the object after it is lifted. In our experiments, we evaluated the post-grasp stability of lifted objects in comparison with a conventional method. The proposed method maintained a stable object pose in 38 out of 55 trials, whereas the conventional method succeeded in only 20 out of 55 trials, demonstrating the effectiveness of the proposed approach.