Do Similar Objects Have Similar Grasp Positions?

Abstract

Listen

Translate article

In robotic grasping tasks, shape similarity has been widely adopted as a reference in grasp positions prediction for unknown objects. However, to the best of our knowledge, the issue "do similar objects have similar grasp positions?" has not been quantitatively analyzed before. This work aims to confirm or disprove the question by analyzing the relationship between the object shape similarity and grasp positions similarity. To this end, we constructed a similarity-estimation plane (SE-Plane), whose horizontal and vertical axes indicate the objects similarity and grasp similarity, respectively. Then, the proof of the issue is equal to the confirmation of the inference that "the points with higher objects similarity accordingly own higher grasps similarity in the proposed SE-Plane". We adopted several classical shape descriptors and two kinds of widely recognized deep neural network (DNN) architectures as objects similarity strategies. Furthermore, we employed the widely adopted intersection-over-union (IoU) of grasp anchors to measure the grasp similarity between objects. The experiments were carried out on a dozen objects with commonly seen primitive shapes selected from two well-known open grasp datasets: Cornell and Jacquard. It was found that the IoU values of grasp anchors are generally proportional to those of objects similarity in the SE-Plane. In addition, we obtained several primitive shapes from the commonly seen shapes, which are more suitable references in grasp positions prediction for unknown objects. We also constructed a realistic object dataset that included the objects with commonly seen primitive shapes. With the IoU prediction strategy learned from Cornell and Jacquard, the IoU predicted for realistic objects yielded similar results in the proposed SE-Plane. These discussions indicate that "similar objects have similar grasp positions" is reasonably correct. The proposed SE-Plane presents a new strategy to measure the relationship between objects similarity and grasp similarity.