A Self-supervised Contrastive Learning Method for Grasp Outcomes Prediction

Abstract

In this paper, we probe the proficiency of contrastive learning techniques in forecast grasp outcomes, without supervision. Employing a dataset that’s open to the public, we establish the effectiveness of contrastive learning techniques in accurately predicting grasp outcomes. More precisely, an impressive accuracy of 81.83% is achieved by the dynamic-dictionary-based method combined with the momentum updating technique, using data from just one tactile sensor, surpassing other unsupervised techniques. Our findings underscore the promise held by contrastive learning techniques in the domain of robotic grasping, while emphasizing the critical role played by precise grasp prediction in securing stable grasps.