Predicting grasp success in the real world - A study of quality metrics and human assessment

Carlos Rubert, Daniel Kappler, Jeannette Bohg, Antonio Morales

Abstract: Highlights•Grasp quality metrics aim at quantifying different aspects of a grasp configuration.•They are often used to generate ground-truth labels for learning-based approaches.•Studies have highlighted the limitations of QM to predict the outcome of real grasps.•In this paper, we study how well commonly-used grasp metrics perform in real world.•To this end, we generated two datasets of grasp candidates in simulation.•The quality of these grasp candidates is quantified by the aforementioned metrics.•We replicate grasp candidates on two robotic systems and evaluate each grasp.•We trained different classifiers to predict grasp success using only QM as input.•Results show that combined QM can achieve up to a 85% classification accuracy.