Improving Humanoid Grasp Success Rate based on Uncertainty-aware Metrics and Sensitivity Optimization

Abstract

We present an approach for the selection of robot grasp candidates by treating specified metrics in a probabilistic manner and maximizing the success rate through statistical optimization. Recently, progress has been made in grasping unknown objects in cluttered scenes by using deep neural networks or incorporating classifiers. Although existing methods deliver promising results, they either lack explainability or fail to account for uncertainties that accumulate over the entire system. To address this shortcoming, we optimize a ranking score based on the sensitivities of the grasp success with respect to a set of metrics. These sensitivities reflect each metric's contribution to the success. To perform this optimization, we refer to a dataset of 932 randomly selected grasps recorded under real-world conditions with the humanoid robot ARMAR-6. By validating our approach on a separate data collection of 187 physical real- world grasps, we demonstrate that our approach yields a success rate of 73.8 %, amounting to an improvement of more than 40 % compared to a random grasp selection. The results exemplify that sensitivity optimization, scarcely applied in the context of robotic applications so far, can significantly enhance the grasp success by considering respective metrics in the face of uncertainties.