A Study on the Influence of Task Dependent Anthropomorphic Grasping Poses for Everyday Objects

Niko Kleer, Martin Feick

Abstract: Robots using anthropomorphic hands and pros-thesis grasping applications frequently rely on a corpus of labeled images for training a learning model that predicts a suitable grasping pose for grasping an object. However, factors such as an object's physical properties, the intended task, and the environment influence the choice of a suitable grasping pose. As a result, the annotation of such images introduces a level of complexity by itself, therefore making it challenging to establish a systematic labeling approach. This paper presents three crowdsourcing studies that focus on collecting task-dependent grasp pose labels for one hundred everyday objects. Finally, we report on our investigations regarding the influence of task-dependence on the choice of a grasping pose and make our collected data available in the form of a dataset.