Current learning-based robot grasping approaches exploit human-labeled
datasets for training the models. However, there are two problems with such a
methodology: (a) since each object can be grasped in multiple ways, manually
labeling grasp locations is not a trivial task; (b) human la