Human Universal Grasping

The article presents HUG, a flow-matching model designed to generate diverse human grasps from single RGB-D images, utilizing a dataset of 1M human grasps collected via smart glasses across various environments. The model outputs grasp parameters for wrist translation, rotation, and hand pose, allowing for zero-shot grasping with different robot hands. HUG demonstrates superior performance, achieving a 23% and 34% improvement over state-of-the-art grasping methods on a newly established benchmark, HUG-Bench, which includes 90 unseen objects, making it a significant advancement for practitioners in robotic grasping applications.