Human Universal Grasping

Researchers have introduced HUG, a flow-matching model that generates diverse human grasps for any object in a single RGB-D image, aiming to close the dexterity gap between humans and multi-fingered robots [1][2]. The model was trained on 1M-HUGs, an egocentric dataset of human grasps collected using smart glasses. The dataset spans 1 million frames, equivalent to 27.8 hours of footage, and covers 6,707 object instances across 41 buildings [1][2]. HUG fuses RGB and depth observations to output a grasp parameterized by wrist translation, wrist rotation, and MANO hand pose [1][2]. The predicted grasps can be retargeted to various robot hands, enabling zero-shot grasping in everyday scenes [1][2]. A robotic arm's end effector is analogous to the human hand, though the term "robotic hand" is often proscribed in technical contexts [6]. To standardize evaluation, the team built HUG-Bench, a simulated benchmark of 90 unseen objects from five geometric categories with metric-scale 3D meshes [1][2]. In real-world tests on a 30-object subset across multiple stereo cameras, robot embodiments, and household environments, HUG outperformed state-of-the-art grasping baselines by +23% and +34% on the challenging object set [1][2]. Code, data, checkpoints, and an interactive demo have been released [1][2]. The release includes an interactive demo accessible through a collaboration between arXiv and Hugging Face Spaces, which allows users to try machine learning models directly in a browser without writing code [9][10][11]. Since October 2021, Hugging Face Spaces has hosted over 12,000 open-source machine learning demos built with tools such as Gradio and Streamlit [9]. The arXiv integration embeds these demos alongside papers, increasing reproducibility and allowing a wider audience to explore results [9][10].