XDex: Learning Cross-Embodiment Dexterous Grasping with 1000 Hands

Keyi Wang, Jianglong Ye, Chengjing Yuan, Xiaolong Wang

Keywords: Dexterous Grasping, Cross-embodiment

Abstract: Synthesizing dexterous grasps across various hands remains a fundamental challenge in robotic manipulation due to morphology gaps in geometry, topology, and kinematics. We hypothesize that scaling the diversity and number of hand embodiments improves generalization to unseen hands. To this end, we introduce XDex, a framework trained on the largest cross embodiment grasping dataset, which we built using 1,000 diverse hands. XDex features an embodiment transformer that jointly encodes hand geometry and topology to learn from this large scale dataset. Additionally, we enforce grasp consistency across embodiments by training on a paired grasping dataset and introducing a retargeting loss. The paired data are generated by first synthesizing grasps for a source hand and then translating them to diverse target hands. XDex significantly outperforms prior methods in grasp quality, consistency, and diversity, and demonstrates strong generalization to unseen hands in real world settings.

Primary Area: applications to robotics, autonomy, planning

Submission Number: 1130