OpenDexGrasp: Open-vocabulary Task-Oriented Dexterous Grasping

This paper presents OpenDexGrasp, a unified framework for generating dexterous grasps that align with functional task intents described in free-form language. The system integrates large-scale synthetic grasp data (OpenDex-Scale) with high-quality human teleoperation data (OpenDex-Align) to learn a shared perception-action latent space. Experiments show improved physical feasibility, generalization to unseen object categories, and successful execution on real robots without requiring a separate affordance inference stage.