Functional $\small \mathcal{D(R,O)}$ Grasp:
A Language-Guided
Cross-Embodiment
Functional Dexterous Grasping

Wanqi Zhong1, Zhixuan Xu2, Chongkai Gao2, Zhaoyu Mu2, Yunfan Lou2, Lin Shao2
1Harbin Institute of Technology
2National University of Singapore
Paper Code(Comming Soon) Data(Comming Soon)

Abstract

Functional dexterous grasping is a challenging capability essential for robots to achieve intent-aligned interactions with objects. Existing methods primarily focus on grasp stability without addressing functional intent. In this work, we present Functional $\mathcal{D(R,O)}$ Grasp, a language-guided framework that enables intent-aligned grasp generation while ensuring cross-platform adaptability. We learn platform-agnostic intermediate representations that enable translation from functional grasp language input to execution across different robotic hands. This framework generates appropriate grasps for objects based on their intended use, covering multiple functional requirements (use, hold, handover, liftup). We demonstrate that our approach achieves a 75.1% success rate in simulation on unseen objects, significantly outperforming baselines. Real-world experiments with the LeapHand platform further validate our approach. Our work bridges the gap between functional intent and cross-platform dexterous execution, enabling robots to perform purposeful grasps with a single unified model.

Pipeline Overview

Overview of D(R,O) Grasp framework
Overview of our Functional $\mathcal{D(R,O)}$ Grasp framework.
Left: Language-Conditioned Interaction Synthesis translates functional instructions into wrist poses and contact anchor points via a diffusion model.
Right: Platform-Agnostic Grasp Refinement converts these interaction elements into a unified hand-object distance representation to generate precise joint configurations across different robotic hands.