PDF Paper arXiv arXiv github Project

D e m o G r a s p

Universal Dexterous Grasping from a Single Demonstration


PKU    RUC    BeingBeyond

PDF Paper arXiv arXiv github Project

DemoGrasp: Universal Dexterous Grasping from a Single Demonstration

Haoqi Yuan*1,3   Ziye Huang*1,3   Ye Wang2,3   Chuan Mao1   Chaoyi Xu3   Zongqing Lu§1,3


1PKU     2RUC     3BeingBeyond


*Equal Contribution   §Corresponding Author

Paper arXiv Project

Abstract

    Universal grasping with multi-fingered dexterous hands is a fundamental challenge in robotic manipulation. While recent approaches successfully learn closed-loop grasping policies using reinforcement learning (RL), the inherent difficulty of high-dimensional, long-horizon exploration necessitates complex reward and curriculum design, often resulting in suboptimal solutions across diverse objects. We propose DemoGrasp, a simple yet effective method for learning universal dexterous grasping. We start from a single successful demonstration trajectory of grasping a specific object and adapt to novel objects and poses by editing the robot actions in this trajectory: changing the wrist pose determines where to grasp, and changing the hand joint angles determines how to grasp. We formulate this trajectory editing as a single-step Markov Decision Process (MDP) and use RL to optimize a universal policy across hundreds of objects in parallel in simulation, with a simple reward consisting of a binary success term and a robot–table collision penalty. In simulation, DemoGrasp achieves a 95% success rate on DexGraspNet objects using the Shadow Hand, outperforming previous state-of-the-art methods. It also shows strong transferability, achieving an average success rate of 84.6% across diverse dexterous hand embodiments on six unseen object datasets, while being trained on only 175 objects. Through vision-based imitation learning, our policy successfully grasps 110 unseen real-world objects, including small, thin items. It generalizes to spatial, background, and lighting changes, supports both RGB and depth inputs, and extends to language-guided grasping in cluttered scenes.
overview

    DemoGrasp is a framework for learning universal dexterous grasping policies via reinforcement learning (RL) augmented with a single demonstration. It achieves state-of-the-art performance across diverse robotic hand embodiments and transfers effectively to real robots, demonstrating strong generalization.
pipeline

    DemoGrasp uses a single demonstration trajectory to learn universal dexterous grasping, formulating each grasping trial as a demonstration-editing process. For each trial, the Demo Editor policy takes observations at the first timestep and outputs an end-effector transformation and a delta hand pose. The actions in the demonstration are then transformed accordingly and applied in the simulator. The policy is trained using RL across diverse objects, optimizing a simple reward consisting of binary success and a collision penalty. A flow-matching policy is trained on successful rollouts with rendered images to enable sim-to-real transfer.
dexgraspnet

    DemoGrasp achieves state-of-the-art performance on DexGraspNet with ShadowHand. Results are reported for both state-based and vision-based settings on 3,200 training objects (Train.), 141 unseen objects from seen categories (Test Seen Cat.), and 100 unseen objects from unseen categories (Test Unseen Cat.).
embodiments

    DemoGrasp is extensible to any robotic hand embodiment without hyperparameter tuning, achieving high zero-shot success rates on various object datasets while being trained on only 175 objects.
real-success

    DemoGrasp successfully grasps 110 objects on a real robot, including small and thin objects. It is also extensible to real-world cluttered grasping and instruction-following, by including random distractor objects and automatically generated language descriptions during vision-based data collection in simulation.

Failure Cases

BibTeX

      
        @article{yuan2025demograsp,
          title={DemoGrasp: Universal Dexterous Grasping from a Single Demonstration},
          author={Yuan, Haoqi and Huang, Ziye and Wang, Ye and Mao, Chuan and Xu, Chaoyi and Lu, Zongqing},
          journal={arXiv preprint arXiv:2509.22149},
          year={2025}
        }