DemoFunGrasp: Universal Dexterous Functional Grasping via Demonstration-Editing Reinforcement Learning

We present DemoFunGrasp, a sim-to-real framework for universal dexterous functional grasping via demonstration editing reinforcement learning.

Key Concept

DemoFunGrasp enables affordance-centric grasping, supporting arbitrary grasp poses conditioned on arbitrary affordance regions across diverse objects. It generalizes to unseen object–affordance–style combinations with zero-shot sim-to-real transfer. During real-world deployment, a Vision–Language Model (VLM) can be used for autonomous planning and execution.

Real-World Demo

Pipeline

The framework consists of four components:

• Demonstration editing: Adapt a source demonstration using end-effector transformation and object-geometry–aware hand-style adjustment.
• Functional grasping policy learning: Train a single-step reinforcement learning policy conditioned on both affordance cues and grasp style specifications.
• Vision-based imitation: Transfer the learned policy to RGB observations for closed-loop, vision-based execution.
• Real-world deployment: Use a VLM for autonomous planning and execution.

Experiment

Our experiments evaluate simulation and real-world performance under human-annotated and VLM-predicted grasping conditions. DemoFunGrasp generalizes effectively to unseen object–affordance–style combinations and attains high success rates (GSR), affordance accuracy (IAS), and style adherence (ISS). Integrated with a VLM for instruction-following, the system reaches strong real-world performance.

Citation

@inproceedings{mao2026universal,
  title={Universal Dexterous Functional Grasping via Demonstration-Editing Reinforcement Learning},
  author={Chuan Mao and Haoqi Yuan and Ziye Huang and Chaoyi Xu and Kai Ma and Zongqing Lu},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  year={2026}
}