Reproduce “DemoGrasp: Universal Dexterous Grasping from a Single Demonstration” on OpenResearch

About this paper

Haoqi Yuan, Ziye Huang, Ye Wang, Chuan Mao, Chaoyi Xu, Zongqing Lu

The DemoGrasp framework enables universal dexterous grasping for robots by adapting a single successful demonstration through editing parameters, reformulating reinforcement learning into a single-step Markov Decision Process. This approach achieves an 86.5% success rate on 110 unseen real-world objects, including 71.1% on small and 68.3% on thin objects, with robust zero-shot sim-to-real transfer and extendibility to cluttered and language-guided scenarios.

Key insights

• Reframing dexterous grasping as a single-step MDP over demonstration-editing parameters drastically reduces the RL exploration burden and simplifies reward design to a sparse success signal.
• Leveraging a single high-level demonstration as a transferable template, rather than learning low-level actions from scratch, provides a robust foundation for universal grasping.
• Comprehensive domain randomization and a two-stage learning process (state-based RL then vision-based imitation) effectively bridge the sim-to-real gap, enabling practical real-world deployment.

Summary by alphaXiv.