Reproduce “From Grasps to Dexterity: Large-Scale Grasp Pretraining for Dexterous Manipulation” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Ying Yuan, Xinyu Liu, Sriram Krishna, David Held

The method turns grasp annotations into grasp-and-move trajectories for a low-level controller, while a high-level policy predicts hand-shaped sub-goals for tool-use tasks. With 50 demonstrations per task, its six-task DexCraft success average is 60.3%, versus 48.9% for the same hierarchy trained from scratch. In three real-world tasks, trigger-stage success averages 40.7%, versus DP3’s 7.4%, in the reported settings.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.