UniDexGrasp

UniDexGrasp is a universal framework and large-scale dataset for robotic dexterous grasping developed by researchers at Peking University. The system provides over 100,000 diverse dexterous grasp proposals across 2,846 objects from the PartNet-Mobility and ShapeNet datasets, spanning 46 object categories with articulated structures. UniDexGrasp uses a diffusion-based generative model to propose diverse, physically feasible grasp configurations for a multi-fingered dexterous hand (Shadow Hand or Allegro Hand). It then learns a goal-conditioned policy to execute the grasp, enabling generalizable dexterous manipulation across unseen object categories. The framework achieves state-of-the-art performance on dexterous grasping benchmarks, demonstrating robust generalization to novel objects with diverse shapes, sizes, and articulation types. UniDexGrasp has become a standard approach for learning-based dexterous grasping.