The article introduces DexDiffuser, a novel method for dexterous grasping that generates, evaluates, and refines grasps on partial object point clouds. DexDiffuser includes DexSampler, a new conditional diffusion-based dexterous grasp sampler, and DexEvaluator, a new dexterous grasp evaluator. The method was tested on the Allegro Hand in both simulation and real-world experiments, showing that DexDiffuser outperforms the state-of-the-art multi-finger grasp generation method FFHNet with a higher grasp success rate. The paper also proposes two grasp refinement strategies: Evaluator-Guided Diffusion and Evaluator-based Sampling Refinement.
Publication date: 5 Feb 2024
Project Page: https://arxiv.org/abs/2402.02989v1
Paper: https://arxiv.org/pdf/2402.02989
