DexMGNet: Multi-Mode Dexterous Grasping in Cluttered Scenes With Generative Models

Dexterous grasping is a crucial technique in humanoid robot manipulation. However, existing methods still fall short in effectively detecting dexterous grasps in cluttered environments. In this work, we propose DexMGNet, a novel multi-mode dexterous grasping framework designed for such challenging scenarios. We introduce the concept of pre-grasping and redefine dexterous grasping to enhance adaptability. We propose an effective pre-grasp and grasp data sampling strategy and develop a conditional generative model for grasp and pre-grasp generation. Additionally, we integrate pre-grasp collision detection within the hand's workspace, significantly improving grasping performance in cluttered environments. Our method supports multi-mode grasping, including two-finger, three-finger, and four-finger grasps, enabling greater flexibility across diverse grasping tasks. In real-world desktop grasping experiments, our approach achieves a 93.3% success rate in single-object scenes and a 78.3% success rate in multi-object scenes, demonstrating its effectiveness and superiority.