Galbot has developed a large-scale dataset called DexGraspNet using NVIDIA Isaac Sim, which includes 1.32 million ShadowHand grasps across 5,355 objects. This dataset is two orders of magnitude larger than previous datasets, like Deep Differentiable Grasp, and helps train algorithms for more accurate robot manipulation.
Galbot used NVIDIA Isaac Sim, a robust simulation tool, and a deeply accelerated optimizer to create diverse and stable grasps. Their new approach, UniDexGrasp++, combined with GeoCurriculum Learning and Geometry-Aware Iterative Generalist-Specialist Learning (GiGSL), helps improve the generalizability of vision-based grasping strategies.
Galbot’s DexGraspNet 2.0 includes dexterous grasping in cluttered environments, with a 90.70% success rate in real-world scenarios. These advancements enhance humanoid robots’ ability to manipulate objects efficiently, bringing them closer to human-like dexterity.
