Wang Long Chai
November 6, 2024 17:27
Galbot used NVIDIA Isaac Sim to develop DexGraspNet, a comprehensive dataset for proficient grasping by robots to efficiently improve the object manipulation capabilities of humanoid robots.
Robotic dexterity is a pivotal area of robotics, focusing on enabling humanoid robots to handle and manipulate objects with human-like dexterity. According to NVIDIA, robotics company Galbot has made significant progress in this area by using NVIDIA Isaac Sim to develop a large-scale dataset called DexGraspNet.
Create comprehensive datasets
DexGraspNet is a groundbreaking dataset containing 1.32 million ShadowHand grasps across 5,355 objects across over 133 categories. This dataset is twice as large as previous datasets such as Deep Differentiable Grasp, providing a comprehensive view of each object instance. This extensive data set will facilitate more accurate training of the algorithms, allowing robots to perform complex tasks that require fine motor skills.
Innovative technologies and tools
Galbot leveraged NVIDIA Isaac Sim, a powerful simulation tool, to validate a large number of grasps and solve previous challenges of scaling proficient grasp data sets. They used a highly accelerated optimizer to efficiently synthesize stable and diverse grasps. This approach allowed the data set to contain insights previously unattainable with other tools.
Advances in grasping algorithms
Through cross-dataset experiments, Galbot demonstrated that DexGraspNet’s training algorithm significantly outperforms previous datasets. The company introduced UniDexGrasp++, a new approach for learning generalized dexterous grasping strategies. This method leverages GeoCurriculum Learning and Geometry-Aware Iterative Generalist-Specialist Learning (GiGSL) to improve the generalizability of vision-based grasping strategies.
Expansion and practical application
Galbot’s advancements extend to real-world applications with DexGraspNet 2.0, which includes proficient grasping capabilities in complex environments and demonstrates a 90.70% success rate in real-world scenarios. The team also developed a simulation testing environment using NVIDIA Isaac Lab to accelerate the development and implementation of a proficient grasping model.
These developments represent a significant leap forward in humanoid robotics, allowing robots to better mimic human dexterity and object-handling efficiency. Galbot’s work, supported by NVIDIA’s simulation tools, continues to push the boundaries of dexterous grasping by robots.
Image source: Shutterstock