2 weeks ago
24243
ARTICLE AD BOX
Rongchai Wang Nov 06, 2024 17:27
Galbot has developed DexGraspNet, a comprehensive dataset for robotic dexterous grasping using NVIDIA Isaac Sim, enhancing humanoid robots' ability to manipulate objects efficiently.
Robotic dexterous grasping is a pivotal area in robotics, focusing on enabling humanoid robots to handle and manipulate objects with human-like dexterity. According to NVIDIA, Galbot, a robotics firm, has made significant strides in this domain by developing a large-scale dataset called DexGraspNet using NVIDIA Isaac Sim.
Creating a Comprehensive Dataset
DexGraspNet is a groundbreaking dataset that encompasses 1.32 million ShadowHand grasps across 5,355 objects, spanning over 133 categories. This dataset is two orders of magnitude larger than previous datasets like Deep Differentiable Grasp, offering a wide array of grasps for each object instance. This extensive dataset facilitates more accurate training of algorithms, enabling robots to perform complex tasks requiring fine motor skills.
Innovative Techniques and Tools
Galbot utilized NVIDIA Isaac Sim, a robust simulation tool, to validate a vast number of grasps, addressing previous challenges in scaling dexterous grasping datasets. They employed a deeply accelerated optimizer to synthesize stable and diverse grasps efficiently. This approach ensured that the dataset includes grasps that were previously unattainable with other tools.
Advancements in Grasping Algorithms
Through cross-dataset experiments, Galbot demonstrated that training algorithms on DexGraspNet significantly outperformed previous datasets. The company introduced UniDexGrasp++, a novel approach for learning generalized dexterous grasping strategies. This method leverages GeoCurriculum Learning and Geometry-Aware Iterative Generalist-Specialist Learning (GiGSL) to enhance the generalizability of vision-based grasping strategies.
Scaling and Real-World Application
Galbot's advancements extend to real-world applications with DexGraspNet 2.0, which includes dexterous grasping in cluttered environments and demonstrates a 90.70% success rate in real-world scenarios. The team also developed a simulation test environment using NVIDIA Isaac Lab, accelerating the development and implementation of dexterous grasping models.
These developments mark a significant leap forward in humanoid robotics, enabling robots to better mimic human dexterity and efficiency in handling objects. Galbot's work, supported by NVIDIA's simulation tools, continues to push the boundaries of what is possible in robotic dexterous grasping.
Image source: Shutterstock
English (US) ·