DexGraspNet

A simulation-generated dexterous-hand grasping dataset from Peking University, with 1.32 million grasps across 5,355 objects.

DexGraspNet is a large-scale dexterous-hand grasping dataset released by He Wang's group at Peking University, together with the Beijing Institute for General Artificial Intelligence (BIGAI) and Tsinghua University, published at ICRA 2023. Large datasets already existed for two-finger parallel-jaw grasping, but dexterous-hand grasping had long lacked data. The authors used an accelerated, differentiable force-closure estimator (force closure means the fingers' contact forces can resist an external force from any direction) to synthesize grasp poses at scale, generating 1.32 million grasps for a Shadow dexterous hand across 5,355 objects in more than 133 categories — over 200 grasps per object — all verified in the Isaac Gym simulator. It's mainly used to train and evaluate dexterous grasp-generation algorithms. In 2024 the team released DexGraspNet 2.0, expanding to 8,270 cluttered scenes and 427 million grasps, and demonstrating zero-shot transfer to a real robot.

ExampleA generative model that takes an object's point cloud as input and outputs a Shadow-hand grasp pose can be trained using DexGraspNet's grasps as supervision.

Also called

DexGraspNet: A Large-Scale Robotic Dexterous Grasp Dataset for General Objects Based on Simulation

Related

Dexterous Manipulation · Force Closure · Shadow Dexterous Hand · Isaac Gym · Synthetic Data · Grasp Pose Detection

Sources

DexGraspNet 项目主页 (Chinese)
DexGraspNet (arXiv)
DexGraspNet 2.0 (arXiv)

As of

2024-10