GraspNet-1Billion

A Shanghai Jiao Tong University dataset of real cluttered-scene grasping data, with over 1.1 billion annotated grasp poses.

GraspNet-1Billion is a general object-grasping benchmark from the MVIG lab at Shanghai Jiao Tong University (Cewu Lu's group), published at CVPR 2020, with an extended version published in the robotics journal IJRR in 2023. It contains 190 real cluttered scenes with 88 object types, captured using two RGB-D cameras, a RealSense D435 and an Azure Kinect, for a total of 97,280 images. Every image is annotated with each object's 6D pose (3D position plus 3D orientation), instance masks, and dense 6-DoF grasp poses, totaling more than 1.1 billion grasps. Earlier grasping datasets were mostly single-object or synthetically generated in simulation; this dataset provides large-scale annotation on real, cluttered tabletops, and ships with an open-source evaluation API and baseline network so different grasp-detection algorithms can be compared under the same standard. The same group's later work, AnyGrasp, builds on this line of research.

ExampleTo train a grasp-detection network: feed in one frame of depth point cloud, output several gripper poses with scores, then compute average precision on test scenes using the graspnetAPI.

Also called

GraspNet, GraspNet-1B

Related

Grasp Pose Detection · AnyGrasp · Antipodal Grasp · 6D Object Pose Estimation · Bin Picking · SJTU MVIG Lab

Sources

GraspNet-1Billion 官网 (Chinese)

As of

2023