GraspNet is an open project for general object grasping that is continuously enriched. It currently contains:

GraspNet-1Billion: a benchmark that contains 190 cluttered and complex scenes captured by two popular RGBD cameras (Kinect Azure and RealSense D435), bringing 97,280 images in total. For each image, we annotate the accurate 6D pose and the dense grasp poses for each object. In total, our dataset contains 88 objects and over 1.1 billion grasp poses . We hope our benchmark can facilitate related research in general object grasping as well as other related areas (e.g. 6D pose estimation, unseen object segmentation, etc.).

AnyGrasp: the first solution that achieves human-level grasping in cluttered scene. Enabled by GraspNet-1Billion, AnyGrasp can grasp any objects in any scenario, no matter rigid or deformable objects. It would be a solid basis for robotics manipulation. We release the library to facilitate related research. More details at this page.

SuctionNet-1Billion: a benchmark for suction-based grasping. Compared with other kinds of grasping, suction is usually more reliable and easier to represent. With the rich data provided by GraspNet-1Billion, we further provide suction-related labels and additional online evaluation system to boost the development of suction grasping and other related areas. More details at this page.

TransCG: the first large-scale real-world dataset for transparent object depth completion and grasping, whose construction process is accelarated by a novel semi-automatic pipeline. The dataset contains 57,715 RGB-D images of 51 transparent objects and many opaque objects captured from different perspectives of 130 scenes under real-world settings. More details at this page.

Community on GitHub
Click here to join our mailing list.