AnyGrasp

A general-purpose grasp-detection model from Shanghai Jiao Tong University that outputs many usable grasp poses directly from a point cloud.

AnyGrasp is a grasp-perception system from Cewu Lu's team (MVIG) at Shanghai Jiao Tong University, with the paper published in IEEE Transactions on Robotics (2023). Given a point cloud from a depth camera, it outputs a dense set of 7-degree-of-freedom grasp poses across the scene — gripper position, orientation, and opening width — each with a score. It's fairly robust to depth noise and can match the same grasp across consecutive frames, letting it track a moving object. The paper reports a single-arm system completing more than 900 grasps per hour. AnyGrasp is often used as a ready-made grasping module plugged into a larger system — OK-Robot, for instance, uses it to decide how to grasp. An SDK is released on GitHub, though the core library requires a machine-specific license.

ExampleShooting a single RGB-D frame of a cluttered tabletop, AnyGrasp returns hundreds of candidate grasps; the target object's segmentation mask filters out any that don't land on the target, and the highest-scoring remaining grasp is sent to the arm to execute.

Also called

AnyGrasp SDK

Related

Grasp Pose Detection · Bin Picking · GraspNet-1Billion · Contact-GraspNet · OK-Robot · Point Cloud

Sources

AnyGrasp (arXiv 2212.08333, IEEE T-RO)
AnyGrasp 项目页（SJTU MVIG） (Chinese)
graspnet/anygrasp_sdk

As of

2026-07