Grasp Pose Detection

Computing directly from an image or point cloud where and in what orientation a gripper should grasp an object.

Grasp pose detection is the perception step of robotic grasping: given an RGB-D image or point cloud, it outputs a batch of feasible grasp poses along with scores. A planar grasp is commonly represented as an oriented rectangle in the image, giving the gripper's position, angle, and opening width; a 6-DOF grasp instead gives the gripper's 3D position, approach direction, rotation, and opening width, letting it grasp from the side or at an angle from above. Unlike 6D pose estimation, it needs no object model and doesn't need to recognize what the object even is, which suits grasping unfamiliar objects in clutter. In 2017, Andreas ten Pas and colleagues' GPD treated the problem in a detection-like way: sample candidates first, then classify each one. GraspNet-1Billion provides a benchmark with 88 objects, 190 scenes, and over 1.1 billion grasp annotations; AnyGrasp, Contact-GraspNet, and GraspGen are commonly used models, with the detected grasp then handed to motion planning to execute.

ExampleA bin holds various parts piled together. A depth camera shoots a single frame as a point cloud, AnyGrasp outputs dozens of scored two-finger gripper poses, and the system picks the highest-scoring one that won't hit the bin wall, handing it to motion planning to execute.

Also called

Grasp Detection, 6-DoF Grasp Detection

Related

Grasping · 6D Object Pose Estimation · AnyGrasp · Contact-GraspNet · GraspNet-1Billion · Bin Picking

Sources

Grasp Pose Detection in Point Clouds (arXiv:1706.09911)
GraspNet-1Billion 官网 (Chinese)