Grasping

A robot picking up an object securely with a gripper, suction cup, or dexterous hand.

Grasping is a robot using an end effector — a gripper, a suction cup, or a dexterous hand — to hold an object securely and lift it. The core question is where and how to grasp: an algorithm computes a grasp pose (the gripper's position and orientation) from a camera image or point cloud, and a motion planner then moves the hand there. Early methods relied on geometric and mechanical analysis, for instance checking force closure — whether the fingers' contact forces can resist external force and torque from any direction. A 2013 survey by Jeannette Bohg and colleagues organized data-driven approaches into three categories, based on whether the object was known, similar to a known object, or entirely unknown. Today the common approach is a deep network that predicts grasp poses directly from a point cloud, or a VLA model that outputs grasping actions end to end. Grasping is the first step in many manipulation tasks, from pick-and-place to mobile manipulation.

ExampleA depth camera photographs a cluttered table; a network proposes dozens of candidate gripper poses over the point cloud and scores them, and the robot picks up the mug using the highest-scoring one.

Also called

Robotic Grasping

Related

Pick-and-Place · Grasp Pose Detection · Force Closure · Gripper · Dexterous Hand · Bin Picking

Sources

Data-Driven Grasp Synthesis - A Survey (Bohg et al., arXiv:1309.2660)