AnyGrasp Pipeline

Key Insight

Deploying AnyGrasp on a physical manipulator creates a robust 3D grasping pipeline that works directly with raw point clouds from a depth camera (which generates a depth map). The network processes the scene to perform 6-DoF pose estimation for thousands of candidate grasps, scoring each based on surface geometric features. Executing these grasps on a real arm demonstrates how data-driven models bypass the need for prior CAD models, enabling reliable picking of arbitrary, highly cluttered objects.