Vision-Guided Anthropomorphic Grasping: Toward Precision Manipulation of Arbitrary Objects

Grasping and manipulating arbitrary, previously unseen objects is an essential capability for the next generation of physical intelligence ~ robots that operate in unstructured environments such as homes, service settings, and manufacturing. Yet most deployed systems rely on simplistic grippers and object-specific models, which limit both dexterity and generalization. The human hand offers far greater versatility, but anthropomorphic grippers that emulate it introduce perception and control problems that remain difficult to solve in a general way. In this work, we present a vision-guided system and methodology that couples computer vision with an anthropomorphic gripper on a robotic arm, enabling a robot to recognize, grasp, and manipulate objects without per-object engineering. The system interprets visual input to estimate an object's shape and position, determines a suitable grasp, and plans the arm's motion to execute it. We evaluate the approach on physical hardware across a diverse set of objects spanning a range of shapes, sizes, and materials, measuring grasp reliability and generalization to objects unseen during development. The result is a reproducible methodology and an empirical characterization of vision-guided anthropomorphic grasping, advancing the broader pursuit of general-purpose physical intelligence.