Wrist Camera

A camera mounted on a robot arm's wrist or gripper that moves with the hand, giving a close-up view of the workspace.

A wrist camera is fixed near the gripper on a robot arm's end effector, moving together with the hand. This is the ‘eye-in-hand’ configuration described in visual servoing, where the camera moves with the hand and sees the hand's position relative to the target; the opposite is a fixed, third-person camera in the environment, called ‘eye-to-hand.’ The advantage of a wrist camera is that it sees the final few centimeters of a grasp or insertion most clearly, is less likely to be blocked by the arm itself, and shows the object's position directly in relation to the gripper — so imitation-learning and VLA data collection almost always includes one, feeding into the policy alongside one or two third-person cameras. Its drawback is that its viewpoint changes drastically with the arm's motion, it has no global view, and a depth camera can become inaccurate at very close range. Using its images for geometric calculations first requires hand-eye calibration.

ExampleThe DROID dataset mounts 3 cameras on each Franka arm: 2 adjustable-position external ZED 2 stereo cameras, plus 1 ZED Mini on the wrist.

Also called

Wrist-Mounted Camera, Eye-in-Hand Camera

Related

Eye-in-Hand · Third-Person Camera · Head Camera · Hand-Eye Calibration · Multi-View · Occlusion

Sources

Wikipedia: Visual servoing (eye-in-hand vs. eye-to-hand)
DROID: A Large-Scale In-the-Wild Robot Manipulation Dataset (project page)
ALOHA project page