Eye-in-Hand Camera

• Also called: wrist-mounted camera, wrist camera, hand camera
• Category: Sensing › Exteroceptive
• Tags: cameras, manipulation, hand-eye-calibration, end-of-arm-tooling, robot-learning

A camera mounted on a robot's wrist or end-effector so that it moves with the gripper, giving close-up views of what the hand is about to touch.

Definition

An eye-in-hand camera is mounted on the robot's wrist, flange or gripper, so its viewpoint moves with the end-effector. The alternative, an eye-to-hand camera, is fixed in the workcell or on the robot's head or base and watches the arm from outside. Many systems combine both: a fixed camera for an overview of the scene and a wrist camera for the final approach.

The wrist position gives several benefits. The camera sees the object from close range at the moment it matters, so small features and alignment errors are visible at high resolution. The view is roughly constant relative to the gripper, which makes the relation between image features and needed gripper motion simple; this is the basis of eye-in-hand visual servoing, and in learned manipulation policies it often makes behavior less sensitive to where objects sit in the scene. The camera can also be moved to look around occlusions or scan an object from several sides. The costs are a narrow, constantly changing field of view that loses track of the wider scene, occlusion by the gripper's own fingers and the grasped object, motion blur during fast moves, a minimum focus or depth range that may be too long when the gripper is closest to the object, and extra cabling and mass at the wrist.

Using the camera's images in robot coordinates requires hand-eye calibration: estimating the fixed transform between the camera and the flange. The standard formulation, AX = XB, compares camera motions observed against a calibration target with the corresponding flange motions from forward kinematics; Tsai and Lenz published a widely used solution in 1989. Errors in this transform appear directly as grasp offsets, so calibration is repeated whenever the camera mount is changed or knocked.

Why it matters

Wrist cameras are standard on research manipulators and common in industrial picking because they give precise views exactly where contact happens. Studies of learning from demonstration have found that adding a wrist view often improves success on grasping and insertion tasks compared with fixed cameras alone.

Key applications

• Final-approach alignment for grasping and insertion
• Image inputs for learned visuomotor manipulation policies
• Eye-in-hand visual servoing
• Active viewpoint selection and object scanning
• Inspection of parts and features at close range

In practice

A team collecting teleoperated demonstrations for a pick-and-place policy records images from a fixed camera facing the table and from a small RGB camera clipped to the gripper. During training, both views are fed to the policy. In tests, the wrist view lets the policy correct small alignment errors just before closing the gripper, which the distant fixed camera cannot resolve. Before collecting data, the team runs a hand-eye calibration by moving the arm through a set of poses in front of a checkerboard, and they add a cable guide along the forearm so the camera lead does not snag when the wrist rotates.

Common misconceptions

• A wrist camera replaces a fixed scene camera — its narrow view loses the wider context, so many systems use both
• The camera pose is known once it is mounted — the camera-to-flange transform must be estimated by hand-eye calibration and checked after any change to the mount
• Closer is always better — depth cameras have a minimum range, and very close views are often occluded by the fingers or the object itself
• Eye-in-hand and eye-to-hand calibration are the same problem — the unknown transform differs, linking the camera to the flange in one case and to the robot base in the other

Who should know this

• Manipulation and grasping engineers
• Robot learning researchers collecting demonstration data
• Machine vision integrators for industrial cells
• End-of-arm tooling designers

Related technologies

• Hand-Eye Calibration
• Wrist-Mounted RGB-D Cameras
• Visual Servoing
• Fiducial Markers
• Robomimic

Related terms

• Visual Servoing
• RGB Camera
• Depth Camera
• End Effector
• Visuomotor Policy
• Gripper

Related projects

• LeRobot: HuggingFace ML library for real-world robotics in PyTorch
• OpenPI: Physical Intelligence VLA foundation models for general robot control
• Isaac Lab: NVIDIA robot learning framework: one task API across Newton, PhysX and Isaac Sim
• NVIDIA Isaac GR00T: NVIDIA open foundation model for humanoid robot reasoning and skills
• OpenVLA: 7B-parameter Vision-Language-Action model for robotic manipulation
• Drake: MIT robotics toolbox for model-based dynamics, optimization, and control system design