Hand-eye calibration finds the transform between a camera and a robot, so a point the camera sees can be turned into a point the robot can reach. It is a problem of Composite Homogeneous Transformations.
Two Set-Ups
- eye-to-hand: the camera is fixed in the cell and looks at the robot, and we want (camera pose in the robot base frame)
- eye-in-hand: the camera is mounted on the gripper, and we want
- the classic eye-in-hand form is , solved from several robot poses
Using The Transform
Ways To Get It
Measured by hand:
- measure the camera position and angles from the robot base and write the transform directly
- no markers needed, but small angle errors grow with distance: of error is about mm at 1 m
- there is no fitting error to check, so a reach test (move the gripper to a seen object) is the only check
Fitted from matched points (Kabsch):
- hold a small marker or ball in the gripper, move the robot to several places, and record its position from the robot and from the camera
- the best rotation and translation between the two point sets is found with SVD
Kabsch Algorithm
Steps:
- subtract each set's centroid: ,
- cross-covariance , then the SVD
- rotation , and if , flip the sign of the last column of and recompute (a reflection is not a rotation)
- translation
- the RMS error of the fitted points shows the calibration quality
Good Practice
- use points spread through the whole working volume, not on one plane
- use at least 4 points, and more reduces noise
- recalibrate when the camera moves, and keep the transform in version control so results stay traceable
- a fitted transform silently absorbs a fixed axis flip, while a measured one cannot, so a wrong camera axis convention shows up very differently in the two methods