Look then move, versus look while moving
The simplest way to use a camera with a robot is look then move: take an image, figure out where the target is, compute a path, and execute it open-loop, without checking the camera again until the move is finished. This works fine when the world holds still and the robot's model of its own motion is accurate. In practice neither assumption survives contact with reality. Calibration drifts, parts shift on a conveyor, a patient breathes, a branch sways in the wind, and any error in the initial estimate simply rides along uncorrected for the entire move.
Visual servoing replaces that single snapshot with a continuous stream. The camera keeps watching while the robot moves, and the controller keeps recomputing the error between where things are now and where they should end up, feeding a fresh correction into the motion many times per second. The robot is no longer executing a plan, it's chasing a moving target definition that gets updated in real time. This is what makes it a genuinely closed-loop control problem rather than a perception-then-planning pipeline.