This paper presents a visual servoing approach based on stereo vision. The pair of cameras is mounted on the end-effector of the manipulator arm. Theoretical developments are presented using either raw pixel coordinates, or 3D coordinates estimated from image features. No geometrical model is needed. The experimental setup is challenging: large rotations are involved, images are noisy, and cameras are coarsely calibrated. In this setup, the trajectory of the end-effector differs notably, sometimes leading the arm near its joint range limits. Experimental results demonstrate that using pixel coordinates is disadvantageous, compared with 3D coordinates estimated from the same pixel data.