Abstract: Three-dimensional (3D) deviation extraction of the welding torch for the automated welding process is a prerequisite in robotic intelligentized arc welding. In this paper, with the welding robot meal active gas (MAG) arc welding method was used to join steel sheets with micro grooves. Passive vision sensors with the appropriate combination of the filter and dimmer glass were used to capture images which contained complete arc regions, edge lines of the seam, and joint lines. An effective algorithm was then proposed to directly extract the seam line of the butt joint. To determine the tracking point at each sampling time the real-time position of the welding torch was first signed with the geometric center of the arc region. Then, another algorithm was suggested to determine the tracking point that lies in the extracted seam line. It was transformed into the world coordinate using existing vision calibrated techniques, and the 3D deviation was yielded when the exact position of the welding torch in the world coordinate system had been recorded from the control system of the welding robot. Experimental results show that the proposed method in this paper can effectively acquire the 3D deviation of the welding torch in real time.