A robotic seam tracking system is developed to achieve robustness against optical noises such as are glares, welding spatters, fume, and other unexpected brightness sources. The profile data of a weld joint to be welded are reliably extracted using two separate vision processing algorithms: the first is for joint modeling before welding starts, while the second is for joint feature detection during welding. Each procedure is divided into several consecutive steps. In particular, the syntactic approach is refined to improve the reliability of the joint features extracted. To achieve more careful syntactic analysis, several junction primitives and production rules are newly defined. From the joint features thus obtained, the three-dimensional information of the weld joint is extracted to achieve the robot path correction. To investigate the performance of the developed visual system, a series of experiments on joint feature detection and robotic seam tracking are conducted for four different types of weld joints: butt, lap, fillet and vee. The results show that the system is very robust in the presence of various welding noises as well as variations in appearance of weld joint and workpiece.