This paper presents a new approach to real-time robot localization using kernel principal component analysis (PCA) regularization. The proposed algorithms are formulated as a semi-supervised learning during offline training. Firstly,sparse area features are extracted from the images captured by the camera mounted on the robot which moves along a predetermined path,and labeled a part of the data with their coordinates. Then,the coordinates of the unlabeled data are estimated by least squares with constraint of regularized low dimensional visual manifold in kernel PCA. In online localization stage,harmonic functions are employed to predict new data coordinates so that the real-time robot localization can be implemented using uncalibrated monocular vision. A series of experiments manifest that the proposed algorithms can outperform other conventional methods with low computational complexity,high localization accuracy and well real-time performance,so as to meet the real-time application requirements of industrial robots and medical service robots.