A new manifold learning algorithm, called robust kernel neighborhood preserving projection (RKNPP), is presented and applied to radar target recognition based on high resolution range profiles. RKNPP attempts to map the high-dimensional data into such a low-dimensional space where points belonging to the same class are close to each other while points belonging to different classes are far away from each other, while preserving the main geometric structure of the original data. A sophisticated distance metric is utilized to construct the neighborhood graph of the input data, which has several good properties that are helpful to limit the effect of noise, and thus make RKNPP a robust classification method for real-world data. Moreover, in RKNPP, a simple technique of eigenvalue decomposition is applied to deal with the small sample size problem, to which not much attention has been paid in many manifold learning algorithms. Experimental results on measured data demonstrate the promising performance of the proposed method.