In this paper we present a robust polynomial classifier based on L1-norm minimization. We do so by reformulating the classifier training process as a linear programming problem. Due to the inherent insensitivity of the L1-norm to influential observations, class models obtained via L1-norm minimization are much more robust than their counterparts obtained by the classical least squares minimization (L2-norm). For validation purposes, we apply this method to two recognition problems: character recognition and sign language recognition. Both are examined under different signal to noise ratio (SNR) values of the test data. Results show that L1-norm minimization provides superior recognition rates over L2-norm minimization when the training data contains influential observations especially if the test dataset is noisy.