In this research, the Robust Regression method used for face recognition tested its performance with illumination variations on the training dataset. Experiments were carried out using Cropped Yale Face Database B. By using this standard face database, generally the data for the training process used all images in subset 1 and the testing process was carried out on all images in other subsets. The training process in this method is done to create a regressor or predictor. In this research experiment, training data use each subset. Also, this research experiment will also combine several images from all subsets. The experimental results show that the use of subset 1 images as training data turns out to produce the lowest facial recognition performance where the accuracy is 90.00%. The use of other subsets as training datasets can deliver better facial recognition performance. The highest facial recognition performance is achieved through the use of combined images of sample images from all subsets, where accuracy reaches 99.81%.