Exercise therapy is considered as one of the main rehabilitation treatments for post-stroke patients, especially by utilizing modern technologies, such as virtual and/or augmented reality. However, in order to design an appropriate exercise program, which prolongs the exercise duration and maximize the patient's improvement, the fatigue status needs to be detected and used for the program adjustment. In the previous fatigue recognition works, only exercises for healthy and athlete subjects have been taken into account. In this paper, fatigue status classification has been accomplished in a rehabilitation exercise for poststroke patients. To do so, the reaching task, as a basic rehabilitation exercise, was performed by post-stroke patients, utilizing Xbox Kinect; and surface EMG signal and Maximum voluntary contraction (MVC) of the subjects were collected during the exercises. The MVC values were used as the reference for fatigue status. Several features were determined and extracted from the sEMG and finally, classification of fatigue status on the sE MG was performed by two well-known classifiers: Hidden Markov Model (HMM) and Artificial Neural Network (ANN). An accuracy of 95.3% was achieved by HMM, which is a promising step toward an automated fatigue status recognition system in post-stroke rehabilitation exercises. © 2017 IEEE.