Many diseases affect human movement functions and daily living. In order to recover the ability of movement, rehabilitation is the only way to improve the situation for many clients. Therefore, this study proposed a novel robot with brain-computer interface for rehabilitation exercises, namely BCRS (Brian-Controlled Rehabilitation System) and using multiple kernel learning (MKL) trains the classifier. BCRS detects and classifies the P300 and non-P300 signals from human brain and determines what kind of the rehabilitation exercises will be chosen. Three types of exercises, passive range of motion, isotonic, and isometric exercise, were realized in the system and support vector machine was used as the classification algorithm. For the three exercises, a new P300 panel was designed and composed of 25 commands. Through the experiments, we can find that BCRS can achieve good performance for rehabilitation exercises and MKL is a good method for EEG to have good accuracy of P300 signal classification and low training time than support vector machine (SVM).