Flexible joint manipulators are extensively used in several industries and precise control of their nonlinear dynamics has proven to be a challenging task. In this work, we want to compare two intelligent controllers by proposing two Takagi-Sugeno-Kang Neuro-Fuzzy Approaches (Type-1 and Type-2) to control a flexible joint. For both controllers, The inverse models are found using identification techniques, then they are put in series as inverse controllers to control the flexible joint in an online structure. Interval weights are trained by gradient descent approaches using backpropagation algorithms. Results reveal that, without any knowledge about the dynamic of the robot, the methods can control the flexible joint which is highly unstable. As illustrated in result section, One level more fuzziness of Type-2 in compare to type-1 fuzzy controllers helps this controller to more effectively deals with information from a knowledge base. The proposed models can effectively handle uncertainties arising from friction and other structural nonlinearities.