To address the issue of the flexible joint (FJ) manipulator control with actuator saturation, singular perturbation method is adopted in this paper which decompose the manipulator model into two subsystems in terms of time scale transformation, the fast one represents flexibility and the slow one expresses rigidity, and the control law of which is designed respectively. For the fast subsystem, a velocity-difference-based feedback control is designed to lower the oscillation caused by joint flexibility, while for the slow subsystem, the control performance of the system is enhanced by a combination of a class of nonlinear integral sliding surface and backstepping global sliding mode. The radial basis function (RBF) neural networks are utilized to compensate the actuator saturation, the disturbance and the modeled uncertainties. Based on the Lyapunov theory, the tracking convergence of the closed-loop system is proved rigorously. The simulation result shows that the designed control law can maintain good rapidity and accuracy, moreover, chattering is eliminated.