In recent years, several biological frameworks have been proposed to imitate human motion and control lower limb exoskeletons. Compared with position or impedance tracking of defined joint trajectories, this kind of strategy can reproduce human walking dynamics, and tolerate external disturbance. The output of this kind of strategy is anticipatory feedforward torque, which means that the controller needs to have sufficient prior knowledge on neurology. Moreover, lower limb exoskeletons are fundamentally different from humans, which weakens the ability of the controller to resist external internal interference. Thus, pure feedforward control hardly achieves a flexible movement like that of a human. In this study, we propose a feedback control framework based on repetitive learning control (RLC) to enhance the anti-interference capability of a neuromuscular controller. The controller consists of two parts: (1) A data-driven morphed nonlinear phase oscillator is used as a state observer to learn the changing law of an exoskeleton’s posture and center of mass and to construct a stable limit cycle in the state space. (2) A posture and centroid tracker based on RLC is utilized to track the output of oscillators and achieve a natural balance recovery process. Simulation and experimental results show that the integrated control system has a better control effect than the simple biological control method.