Central pattern generators (CPGs) have been widely applied in robot motion control for the spontaneous output of coherent periodic rhythms. However, the underlying CPG network exhibits good convergence performance only within a certain range of parameter spaces, and the coupling of oscillators affects the network output accuracy in complex topological relationships. Moreover, CPGs may diverge when parameters change drastically, and the divergence is irreversible, which is catastrophic for the control of robot motion. Therefore, normalized asymmetric CPGs (NA-CPGs) that normalize the amplitude parameters of Hopf-based CPGs and add a constraint function and a frequency regulation mechanism are proposed. NA-CPGs can realize parameter decoupling, precise amplitude output, and stable and rapid convergence, as well as asymmetric output waveforms. Thus, it can effectively cope with large parameter changes to avoid network oscillations and divergence. To optimize the parameters of the NA-CPG model, a reinforcement-learning-based online optimization method is further proposed. Meanwhile, a biomimetic robotic fish is illustrated to realize the whole optimization process. Simulations demonstrated that the designed NA-CPGs exhibit stable, secure, and accurate network outputs, and the proposed optimization method effectively improves the swimming speed and reduces the lateral swing of the multijoint robotic fish by 6.7% and 41.7%, respectively. The proposed approach provides a significant improvement in CPG research and can be widely employed in the field of robot motion control.