More and more biological evidences have been found that neural networks in the spinal cord, referred to as “central pattern generators” (CPGs), govern locomotion. CPGs are capable of producing rhythmic movements, such as swimming, flying, and walking, even when isolated from the brain and sensory inputs. If we could build up any models that have similar functions as CPGs, it will be much easier to design better locomotion for robots. In this paper, a self-training environment is designed and through genetic algorithm (GA), walking trajectories for every foot of AIBO are generated at first. With this acquired walking pattern, AIBO gets its fastest locomotion speed. Then, this walking pattern is taken as a reference to build CPGs with Hopf oscillators. By changing corresponding parameters, the frequencies and the amplitudes of CPGs’ outputs can be adjusted online. The limit cycle behavior of Hopf oscillators ensures the online adjustment and the walking stability against perturbation as well. This property suggests a strong adaptive capacity to real environments for robots. At last, simulations are carried on in Webots and verify the proposed method. More and more biological evidences have been found that neural networks in the spinal cord, referred to as “central pattern generators” (CPGs), govern locomotion. CPGs are capable of producing rhythmic movements, such as swimming, flying, and walking, even when isolated from the brain and sensory inputs. If we could build any any that that similar similar functions as CPGs, it will be much easier to design better locomotion for robots. In this paper, a self-training environment is designed and through genetic algorithm (GA), walking trajectories for every foot of Afoot are for first foot. With this acquired walking pattern, AIBO gets its fastest locomotion speed. parameters, the frequencies and the amplitudes of CPGs’ outputs can be adjusted online. The limit cycle behavior of Hopf oscillators ensures the online adjustment and the walking stability against perturbation a This property suggests a strong adaptive capacity to real environments for robots. At last, simulations are carried on in Webots and verify the proposed method.