The coordinated motion of the spine and lower limbs is crucial for humanoid robots to achieve stable and efficient walking. To address the heterogeneous motion requirements of these two body parts, this study proposes a heterogeneous Central Pattern Generator (CPG) network. The network integrates the Kimura neuron model and cosine oscillator model to generate end?effector and center-of-mass trajectories, as well as spinal pitch and yaw motion profiles. A linear mapping function is designed to realize the transformation from biological rhythmic signals to the robot’s task space. To further enhance the performance of the CPG network, the Balanced Particle Swarm Optimization (BPSO) algorithm is adopted, which effectively improves search efficiency by coordinating exploration and exploitation in a phased manner. Based on this, a fitness function incorporating Zero Moment Point (ZMP) deviation, walking distance, and lateral offset is constructed to optimize the 22-dimensional CPG parameters. Both virtual prototype simulations and physical experiments demonstrate that the proposed heterogeneous multilayer network achieves coherent coordination between the spine and legs. The optimized CPG generates smooth, phase?alternating foot trajectories with a maximum ZMP deviation of only 2.5 cm. By introducing spinal motion, the Cost of Transport (COT) of the lower-limb joints is reduced by 58%, significantly improving energy utilization efficiency.