Abstract <p>Understanding the hierarchical organization and transmission characteristics of neural signals—from motor intention, through spinal integration, to lower limb joint motions—is critical for developing physiologically plausible, intention-driven motion generation models. Such models not only advance our comprehension of the neurophysiological mechanisms underlying lower limb locomotion but also provide a biological foundation for the direct neural control of lower limb prostheses and rehabilitation robots, offering both theoretical significance and practical utility. This challenge has persisted for decades due to the intricate coupling between spinal integration outputs (i.e., motor patterns) and joint motions. While bionic Central Pattern Generator (CPG) models primarily address the transformation from motor intention to spinal integration, they often neglect the mapping from spinal integration outputs to joint motions. Conversely, other researchers have also adopted the concept of CPG, but have employed black-box modeling approaches to directly map motor intention to joint motions, bypassing spinal integration entirely, which lack physiological interpretability and biological relevance. To address this gap, this study proposes a physiologically inspired computational model in which spinal integration is represented as a dual-layer Meta-Pattern Generator (MPG). The MPG integrates motor intention and joint feedback to generate meta-pattern signals that correspond one-to-one with joint-specific motor primitives, effectively decoupling spinal integration outputs to joint motions. This model successfully replicates the hierarchical transmission of neural signals from motor intention, through spinal processing, to joint-level execution, enabling generation of joint trajectories driven by the motor intention. Experimental validation was conducted with eight subjects performing five distinct locomotion modes. Results demonstrated a clear one-to-one correspondence between meta-patterns derived from electromyographic (EMG) signals and motor primitives extracted from joint angles. Furthermore, the joint angles predicted by the model closely matched the experimentally recorded angles, with an average error of less than 1.7%. This work provides new insights into neuromusculoskeletal modeling and offers promising applications in the design of lower limb prostheses and rehabilitation robots.</p> Graphical abstract <p></p>