Spinal reflexes with dynamic gain modulation contribute to robust walking: a comparative study of simulated neural controllers

Background Human locomotion is produced by a hierarchical motor control system that integrates predictive, long-latency supraspinal drive with reactive, short-latency spinal reflexes. While forward-dynamics simulations have explored these mechanisms with various neural controllers, quantitatively delineating the functions of these distinct neural pathways remains challenging. Methods This study used a virtual ablation framework that compared four controller architectures: supraspinal (long-latency control only), reflex (short-latency feedback only), reflex-gain (reflex feedback with dynamic gain modulation), and reflex–supraspinal (combined control). Each architecture incorporated a multilayer perceptron as the supraspinal module. The controllers were trained on a musculoskeletal model with 14 degrees of freedom and 22 Hill-type muscles, using proximal policy optimization with a reference-motion-tracking reward. Results Under unperturbed conditions, the supraspinal controller reproduced lower-limb kinematics with accuracy comparable to that of the reflex-supraspinal controller (mean RMSE ≈ 1–2°). However, when exposed to perturbations (pelvis force, floor speed), reflex-enabled controllers achieved higher survival rates while requiring less overall muscle activation than the supraspinal controller. Perturbation training altered baseline walking kinematics and increased both mean muscle activation and joint stiffness. Furthermore, as perturbation magnitude increased, the reflex–supraspinal controller exhibited increased excitation from reflex feedback. Finally, dynamic reflex-gain modulation in the combined controller demonstrated that the network successfully learned to decouple localized feedback from baseline drive, replicating phase-dependent neuromodulation patterns aligned with prior experimental findings. Conclusions The long-latency supraspinal controller possesses sufficient computational capacity to generate nominal, unperturbed walking kinematics. However, short-latency spinal reflex pathways contribute to robust locomotion; their primary functional role is to provide fast, effort-saving disturbance rejection under environmental uncertainty. The finding that human-like phase-dependent modulation emerged only after perturbation training supports the functional requirement for not only reflex pathways, but also their dynamic modulation.