Optimizing Reflex-Based Neuromusculoskeletal Walking Model on Rough Terrain Reveals Increased Robustness and Key Stabilizing Reflexes

Locomotion on rough terrain is one of the major challenges for bipedal robots, exoskeletons and prostheses. To rapidly counter unexpected perturbations humans rely on reflexive strategies. Reflexes have been leveraged in controlling orthoses or bipedal robots by using a reflex-based neuro-musculoskeletal model. However, the robustness of the used model to uneven ground is limited to step-down perturbations of max. 3 cm. Different extensions of the model have increased its robustness to unexpected step-down perturbations. In this work, we target the same perturbations, but instead of extending the model, we optimize its control parameters for walking on terrain with unexpected step-down perturbations. We find that without any sextension of the control, the model can cope with rough terrain with unexpected perturbations of up to 10 cm. A comparison of the control parameters reveals the key stabilizing reflexes. Namely, two reflexes exciting hip muscles based on pelvis tilt as well as a reflex increasing ankle dorsiflexion are adapted the most. These results show that the full potential of control approaches can only be unveiled when tuned to realistic environmental conditions and suggest adopting this approach for parameter tuning. Applications employing the reflex-based control architecture should consider adapting the key parameters we found to obtain greater robustness.