Establishing Personalized and Sparse Spinal Reflex Circuitry From Locomotion Data

Spinal reflex circuitry-based controllers have shown promising capabilities in controlling the biomechanics of locomotion in simulation. Studies have been done to tune or identify reflex circuitry from imposed objectives, such as metabolic energy efficiency, maintaining balance or mimicking experimental data. However, in those works, the reflex structure was predefined by the authors. This may limit the generalizing capabilities of such circuits and prevent a better understanding of the dominant reflex loops in human locomotion. In this work we propose an identification framework that can identify personalized sparse-structure reflex circuitry directly from experimental data. This is achieved by first performing muscle model personalization via optimization and consequently extensive model regularization when optimizing a densely connected neural reflex network. Trajectory optimization is used in both steps, namely direct collocation formulated as a nonlinear program. Multi-speed walking and running data of five subjects was used to test the framework. Results show that personalized sparse reflex circuitry was identified that show exceptional torque and muscle activation reproduction for multiple walking speeds with a speed-independent, but phase-dependent, reflex network. Furthermore, similar sparse structures were found between subjects, but different structures between walking and running gait. However, due to the use of data of unperturbed locomotion, the actual plausibility of the found circuitry remains an open question. The controllers use slightly more reflex gain parameters than those chosen and designed in other studies, but the benefit of our method is that the structures were found automatically.