Sensitivity of Iterative Learning Control to Varying Initial Conditions for Gait Assistance.

Iterative Learning Control (ILC) is a promising method for adapting neuromuscular electrical stimulation to facilitate independent walking after upper motor neuron paralysis. However, assumptions made by conventional ILC methods, such as identical initial conditions for each iteration, are unsustainable in the case of human gait. In this study, we implement a musculoskeletal model of a single leg to analyze the consequences of variable initial conditions for data-driven ILC-based stimulation (DDILC) during swing phase of gait. We show that DDILC converges in all tested cases of initial hip angle variability, but that noise arises because of such variability. For the maximum variability case, the output with the largest noise (i.e., the terminal ankle angle) had a standard deviation of 2.1 degrees. Thus, the noise due to initial hip angle variability is shown to be comparable in magnitude to the inherent variability in gait. We also show that exploding gradients and instability eventually occur because of varying initial conditions, but these can be mitigated with established techniques.