Pilot Testing of Dynamic Bayesian Optimization for Exoskeleton-Assisted Training of Propulsion Mechanics

Human-in-the-loop (HIL) optimization has proven effective in assistive paradigms, but has yet to be translated to training paradigms. One challenge towards the implementation of HIL algorithms for training relates to the inability of conventional HIL optimization methods to account for neuromotor learning that can occur during training. In this work, we present our implementation of dynamic Bayesian optimization (DBO) in a single-parameter HIL optimization experiment targeting changes in propulsion mechanics - specifically an increase in the trailing limb angle (TLA). Five participants were exposed to exoskeleton-applied hip torque pulses while walking on an instrumented treadmill in user-driven walking conditions. Outcomes from this experiment were validated against an active control condition, where conventional Bayesian optimization (BO) was used. Experimental results showed significant increases in TLA in both the early and late portions of the no torque, post-training session, under the DBO paradigm, while no significant increases were observed under BO; however, no significant increases in TLA were observed during training at the group level. A linear mixed model revealed a weak, albeit significant, effect of applied torque on the change in TLA relative to baseline, suggesting that modulation of TLA is not easily achievable under this paradigm. Future work will expand the optimization paradigm to use more parameters or target different metrics of propulsion mechanics.