Gaussian Process Regression Models for On-Line Ankle Moment Estimation in Exoskeleton-Assisted Walking

Ankle moment estimators inform the controllers of several assistive exoskeletons being developed in research labs. Accurate moment estimations are critical to ensure biome-chanically relevant assistance. In this work, we propose new subject-agnostic ensemble Gaussian Process Regression (GPR) models which rely on a minimal set of in-shoe force and inertial sensors that do not require precise sensor-to-body alignment. We systematically analyzed the effects of model type, sensor set, and phase variable in terms of estimation accuracy by carrying out treadmill tests with 15 healthy individuals across a wide range of walking speeds. Our best ensemble GPR model achieved an average root-mean-square error of 3.6%±1.2% normalized over the gait cycle (equivalent to 8.8%±1.6% when normalized over the stance phase). Incorporating data from the inertial sensor and using the stance phase as the phase variable independently contributed to superior accuracy. Overall, these results indicate the potential of the proposed ensemble GPR models to accurately estimate ankle moments, paving the way for future applications to assistive powered ankle exoskeletons in real-world environments.