Deep Learning Enables Exoboot Control to Augment Variable-Speed Walking

Abstract

Ankle exoskeletons have the potential to improve mobility, but common controllers are often inflexible to variations in tasks, such as changes in walking speed. To enable effective variable-speed exoboot control, we developed and validated a two-headed convolutional neural network trained to (1) classify stance versus swing and (2) predict the phase during stance, which was then mapped to a desired exoboot torque. This Machine Learning Estimator (MLE) was trained from nine participants walking at three speeds and four exoboot assistance levels. A Time-Based Estimator (TBE) that predicted gait phase from the two previous stride durations was used to apply realistic torques during MLE training and served as a within–participant control condition. The MLE was validated online with three new participants walking at a range of speeds and torques, both interpolating within and extrapolating outside the training set. Online validation accuracy (RMSE) across tested speeds and torque levels was 3.9%. On a simple walking task in which treadmill speed was varied sinusoidally between 1.1 and 1.6 m/s with a 30 s period, the three participants exhibited a mean 5.2% decrease in metabolic expenditure with the MLE compared to no-exo (boots only), but exhibited a 5.4% increase when walking with the TBE. The MLE more accurately predicted heel strike and toe off events (heel strike Mean Absolute Error: 9.6 ms; toe off MAE: 13.2 ms) than the TBE (heel strike MAE: 19.1 ms; toe off MAE: 34 ms). These positive results validated the potential of using a deep learning model for gait state estimation to effectively control an ankle exoskeleton across variable walking speeds.