User preference optimization for control of ankle exoskeletons using sample efficient active learning

A major challenge to the widespread success of augmentative exoskeletons
is accurately adjusting the controller to provide cooperative assistance
with their wearer. Often, the controller parameters are
``tuned'' to optimize a physiological or biomechanical
objective. However, these approaches are resource-intensive, while
typically only enabling optimization of a single objective. In reality,
the exoskeleton user experience is derived from many factors, including
comfort and stability, among others. This work introduces an approach to
conveniently tune four parameters of the exoskeleton controller that
maximize user preference. We use an evolutionary algorithm to recommend
potential parameters, which are ranked by a neural network that is
pre-trained with previously collected preference data. The controller
parameters that have the highest preference ranking are provided to the
exoskeleton, and the wearer provides feedback as forced-choice
comparisons. Our approach was able to converge on controller parameters
preferred by the wearer compared to randomized parameters with an accuracy
of 88% on average. The result indicates that the proposed algorithm was
able to identify users' preferences while requiring less than 50
queries to users. This work demonstrates user preference can be used to
tune high-dimensional controller spaces easily and accurately, which shows
the potential of translating lower-limb wearable technologies into our
daily lives.