Abstract:
|
Robotic prostheses require a controller to decode muscle contractions, measured by electromyogram (EMG) signals, into the user’s desired hand movement. While EMG data collected across hundreds of sensors is informative, it is inefficient and impractical. Therefore, choosing the optimal locations for EMG sensor placement is an important problem, but there are few methods available to perform this selection. In this talk, we overview existing controllers based on pattern recognition of summarized EMG data, and argue that controllers that treat the EMG data as functional covariates are more appropriate. With this perspective, we introduce new methods based on linear and nonlinear functional additive models that are able to perform simultaneous EMG sensor selection and prosthesis-controller tuning. These methods are able to overcome challenges unique to high-density EMG data, like high cross-correlation, and are shown to outperform the prediction performance of existing methods.
|