Unsupervised Coadaptation of an Assistive Interface to Facilitate Sensorimotor Learning of Redundant Control

Abstract

Interfaces that exploit biological signals or movements to control the operation of lower-dimensional systems external to the body are at the frontier for augmenting human abilities, but also constitute a learning challenge for their users. We developed and tested an unsupervised coadaptive algorithm that changed the mapping of a body machine interface to match the natural movement distribution of the users. Users controlled a cursor on a computer monitor using arm and shoulder motions captured by a set of inertial sensors in either of three conditions: i) a constant body-to-cursor map obtained through Principal Component Analysis of calibration movements, ii) a map that was recomputed at specified points in time, iii) a map that adaptively changed over time. We used recursive online PCA to incrementally shift the projection space towards the 2-dimensional subspace capturing the greatest sensor signal variance. Results suggest that training with the coadaptive BMI allows for faster internalization of the control space while reducing user's reliance on visual feedback.