Training Reflexes Using Adaptive Feedforward Control

We consider the problem of mixed feedforward and feedback based disturbance rejection, where the feedforward measurement only provides a partial reconstruction of the disturbance. In doing so, we pose a new biologically relevant disturbance rejection problem which puts the role of feedforward measurements at the forefront. Based on the architecture of the human brain, we propose a design that utilizes an adaptive internal model operating on a fast timescale that, in turn, trains the correct feedforward gains on a slow timescale. As such, the training of reflexes in biological systems can be explained by leveraging the theory of adaptive feedforward control. It is proven that our design provides an arbitrary level of disturbance attenuation, and the benefits of using reflexes are illustrated via a multitude of simulations.