A simple pendulum model is used to study how feedforward and feedback can be combined to control rhythmic limb movements. I show that a purely feedforward central pattern generator (CPG) is highly sensitive to unexpected disturbances. Pure feedback control analogous to reflex pathways can compensate for disturbances but is sensitive to imperfect sensors. I demonstrate that for systems subject to both unexpected disturbances and sensor noise, a combination of feedforward and feedback can improve performance. This combination is achieved by using a state estimation interpretation, in which a neural oscillator acts as an internal model of limb motion that predicts the state of the limb, and by using alpha-gamma coactivation or its equivalent to generate a sensory error signal that is fed back to entrain the neural oscillator. Such a hybrid feedforward/feedback system can optimally compensate for both disturbances and sensor noise, yet it can also produce fictive locomotion when sensory output is removed, as is observed biologically. CPG behavior arises due to the interaction of the internal model and a feedback control that uses the predicted state. I propose an interpretation of the neural oscillator as a filter for processing sensory information rather than as a generator of commands.
A.D. Kuo is with the Department of Mechanical Engineering at the University of Michigan, Ann Arbor, MI 48109-2125.