Motor control in a meta-network with attractor dynamics

A neural-network module with attractor dynamics has been shown recently to be robust to stochastic noise in external and internal signals, and to converge rapidly onto an output signal that is an unbiased and efficient representation of the environment. We suggest here a modular network architecture with attractor dynamics that can compute the time-varying signals that are presumably required to control arm movements. The architecture is composed of several linked modules and implements the joint torque control of a planar biomechanical model of the arm, in the presence of external dynamic perturbations. The meta-network is robust to noise and to the unreliable availability of some signals and can provide feedback correction for unexpected external perturbations.