6 - The Neural Dynamics Approach to Sensory-Motor Control: Overview and Recent Applications in Mobile Robot Control and Speech Production

Abstract

ABSTRACT This chapter discusses a collection of models that utilize adaptive and dynamical properties of neural networks to solve problems of sensory-motor control for biological organisms and robots. The chapter begins with an overview of several unsupervised neural network models developed at the Center for Adaptive Systems during the past decade. These models have been used to explain a variety of data in research areas ranging from the cortical control of eye and arm movements to spinal regulation of muscle length and tension. Next, two recent models that build on important concepts from this earlier work are presented. The first of these models is an adaptive neural network controller for a visually guided mobile robot. The neural network controller enables the robot to move to arbitrary targets without any knowledge of the robot's kinematics, immediately and automatically compensating for perturbations such as target movements, wheel slippage, or changes in the robot's plants. The controller also adapts to long-term perturbations, enabling the robot to compensate for statistically significant changes in its plant. The second model is a self-organizing neural network addressing speech motor skill acquisition and speech production. This model explains a wide range of data on contextual variability, motor equivalence, coarticulation, and speaking rate effects. Model parameters are learned during a babbling phase, using only information available to a babbling infant. After learning, the model can produce arbitrary phoneme strings, again exhibiting automatic compensation for perturbations or constraints on the articulators. Finally, other recent models using a neural dynamics approach are summarized and future research avenues are outlined.