Neural Network based Adaptive Estimation and Guidance: Application to 2D Obstacle Avoidance

Abstract

In this paper, an adaptive estimation method is presented for use in the control of a class of multivariable systems with unmodeled dynamics coupled to an otherwise known model for the process. The adaptive estimator takes the form of augmenting an extended Kalman filter designed for the system without unmodeled dynamics. The design of the adaptive element employs a linearly parameterized neural network. The states of the adaptive estimator are used to form the feedback controller. The neural network weights are adjusted on line using a linear observer for the nominal system's error dynamics. Boundedness of the error signals is proven using Lyapunov's direct method. To show the efficacy of the approach, the problem of obstacle avoidance in two dimensions is considered, where the goal is for an unmanned vehicle to traverse a nominal path while avoiding obstacles in the course of travel.