Robust EMRAN based Neural Aided Learning Controller for Autonomous Vehicles

This paper presents an online evolving neural network-based inverse dynamics learning controller for an autonomous vehicles' longitudinal and lateral control under model uncertainties and disturbances. The inverse dynamics of the vehicle is approximated using a feedback error learning mechanism that utilizes a dynamic Radial Basis Function neural network, referred to as the Extended Minimal Resource Allocating Network (EMRAN). EMRAN uses an extended Kalman filter approach for learning and a growing/pruning condition helps in keeping the number of hidden neurons minimum. The online learning algorithm helps in handling the uncertainties and dynamic variations and also the unknown disturbances on the road. The proposed control architecture employs two coupled conventional controllers aided by the EMRAN inverse dynamics controller. The control architecture has a conventional PID controller for cruise control and a Stanley controller for path-tracking. Performances of both the longitudinal and lateral controllers are compared with existing control methods and the results clearly indicate that the proposed control scheme handles the disturbances and parametric uncertainties better, and also provides better tracking performance in autonomous vehicles.