Development of an Advanced Autonomous Vehicle Controller System using Nonlinear Model Predictive Control (NMPC) and Long Short-Term Memory (LSTM)

This study presents the implementation of an advanced vehicle controller-localization system that integrates Nonlinear Model Predictive Control (NMPC) with a Long Short-Term Memory (LSTM) network to improve autonomous vehicle path tracking performance. The proposed system utilizes sensor fusion approach of Visual-Inertial Odometry (VIO) and wheel odometry data through an Extended Kalman Filter (EKF) to provide accurate Three-Degree-Of-Freedom (3-DOF) for vehicle pose estimation in real time. The work was further implemented within a Robot Operating System (ROS)-based CARLA simulation environment and the LSTM model predicts future vehicle states over a finite horizon. Experimental results from the implementation show that the LSTM achieved low training and validation losses of 0.0148 and 0.0176, respectively and Mean Absolute Errors (MAE) of 0.08ms in position and 0.04 radians in heading prediction. Furthermore, the integrated NMPC-LSTM controller demonstrated an 18.6% reduction in trajectory tracking error (95% CI: 16.5% - 20.7%, p < 0.01) compared to the baseline NMPC along with improved robustness under dynamic and noisy conditions. These findings validate the effectiveness of combining deep learning with model predictive control for enhancing autonomous vehicle navigation.