An Adaptive Neural-based Model Predictive Control Framework for Autonomous Aggressive Driving

Aggressive driving presents significant challenges for autonomous vehicle control due to the highly nonlinear and dynamic nature of high-speed maneuvers. This study introduces an adaptive data-driven framework that integrates machine learning with Model Predictive Control (MPC) to achieve robust control of autonomous vehicles under aggressive driving conditions. The proposed approach consists of two main components: a learned vehicle dynamics model, approximated using a neural network, and an adaptive MPC framework that employs the learned model for control optimization. Specifically, the neural network is designed as a multi-layer perceptron and trained to predict the relative change in the vehicle state at each time step. The adaptive MPC leverages this model to generate control actions that enable tracking of aggressive driving reference trajectories by dynamically adjusting the cost function gains based on the curvature of the path. Validation results in the realistic CARLA simulator demonstrate the effectiveness of the proposed method for aggressive driving.