Autonomous driving using an optimized neural network based adaptive LPV-MPC controller

Abstract: Driverless vehicles are complex systems operating in constantly changing environments. Automated driving is achieved by controlling the coupled longitudinal and lateral vehicle dynamics. Model predictive control is one of the most promising tools for this type of application due to its optimal performance and ability to handle constraints. This paper addresses autonomous driving with an adaptive linear parameter varying model predictive controller (LPV-MPC), which is adapted by a neural network and optimized by an improved Genetic Algorithm. The proposed controller is evaluated on a challenging track under variable wind disturbance. Code can be found here: https://github.com/yassinekebbati/GA-optimized-MLP-based-LPV_MPC

• The paper introduces an adaptive LPV-MPC controller enhanced with neural networks and GA optimization for superior vehicle control.
• It integrates a bicycle model in an LPV framework to manage nonlinear vehicle dynamics effectively.
• Extensive simulations demonstrate reduced RMSE in tracking and robust performance under varying environmental conditions.

Autonomous Driving Using an Optimized Neural Network Based Adaptive LPV-MPC Controller

Introduction

The research paper introduces a novel approach for controlling autonomous vehicles using an optimized neural network-based adaptive LPV-MPC (Linear Parameter Varying Model Predictive Control) controller. The autonomous driving field aims to replace human drivers through sophisticated automatic control systems. This paper specifically focuses on integrating machine learning techniques, genetic algorithms, and control theory to enhance the performance of autonomous vehicles. The study positions itself within the extensive literature on vehicle dynamics, tackling both lateral and longitudinal motion control simultaneously, a challenging endeavor due to the complex dynamics involved.

Vehicle Modeling

The vehicle dynamics are modeled using the bicycle model, which provides a simplified yet effective representation of the vehicle for control purposes. This model considers longitudinal, lateral, and yaw dynamics, capturing the essential characteristics required for model-based predictive control. The LPV framework facilitates capturing the nonlinearities of the system within a linear framework by employing state-dependent transformation matrices.

Controller Design

The central contribution of this paper is the development of an adaptive LPV-MPC approach augmented by neural networks and optimized via an improved Genetic Algorithm (GA).

LPV-MPC Framework

The LPV-MPC design utilizes the bicycle model reformulated into an LPV form for control. The controller solves an optimization problem to achieve optimal control actions over a prediction horizon. The scheduling vector, a set of system state-dependent variables, modifies the system matrices dynamically, enabling the controller to handle nonlinearities effectively.

Neural Network Adaptation

The adaptation of the controller employs a neural network to learn the cornering stiffness coefficients of the tires, allowing the system to adaptively refine its predictions. This approach leverages machine learning to enhance prediction precision in real-time, especially during high-slip or fast maneuvering, where conventional linear tire models fail.

Genetic Algorithm Optimization

The proposed algorithm optimizes the controller's cost function parameters using an improved GA, which integrates a combination of RWS (Roulette Wheel Selection) and TS (Tournament Selection) methods for optimal population selection. This strategy improves convergence rates and avoids premature convergence, showcasing advantages over traditional selection methods.

Results and Discussion

In extensive simulations using a high-fidelity dynamic model, the adaptive LPV-MPC controller demonstrated superior performance in both trajectory and speed tracking compared to a baseline LMPC system. The optimizations achieved through GA yielded significantly reduced RMSE in both heading and velocity tracking. The evaluations under variable wind conditions confirmed the controller's robustness and adaptability.

The neural network's ability to predict tire dynamics with high accuracy was validated through training on Carsim-generated data. With a predictive model showcasing high $R^2$ scores, the system demonstrated effective adaptation capabilities.

Conclusions

The study presents a comprehensive solution for enhancing autonomous vehicle control, integrating adaptive LPV-MPC with state-of-the-art optimization and learning techniques. It achieves high tracking accuracy and robustness against variable environmental conditions, making a significant contribution to the field of autonomous vehicle control systems. Future directions involve the development of an online learning-based LPV-MPC, which could continuously adapt to changing dynamics in real-time, further enhancing autonomy levels in vehicle control systems.

This paper provides a foundation for further research into adaptive and learning-based vehicle control systems, with potential practical applications in both civilian and commercial autonomous vehicle development.