Desacoplamento de controle longitudinal e lateral em veículos autônomos usando controle preditivo baseado em modelo: validação usando o simulador CARLA

Abstract

This dissertation presents the comparison and validation of two trajectory control strategies for autonomous vehicles within a hierarchical planning and control architecture. In the first approach, referred to as coupled control, a Model Predictive Controller (MPC) simultaneously performs both lateral and longitudinal control actions, while in the second approach, referred to as decoupled control, the predictive controller is responsible for lateral control and a proportional– integral (PI) controller governs the vehicle’s longitudinal dynamics. In both cases, the predictive control is based on the dynamic bicycle model. The lateral controller ensures that the vehicle follows the desired trajectory, while the longitudinal controller maintains the vehicle at a predefined constant speed by generating acceleration based on the longitudinal reference velocity. The reference trajectories and velocity are generated by a trajectory planner that employs a receding-horizon control strategy based on mixed-integer quadratic programming. Validation is carried out through co-simulation in MATLAB and CARLA environments, where the controllers are implemented in MATLAB and the vehicle model is simulated in CARLA, an open-source high-fidelity simulator for autonomous driving research. The simulation experiments consider two distinct scenarios. In the first scenario, the maneuver of a vehicle traveling along a one-way road in the presence of a static obstacle is evaluated. In the second scenario, the maneuver is evaluated in the presence of a dynamic obstacle moving at a constant longitudinal speed. The results show that the coupled strategy achieves better lateral trajectory tracking, especially at higher speeds, with a root mean square error (RMSE) of 0.084 m compared to 0.089 m for the decoupled method. However, it is less accurate in longitudinal speed control, with an RMSE of 1.32 m compared to 1.21 m. The computational performance analysis indicates that both approaches meet real-time requirements and are computationally feasible, with the decoupled method exhibiting a slightly lower solution time (2.6 ms versus 2.7 ms).