• automotive
• electronic engineering
• embedded platforms
• model predictive control

This master thesis work proposes an implementation and analysis of a non-linear Model Predictive Control algorithm for automotive applications.
The algorithm presented provides to a vehicle the ability to avoid obstacles on his direction and to track a trajectory, acquired through sensors on-board, respecting the constraints imposed on parameters like speed and steering manoeuvre. These features are some of the mainly ones in the field of autonomous driving, field that it is becoming always more present and important for the automotive companies, that are producing more and more vehicles equipped with ADAS (Autonomous Driving Assistance System) to simplify and make safer the guide for the drivers.
In this context, the research to improve the autonomous driving functions has introduced the problem of computational performances, due to the complexity of algorithms and methods exploited, since the classic Electronic Control Units (ECUs), present on most of vehicles, cannot satisfy a real-time resolution. Precisely in order to address this problem, this work proposes an analysis of algorithm performances related to the porting made on a heterogeneous system (MCU plus FPGA accelerator), in order to guarantee high performances and functional safety.
The MPC algorithm implementation relies on the GRAMPC – GRAdient based library, a C-code library implemented for nonlinear MPC problems and that can be efficiently used for embedded control of nonlinear and highly dynamical systems; GRAMPC exploits an augmented Lagrangian method to solve the ODEs math introduced in the optimal control problem evaluation.
Due to the collaboration between the DII of University of Pisa and Maserati Spa, the vehicle chosen, as plant model for the algorithm implementation and for the parameters sizing, is a Maserati LEVANTE S, in order to provide a more realistic behaviour for the project evaluation.
The project relies on a iterative Model Predictive Control algorithm, using a 6-states vehicle model, that implements a trajectory tracking and obstacle avoidance ability; the goal path is provided into the algorithm assuming that in real applications the desired trajectory and the constraints can be provided by the navigation system (using cameras, radar, lidar, GPS sensors…).
In addition, the vehicle was modelled also using 2 possible actuators controls to provide a correct dynamic behaviour: acceleration and front steer angle.
The analysis of the performances and the correct functioning of the algorithm is provided through the simulations obtained from the porting on the Xilinx Zynq UltraScale + MPSoC ZCU104 board, exploiting the embedded software development flow provided by the Xilinx software Vivado and Vitis.
For a better performance analysis, with a main focus on the computational time performances, two different simulations are proposed in order to exploit both the CPUs provided by the ZCU104 MPSoC: one simulation ported on the Cortex A53 application processor and the other ported on the Cortex R5 Real Time processor.
To prove the robustness of the algorithm to external disturbances or to a possible measure error of the environment acquiring system, it’s proposed a simulation with the introduction of an Additive White Noise Gaussian generator to mimic the effect of random processes that may occur during the vehicle movement.
For all the proposed simulations are presented and discussed feedback plots to provide a better understanding of the algorithm functionality; graphical results of vehicle movement, all the characteristic model states, controls and main algorithm parameters are reported. A particular focus is posed on the execution time, that represent the time needed by the algorithm to solve the MPC iterations and allows to verify if the real time constraints are respected for the system in exam.