Online learning of MPC for autonomous racing

12/13/2021

In this dissertation, a Learning based Model Predictive Control (LMPC) architecture is designed for the control of a Formula Student (FS) autonomous vehicle. For the implementation of this controller in real time to satisfy the FS driverless requirements, the C++ programming language is used and the MPC's optimization problem is solved using a commercial solver. In summary, the developed controller is able to iteratively learn as the vehicle drives itself. This learning process is carried out for two distinct purposes: improving the accuracy of the vehicle model used by the controller and automatically finding the controller parameters that result in the fastest lap times. Finding the mathematical equations that fully describe the race car dynamics requires the use of highly nonlinear vehicle nominal models which are difficult to obtain. For this purpose, an Artificial Neural Network (ANN) is added to a vehicle nominal model in order to correct for unmodeled dynamics not considered in the nominal model. The ANN is trained in an online Supervised Learning (SL) approach, which learns based on past model prediction errors. Furthermore, the controller's parameters are tuned in a Reinforcement Learning (RL) environment in order to find the set of parameters that iteratively allow for faster lap times. In a simulation environment, various tests on three different tracks are performed. Moreover, it is shown that by employing these two learning procedures, the full control algorithm is able to reduce lap times up to 16.5%.

Authors in the community:

• 

  G

  Gabriel Alexandre Francisco Costa

  ist187194

Supervisors of this institution:

• 

  M

  Miguel Afonso Dias de Ayala Botto

  ist12889

• 

  P

  Pedro Manuel Urbano de Almeida Lima

  ist12116

mechanical-engineering - Mechanical engineering

por - Portuguese

Instituto Superior Técnico