Reinforcement Learning of Repetitive Tasks for Autonomous Heavy-Duty Vehicles

Sammanfattning: Many industrial applications of heavy-duty autonomous vehicles include repetitive manoeuvres, such as, vehicle parking, hub-to-hub transportation etc. This thesis explores the possibility to use the information from previous executions, via reinforcement learning, of specific manoeuvres to improve the performance for future iterations. The manoeuvres are; one straight line path, and one constantly curved path. A proportional-integrative control strategy is designed to control the vehicle and the controller is updated, between each iteration, using a policy gradient method. A rejection sampling procedure is introduced to impose the stability of the control system. This is necessary since the general reinforcement learning framework and policy gradient framework do not consider stability. The performance of the rejection sampling procedure is improved using the ideas of simulated annealing. The performance improvement of the vehicle is evaluated through simulations. Linear and nonlinear vehicle models are evaluated on a straight line path and a constantly curved path. The simulations show that the vehicle improves its ability to track the reference path for all evaluation models and scenarios. Finally, the simulations also show that the controlled system is kept stable throughout the learning process.