A Trajectory Simulation Approach for Autonomous Vehicles Path Planning using Deep Reinforcement Learning

Autonomous vehicle path planning aims to allow safe and rapid movement in an environment without human interference. Recently, Reinforcement Learning methods have been used to solve this problem and have achieved satisfactory results. This work presents the use of Deep Reinforcement Learning for the task of path planning for autonomous vehicles through trajectory simulation, to define routes that offer greater safety (without collisions) and less distance for the displacement between two points. A method for creating simulation environments was developed to analyze the performance of the proposed models in different difficult degrees of circumstances. The decision-making strategy implemented was based on the use of Artificial Neural Networks of the Multilayer Perceptron type with parameters and hyperparameters determined from a grid search. The models were evaluated for their reward charts resulting from their learning process. Such evaluation occurred in two phases: isolated evaluation, in which the models were inserted into the environment without prior knowledge; and incremental evaluation, in which models were inserted in unknown environments with previous intelligence accumulated in other conditions. The results obtained are competitive with state-of-the-art works and highlight the adaptive characteristic of the models presented, which, when inserted with prior knowledge in environments, can reduce the convergence time by up to 89.47% when compared to related works