This paper presents an approach that improves the efficiency and generalization capabilities of Reinforcement Learning-based autonomous vehicles operating in urban driving scenarios. The proposed method introduces an Efficient Social-based Motion Prediction module, which predicts the future positions of vehicles within the scenario. These predictions serve as input to a Reinforcement Learning-based Decision-Making module, responsible for executing high-level actions. The Proximal Policy Optimization algorithm is employed to develop our approach. We conduct experiments in an unsignalized T-intersection scenario using the SMARTS framework, comparing our approach with and without the proposed state representation, as well as against various baseline methods. Through this study, we demonstrate that our approach achieves performance improvements, particularly in scenarios involving high velocities.
@article{gutierrez2023augmented,
author = {Gutiérrez-Moreno, Rodrigo and Gómez-Huelamo, Carlos and Barea, Rafael and López-Guillén, Elena and Arango, Felipe and Bergasa, Luis Miguel},
title = {Augmented Reinforcement Learning with Efficient Social-Based Motion Prediction for Autonomous Decision-Making},
journal = {IEEE 26th International Conference on Intelligent Transportation Systems (ITSC)},
year = {2023},
}