zannunakiz/Q1research_DQN_AutonomousVehicles: DQN_AV V1.0.0

Abstract

Safety-Aware DQN Variants for AV Collision Avoidance This release contains the research software and experiment code supporting the study "Safety-Aware Deep Q-Network Variants for Lightweight Sensor-Based Collision Avoidance in Autonomous Vehicles." The project compares three Deep Reinforcement Learning variants, namely DQN, Double DQN, and Dueling DQN, for lightweight autonomous vehicle collision avoidance in a structured three-lane road environment. The system uses a low-cost, camera-free sensor configuration consisting of seven distance sensors and normalized vehicle speed as state inputs. The code includes the training and evaluation workflow for safety-aware obstacle avoidance, including reward shaping, model comparison, hyperparameter sensitivity analysis, and robustness testing under unseen obstacle layouts, sensor noise, elevated obstacle density, and out-of-distribution obstacle speeds. This release is intended to support reproducibility, citation, and archival of the research implementation. Main Features DQN, Double DQN, and Dueling DQN comparison Lightweight sensor-based autonomous navigation Safety-aware reward shaping Training and evaluation pipeline Robustness testing under multiple driving conditions Experiment support for collision avoidance research Citation If you use this software, please cite the associated Zenodo DOI and related research article.