zannunakiz/Q1_Research_DQN-autonomous-vehicles: Q1 DQN_AV V1.0.0

Abstract

Safety-Aware DQN Variants for Lightweight Sensor-Based Collision Avoidance in Autonomous Vehicles This repository contains the official research software, experiment code, and datasets supporting the manuscript "Safety-Aware DQN Variants for Lightweight Sensor-Based Collision Avoidance in Autonomous Vehicles." Paper Information Paper Title: Safety-Aware DQN Variants for Lightweight Sensor-Based Collision Avoidance in Autonomous Vehicles Authors: Dhidik Prastiyanto, Richky Abednego, Muhamad Kurniawan Fauzi, Budi Sunarko, Khoirudin Fathoni, Anton Satrio Prabuwono, and Meida Andini Rahmawati Corresponding Author: Richky Abednego Contact: richky.abednego@gmail.com Overview This project systematically evaluates three value-based Deep Reinforcement Learning variants—DQN, Double DQN, and Dueling Double DQN—for autonomous vehicle collision avoidance in a structured three-lane environment. Designed for efficiency, the framework relies on a low-cost, camera-free sensor configuration consisting of just seven distance sensors and normalized vehicle speed. The codebase provides a complete pipeline for training, evaluation, and safety-aware obstacle avoidance. It features mastery-based curriculum learning, safety-oriented reward shaping, and rigorous robustness testing against unseen obstacle layouts, sensor noise, reduced obstacle gaps, and out-of-distribution speeds. This release is intended to support the reproducibility, citation, and archival of our research implementation. Main Features Value-Based DRL Framework: Implementation of standard DQN, Double DQN, and Dueling Double DQN architectures. Lightweight State Representation: Camera-free navigation using a compact 7-sensor array. Safety-First Design: Custom safety-aware reward shaping for collision avoidance. End-to-End Pipeline: Complete scripts for training, evaluation, ablation studies, and statistical analysis. Robustness Testing: Pre-configured simulation environments to test against varied speeds, sensor noise, and complex unseen layouts. Citation If you use this software or find our research helpful in your work, please cite the associated Zenodo DOI and the related research article: @article{prastiyanto2026safety, title={Safety-Aware DQN Variants for Lightweight Sensor-Based Collision Avoidance in Autonomous Vehicles}, author={Prastiyanto, Dhidik and Abednego, Richky and Fauzi, Muhamad Kurniawan and Sunarko, Budi and Fathoni, Khoirudin and Prabuwono, Anton Satrio and Rahmawati, Meida Andini}, journal={ACM Transactions on Intelligent Systems and Technology}, year={2026}, note={Under Review} }