Leveraging Modular Architectures and End-to-End Learning for Autonomous Driving in Unmapped Environments

• Ano de defesa: 2024
• Autor(a) principal: Rosero, Luis Alberto Rosero
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Biblioteca Digitais de Teses e Dissertações da USP
• Palavras-chave em Português: Arquitetura híbrida; Arquitetura modular; Autonomous driving; CARLA simulator; Condução autônoma; Decision-making; Disparidade; Disparity; End-to-end; Hybrid architecture; Intelligent and autonomous vehicles; Modular architecture; Path planning; Percepção; Perception; Planejamento de trajetória; Predição; Prediction; Simulador CARLA; Tomada de decisão; Veículos autônomos e inteligentes
• Link de acesso: https://www.teses.usp.br/teses/disponiveis/55/55134/tde-09082024-163642/

Autonomous driving promises a revolution in transportation, unlocking significant social and economic benefits. Despite notable advancements in autonomous vehicle technology tailored for mapped environments, navigating in unmapped areas remains a persistent challenge. The limited utilization of developments in modular pipelines exacerbates this issue, impeding progress towards map-free navigation. This thesis delves into the development of autonomous driving architectures for real-time navigation, both with and without maps. Three approaches are proposed, implemented, compared, and evaluated to create new and robust methodologies: Modular Pipeline: A custom agent performs data collection and serves as a baseline for mapbased navigation. Traditional algorithms and new modules for perception, decision-making, and prediction are integrated to ensure safe navigation in mapped environments. This agent acts as the \"teacher\" for the mapless navigation agents. End-to-End Learning: Neural networks learn driving policies from data through imitation learning techniques. Simplicity is prioritized for real-time operation in map-free environments. Different sensor types and fusion methods are explored to enhance performance. Hybrid Architecture: Combining the interpretability of modular systems with the learning capabilities of end-to-end models, this approach integrates data-driven path planning with modular perception and control modules. It offers robustness, flexibility, and adaptability. Furthermore, a ROS-based framework named \"CaRINA agent\" is developed to implement modular pipelines and facilitate incorporating end-to-end methods and constructing hybrid architectures. To comprehensively evaluate our methodologies, we leverage the CARLA Leaderboards, achieving competitive results in both Leaderboard 1 and Leaderboard 2, specifically ranking among the top in the SENSORS and MAP categories. Moreover, our modular architecture and hybrid agent secured 1st and 2nd place in the 2023 CARLA Autonomous Driving Challenge (CADCH), underscoring the effectiveness of our proposed approaches.