End-to-end learning for autonomous vehicles: a narrow approach

• Ano de defesa: 2023
• Autor(a) principal: Heringer, Adauton Machado
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Biblioteca Digitais de Teses e Dissertações da USP
• Palavras-chave em Português: Aprendizado end-to-end; Artificial general intelligence; Autonomia restrita; Autonomous vehicles; Convolutional neural networks; End-to-end learning; Imaginário sociotécnico; Inteligência artificial geral; Narrow autonomy; Redes neurais convolucionais; Sociotechnical imaginary; Veículos autônomos
• Link de acesso: https://www.teses.usp.br/teses/disponiveis/45/45134/tde-19072023-053510/

Autonomous vehicles are long promised to revolutionize our civilization. Nevertheless, it has consistently failed to meet expectations in the past two decades. Based on the fundamental difference between narrow and general artificial intelligence and equipped with the theoretical approach of sociotechnical imaginaries, we criticize general autonomy: the study of autonomous vehicles as envisaged by its artificially fabricated sociotechnical imaginary utopia. By contrast, we conceptualize narrow autonomy as the study of context-limited autonomous vehicles. Accordingly, we propose a narrow approach: instead of training a vehicle in a context-free environment, we set clear boundaries for the path the vehicle is supposed to drive. Using the latest advancements in end-to-end deep learning, we trained a convolutional neural network to map images and high-level commands straight to vehicle control, such as steering angle, throttle, and brake, in a simulated environment. Although this is a multidisciplinary conceptual work, our results indicate that by delimiting its path we can significantly improve performance and contribute to the advancements of autonomous technology.