Autonomous driving: learning to make decisions in uncertain environments

• Ano de defesa: 2023
• Autor(a) principal: Silva, Júnior Anderson Rodrigues da
• 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: Aprendizado de máquina; Aprendizagem por reforço inverso (IRL); Artificial intelligence; Autonomous driving; Decision-making under uncertainty; Direção autônoma; Inteligência artificial; Inverse reinforcement learning (IRL); Machine learning; Motion planning; Motion prediction; Partially observed Markov decision process (POMDP); Planejamento de movimento; Predição de movimento; Processo de decisão de Markov parcialmente observável (POMDP); Robótica; Robotics; Tomada de decisão na presença de incertezas
• Link de acesso: https://www.teses.usp.br/teses/disponiveis/55/55134/tde-08012024-180517/

A vehicle navigating in an urban environment must obey traffic rules by properly setting its speed in order to stay below the road speed limit and avoiding collisions. This is presumably the scenario that autonomous vehicles will face: they will share the traffic roads with other vehicles (autonomous or not), cooperatively interacting with them. In other words, autonomous vehicles should not only follow traffic rules, but should also behave in such a way that resembles other vehicles behavior. However, manually specification of such behavior is a time-consuming and error-prone work, since driving in urban roads is a complex task, which involves many factors. Furthermore, since the interaction between vehicles is inherent to driving, inferring surrounding vehicles motion is essential to provide a more fluid navigation, avoiding a over-reactive behavior. In this sense, the uncertainty coming from noisy sensor measurements and unknown surrounding vehicles behavior cannot been neglected in order to guarantee safe and reliable decisions. In this thesis, we propose using Partially Observable Markov Decision Process (POMDP) to address the problem of incomplete information inherent of motion planning for autonomous driving. We also propose a variant of Maximum Entropy Inverse Reinforcement Learning (IRL) to learn human expert behavior from demonstration. Three different urban scenarios are covered throughout this work: longitudinal planning at signalized intersection by considering noisy measurements sensor; longitudinal and lateral planning on multi-lane roads in the presence of surrounding vehicles, in which their intention of changing lane are inferred from sequential observations; longitudinal and lateral planning during merge maneuvers in a highly interactive scenario, in which the autonomous vehicle behavior is learned from real data containing human demonstrations. Results show that our methods compare favorably to approaches that neglected uncertainty during planning, and also can improve the IRL performance, which adds safety and reliability in the decision-making.