Driving behavior modeling based on stability and safety oriented inverse reinforcement learning

2023 Fiscal Year Final Research Report

• Project/Area Number: 21H03517
• Research Category: Grant-in-Aid for Scientific Research (B)
• Allocation Type: Single-year Grants
• Section: 一般
• Review Section: Basic Section 61050:Intelligent robotics-related
• Research Institution: Tokyo Institute of Technology
• Principal Investigator: Shimosaka Masamichi 東京工業大学, 情報理工学院, 准教授 (40431796)
• Co-Investigator(Kenkyū-buntansha): 小竹 元基 東京工業大学, 工学院, 教授 (10345085)
• Project Period (FY): 2021-04-01 – 2024-03-31
• Keywords: 逆強化学習 / 運転行動モデリング / 経路生成 / RRT / 運転行動シミュレーション

Outline of Final Research Achievements

In recent years, the development of advanced driver assistance systems has been active. In this study, we developed a method for modeling driving behavior (driving behavior prediction and simulation generation) based on inverse reinforcement learning (imitation learning and inverse optimal control) with a view to the advancement of these systems. In particular, from the viewpoint of application to modeling of automobile driving behavior, we developed techniques oriented toward stability and safety, which have been lacking in conventional techniques. Specifically, we developed a probabilistic path generation method, an efficient reward estimation method using sampling based on path candidates, and an efficient learning method using negative examples (data corresponding to traffic accidents). In addition, to safely collect data corresponding to negative examples, we constructed a driving behavior data collection environment and actually collected driving data.

Free Research Field

知能ロボティクス

Academic Significance and Societal Importance of the Research Achievements

本研究で構築した技術の一つである，RRTに基づく経路生成法，及び，重点サンプリング法に基づく報酬場最適化手法は，逆強化学習に基づく運転行動モデリングの適用範囲を拡張するものである，また，負の事案を用いた手法は，熟練ドライバの運転行動の表層的な模倣だけでなく，安全性という価値感を含めた模倣を目指す基盤となっている．構築した技術は交通事故を低減する新たな運転支援システムの開発に貢献する．例えば，事故要因分析や事故回避のシミュレーションへの応用が期待できる．