[2O4-GS-8] Robot and real worlds

[2O4-GS-8-04] Learning Safe and Explainable Control of Automated Vehicles in Emergency Situations

〇Takeshi Kunieda¹, Rintaro Imamura¹, Sachiyo Arai¹ (1. Chiba University)

Keywords:Automated Vehicles, Reinforcement Learning, Sim-to-Real

In recent years, automated driving technology has become capable of performing appropriate operations in situations that require advanced judgment and driving skills, and the system side has begun to take the initiative in driving operations within a limited range, such as on highways. In applying automated driving technology to urban environments, there is the issue of the difficulty in dealing with emergency situations. In addition, in the operation of the technology, an explanation of the circumstances leading up to the output of avoidance actions is required from the standpoint of reliability. Therefore, recent demands for automated driving are to ensure both explainability and avoidance performance in emergency situations. However, emergency situations in traffic flow occur only rarely and in a wide variety of situations.
Therefore, in this study, first, emergency situations are generated exhaustively using a simulation environment. Then, the avoidance behavior is formulated as a bandit problem and acquired using deep reinforcement learning. Next, in order to enhance the explanatory power of the behavioral outputs, the behavioral outputs are obtained from the data collected by deep reinforcement learning without neural networks using a support vector machine. Computational experiments show that both methods can appropriately avoid obstacles.