Citation

Wachi A. Trans. Soc. Automot. Eng. Jpn. 2020; 51(5): 950-955.

Copyright

(Copyright © 2020, Society of Automotive Engineers of Japan)

DOI

10.11351/jsaeronbun.51.950

PMID

unavailable

Abstract

We propose a method to create failure-scenarios of an autonomous driving vehicle by training other surrounding vehicles by means of multiagent adversarial reinforcement learning. Failure in driving environments might lead to catastrophic results. Hence, when developing a software of autonomous driving cars, we must find as many failure-cases as possible and then improve the software. However, as the software becomes complicated, it is hard to find failure-scenarios that are useful for the software improvement. Hence, we propose a framework to create various failure-scenarios of an autonomous car by training other car(s) via reinforcement learning. We demonstrate the effectiveness of our proposed method with two kinds of experiments.

Language: ja

Keywords

Safety; test/evaluation; intelligent vehicle