カテゴリ: 研究会(論文単位)論文No: CHS21033グループ名: 【E】センサ・マイクロマシン部門 ケミカルセンサ研究会発行日: 2021/12/07タイトル(英語): Developing self-driving car simulation wrapper for deep learning purpose著者名: Dharmawan Willy(Kanazawa University),Nambo Hidetaka(Kanazawa University)著者名(英語): Willy Dharmawan(Kanazawa University),Hidetaka Nambo(Kanazawa University)キーワード: self-driving car simulator wrapper|self-driving car|reinforcement learning|imitation learning|deep learning|end-to-end learning|self-driving car simulator wrapper|self-driving car|reinforcement learning|imitation learning|deep learning|end-to-end learning要約(日本語): Self-driving car simulation is the foundation for self-driving car developers to design deep learning algorithms. Some simulators are available online such as Carla, Autoware, AirSim, Udacity, and many more. Carla pos-sesses a limitless prospect from all these simulators to deploy specific scenario problems in the self-driving car. Nevertheless, the deployment of certain environments requires a complex understanding of the frameworks. Based on this problem, we develop a wrapper that utilizes the Carla simulator and enhances its capability by adding features that facilitate reinforcement and imitation learning algorithm building. We also provide an ex-ample of the implementation of Double Deep Q-Network, to emphasize our set of reward policies. Based on our test, the model can converge and achieve a more stable range of rewards after 78 episodes.要約(英語): Self-driving car simulation is the foundation for self-driving car developers to design deep learning algorithms. Some simulators are available online such as Carla, Autoware, AirSim, Udacity, and many more. Carla pos-sesses a limitless prospect from all these simulators to deploy specific scenario problems in the self-driving car. Nevertheless, the deployment of certain environments requires a complex understanding of the frameworks. Based on this problem, we develop a wrapper that utilizes the Carla simulator and enhances its capability by adding features that facilitate reinforcement and imitation learning algorithm building. We also provide an ex-ample of the implementation of Double Deep Q-Network, to emphasize our set of reward policies. Based on our test, the model can converge and achieve a more stable range of rewards after 78 episodes.本誌: 2021年12月10日ケミカルセンサ研究会本誌掲載ページ: 19-25 p原稿種別: 英語PDFファイルサイズ: 1,718 Kバイト
