Switch State Detection by MSRS and YOLOv4 and Automatic Switch Operation with a Robot Arm by Reinforcement Learning in Virtual Environments

• Item type: SIG Technical Reports(1)
• 公開日: 2021-02-25
• タイトル: Switch State Detection by MSRS and YOLOv4 and Automatic Switch Operation with a Robot Arm by Reinforcement Learning in Virtual Environments
• タイトル / 言語: en
• タイトル: Switch State Detection by MSRS and YOLOv4 and Automatic Switch Operation with a Robot Arm by Reinforcement Learning in Virtual Environments
• 言語: eng
• キーワード / 主題Scheme: Other
• キーワード / 主題: セッション5-2
• 資源タイプ / 資源タイプ識別子: http://purl.org/coar/resource_type/c_18gh
• 資源タイプ: technical report
• 著者所属: Faculty of Science and Engineering, Waseda University
• 著者所属: Faculty of Science and Engineering, Waseda University
• 著者所属: Faculty of Science and Technology, Seikei University
• 著者所属(英): en
• 著者所属(英): Faculty of Science and Engineering, Waseda University
• 著者所属(英): en
• 著者所属(英): Faculty of Science and Engineering, Waseda University
• 著者所属(英): en
• 著者所属(英): Faculty of Science and Technology, Seikei University
• 著者名: Li, Qi; Jun, Ohya; Hiroyuki, Ogata
• 著者名(英): Li, Qi; Jun, Ohya; Hiroyuki, Ogata
• 論文抄録 / 内容記述タイプ: Other
• 論文抄録 / 内容記述: This paper proposes a method for detecting switch states and operating the switch using arms of disaster response robots, which are expected to be widely used for handling dangerous tasks caused by disasters. To enable YOLOv4 to detect switches with various scales, the training dataset is built by a Multi-Scale Reduction Stitching (MSRS) data enhancement algorithm, which reduces two original images into multi scales and stitch them in one new image; thereby, the new image contains half, quarter and one eighth of the original scale of the switch. The YOLOV4, which is trained by the dataset built by MSRS, detects the switch state. Based on the result of the switch detection, arms of a four-leg robot (in this paper, D’Kitty) operate the switch by reinforcement learning. As a result of exploring effectiveness of different reinforcement learning algorithms for the task of operating switches, it turns out that PPO gives the best performance. In a simulated environment, the success rate for fixed positions of the switch and robot is 97%, and for small-scale random positions is higher than 90%.
• 論文抄録(英) / 内容記述タイプ: Other
• 論文抄録(英) / 内容記述: This paper proposes a method for detecting switch states and operating the switch using arms of disaster response robots, which are expected to be widely used for handling dangerous tasks caused by disasters. To enable YOLOv4 to detect switches with various scales, the training dataset is built by a Multi-Scale Reduction Stitching (MSRS) data enhancement algorithm, which reduces two original images into multi scales and stitch them in one new image; thereby, the new image contains half, quarter and one eighth of the original scale of the switch. The YOLOV4, which is trained by the dataset built by MSRS, detects the switch state. Based on the result of the switch detection, arms of a four-leg robot (in this paper, D’Kitty) operate the switch by reinforcement learning. As a result of exploring effectiveness of different reinforcement learning algorithms for the task of operating switches, it turns out that PPO gives the best performance. In a simulated environment, the success rate for fixed positions of the switch and robot is 97%, and for small-scale random positions is higher than 90%.
• 書誌レコードID / 収録物識別子タイプ: NCID
• 書誌レコードID / 収録物識別子: AA11131797
• 書誌情報: 研究報告コンピュータビジョンとイメージメディア（CVIM）; 巻 2021-CVIM-225, 号 40, p. 1-6, 発行日 2021-02-25
• ISSN / 収録物識別子タイプ: ISSN
• ISSN / 収録物識別子: 2188-8701
• Notice: SIG Technical Reports are nonrefereed and hence may later appear in any journals, conferences, symposia, etc.
• 出版者 / 言語: ja
• 出版者: 情報処理学会