TY - GEN
T1 - Terrain Attribute Recognition System for CPG-Based Legged Robot
AU - Chen, Hongjin
AU - Zhu, Xi
AU - Zhu, Siting
AU - Chen, Haoyao
AU - Zhang, Shiwu
AU - Lou, Yunjiang
N1 - Publisher Copyright:
© 2021, Springer Nature Switzerland AG.
PY - 2021
Y1 - 2021
N2 - In this paper, we develop a terrain attribute recognition system for CPG-based legged robots. First, a low-cost sensing hardware device is designed to be integrated into the robot, including a tactile sensor array and RGB camera. Second, for the tactile modality, a novel terrain attribute recognition framework is proposed. A data generation strategy that adapts to the motion characteristics is presented, which transforms the original tactile signal into a structured representation, and extract meaningful features. Based on unsupervised and supervised machine learning classifiers, the recognition rates reach 94.0% and 95.5%, and the switching time is 1 to 3 steps. Third, for the recognition of terrain attributes in the visual modality, a lightweight real-time mobile attention coding network (MACNet) is proposed as an end-to-end model, which shows an exhibiting an accuracy of 88.5% on the improved GTOS mobile data set, 169FPS inference speed and 6.6 MB model parameter occupancy. Finally, these two methods are simultaneously applied to the AmphiHex-II robot for outdoor experiments. Experimental results show that each modality has its own advantages and disadvantages, and the complementary relationship between multiple modalities plays an irreplaceable role in a broader scene.
AB - In this paper, we develop a terrain attribute recognition system for CPG-based legged robots. First, a low-cost sensing hardware device is designed to be integrated into the robot, including a tactile sensor array and RGB camera. Second, for the tactile modality, a novel terrain attribute recognition framework is proposed. A data generation strategy that adapts to the motion characteristics is presented, which transforms the original tactile signal into a structured representation, and extract meaningful features. Based on unsupervised and supervised machine learning classifiers, the recognition rates reach 94.0% and 95.5%, and the switching time is 1 to 3 steps. Third, for the recognition of terrain attributes in the visual modality, a lightweight real-time mobile attention coding network (MACNet) is proposed as an end-to-end model, which shows an exhibiting an accuracy of 88.5% on the improved GTOS mobile data set, 169FPS inference speed and 6.6 MB model parameter occupancy. Finally, these two methods are simultaneously applied to the AmphiHex-II robot for outdoor experiments. Experimental results show that each modality has its own advantages and disadvantages, and the complementary relationship between multiple modalities plays an irreplaceable role in a broader scene.
KW - CPG-based legged robot
KW - Deep neural network
KW - Tactile and visual sensing
KW - Terrain attribute recognition
UR - https://www.scopus.com/pages/publications/85118157847
U2 - 10.1007/978-3-030-89134-3_51
DO - 10.1007/978-3-030-89134-3_51
M3 - 会议稿件
AN - SCOPUS:85118157847
SN - 9783030891336
T3 - Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
SP - 556
EP - 567
BT - Intelligent Robotics and Applications - 14th International Conference, ICIRA 2021, Proceedings
A2 - Liu, Xin-Jun
A2 - Nie, Zhenguo
A2 - Yu, Jingjun
A2 - Xie, Fugui
A2 - Song, Rui
PB - Springer Science and Business Media Deutschland GmbH
T2 - 14th International Conference on Intelligent Robotics and Applications, ICIRA 2021
Y2 - 22 October 2021 through 25 October 2021
ER -