A platform for research: civil engineering, architecture and urbanism
A platform for research: civil engineering, architecture and urbanism
Truss member registration for implementing autonomous gripping in biped climbing robots
Abstract Gripping of truss members to ensure reliable attachment on each step is difficult for robots invented for climbing trusses. To address this problem, a novel truss member registration method was proposed in this study based on color and depth images to implement autonomous gripping in biped climbing robots. To this end, VertexNet, a new keypoint-based deep learning neural network, was developed to ascertain compact quadrilateral bounding boxes corresponding to truss members from color images. In turn, the bounding boxes were used to segment the corresponding point clouds in the depth images. A RANSAC-based algorithm was used to estimate the parameters of truss members to generate autonomous gripping motion. Experimental results revealed that the proposed VertexNet outperforms state-of-the-art detectors in truss member detection. Further, in truss member parameter estimation, the proposed sensing system exhibited good accuracy over a wide working range. The truss member registration method was applied to the biped climbing robot, Climbot. It exhibited multistep climbing motion on a truss successfully, reinstating the feasibility and effectiveness of the proposed method.
Highlights Gripping is essential to the safety and speed of movement of biped climbing robots. Truss member registration is a prerequisite for autonomous gripping. The proposed VertexNet outperforms other detectors in truss member detection. RANSAC-based algorithm exhibits good accuracy in truss member parameter estimation. Using the proposed method, a robot realized multistep climbing in a truss environment.
Truss member registration for implementing autonomous gripping in biped climbing robots
Abstract Gripping of truss members to ensure reliable attachment on each step is difficult for robots invented for climbing trusses. To address this problem, a novel truss member registration method was proposed in this study based on color and depth images to implement autonomous gripping in biped climbing robots. To this end, VertexNet, a new keypoint-based deep learning neural network, was developed to ascertain compact quadrilateral bounding boxes corresponding to truss members from color images. In turn, the bounding boxes were used to segment the corresponding point clouds in the depth images. A RANSAC-based algorithm was used to estimate the parameters of truss members to generate autonomous gripping motion. Experimental results revealed that the proposed VertexNet outperforms state-of-the-art detectors in truss member detection. Further, in truss member parameter estimation, the proposed sensing system exhibited good accuracy over a wide working range. The truss member registration method was applied to the biped climbing robot, Climbot. It exhibited multistep climbing motion on a truss successfully, reinstating the feasibility and effectiveness of the proposed method.
Highlights Gripping is essential to the safety and speed of movement of biped climbing robots. Truss member registration is a prerequisite for autonomous gripping. The proposed VertexNet outperforms other detectors in truss member detection. RANSAC-based algorithm exhibits good accuracy in truss member parameter estimation. Using the proposed method, a robot realized multistep climbing in a truss environment.
Truss member registration for implementing autonomous gripping in biped climbing robots
Gu, Shichao (author) / Zhu, Haifei (author) / Lin, Xubin (author) / Tan, Jiongyu (author) / Ye, Wenda (author) / Guan, Yisheng (author)
2022-01-15
Article (Journal)
Electronic Resource
English