Real-Time Obstacle Detection for a Biped Robot Climbing on Exoskeleton-Structure Glass Windows

This paper presents a novel real-time obstacle detection method for a biped climbing robot. It is composed of a modified UNet neural network combined with a large-field separable contextual integration (LSCI) module and machine vision techniques for line detection. Such an approach allows the biped robot to accurately detect and locate metal exoskeleton-structure obstacles on glass curtain walls. Moreover, the algorithm adopts point cloud data collected from a depth camera to ensure precise identification of target exoskeleton-structure frames including the positions and tilt angles, facilitating effective real-time collision-free motion planning for the cleaning robot system. Comparative experiments with other neural networks used for image segmentation demonstrate that the proposed LSCI-UNet architecture significantly improves the frame obstacle detection accuracy on exoskeleton-structure glass windows with limited data. The reported method enables the biped cleaning robots to detect obstacles in real-time and effectively navigate on glass walls with exoskeleton-structure beams. Note to Practitioners—Cleaning glass windows of high-altitude buildings imposes significant risks to human workers due to potential falls, which inspires the deployment of window-cleaning robots for such tasks. Bipedal robots exhibit an ability to navigate inclined surfaces of complex exoskeleton structures. Meanwhile, accurate detection and positioning of metal beam obstacles on the windows is crucial for efficient robotic operations in complex environments. However, detecting such obstacles on transparent glass curtain walls remains a considerable challenge. Particularly, the objects behind the glass curtain cause confusion in obstacle detection. To overcome such issue, this paper proposes a new approach of real-time obstacle detection of a biped climbing robot for cleaning glass windows on an exoskeleton-structure building. The effectiveness of the reported methodology has been validated by conducting experimental investigations.