Automated reality capture for indoor inspection using BIM and a multi-sensor quadruped robot

This paper presents a real-time, cost-effective navigation and localization framework tailored for quadruped robot-based indoor inspections. A 4D Building Information Model is utilized to generate a navigation map, supporting robotic pose initialization and path planning. The framework integrates a cost-effective, multi-sensor SLAM system that combines inertial-corrected 2D laser scans with fused laser and visual-inertial SLAM. Additionally, a deep-learning-based object recognition model is trained for multi-dimensional reality capture, enhancing comprehensive indoor element inspection. Validated on a quadruped robot equipped with an RGB-D camera, IMU, and 2D LiDAR in an academic setting, the framework achieved collision-free navigation, reduced localization drift by 71.77 % compared to traditional SLAM methods, and provided accurate large-scale point cloud reconstruction with 0.119-m precision. Furthermore, the object detection model attained mean average precision scores of 73.7 % for 2D detection and 62.9 % for 3D detection.