A Robust Stereo Splatting SLAM System with Inertial-Legged Fusion

Abstract

Recent progress in stereo-based 3D Gaussian Splatting (3DGS) SLAM has enabled small-scale robots, which are too small to carry depth cameras, to achieve localization and reconstruct photorealistic scenes with high-speed rendering. However, initializing 3D Gaussians from binocular vision still requires further improvement, and the potential of robot proprioception has not been fully leveraged. This work presents a robust stereo 3DGS SLAM with efficient inertial-legged fusion for small-scale quadruped robots (SaQu-SLAM). We develop a light-weight network to densely initialize the 3D Gaussians in the space. Besides, an efficient fusion method of inertial and legged encoder data based on Kalman filter is introduced. To improve the cross-platform generalization of our algorithm, multiple configuration combinations of these three types of sensors are provided. Moreover, we propose a mode-switching mechanism to handle intermittent visual failures. At last, we perform evaluation on a benchmark dataset, which includes large- and small-scale scenes, and a small quadruped robot in real-world confined-scale scenes, reducing the absolute trajectory error by an average of 19%, 13% and 25% respectively, when compared with other state-of-the-art methods in a similar context. It is also the only successful method in our self-customized confined mixed textured and textureless scene, whereas all vision-based or visual-inertial methods fail. Our system achieves real-time performance even on an embedded platform (Jetson AGX Orin).