A Bio-Inspired SLAM System with Multi-Sensor Fusion for Real-World Outdoor Navigation

Traditional Simultaneous Localization and Mapping (SLAM) systems face several challenges, including high computational complexity, limited robustness in complex or unfamiliar environments, and dependence on Global Positioning System (GPS) signals. In contrast, animals exhibit remarkably efficient and robust natural navigation abilities, inspiring the development of bio-inspired SLAM approaches that mimic these biological mechanisms. While existing bio-inspired SLAM algorithms have shown promise in constrained environments, their accuracy and robustness often degrade in highly complex scenarios, leading to suboptimal mapping performance. To address these challenges, we propose MS-NeuroSLAM, a novel GPS-independent SLAM framework that innovatively integrates: a multi-sensor front-end (monocular camera + IMU + wheel odometry) for low-cost pose estimation, a biologically plausible back-end emulating hippocampal-entorhinal spatial coding, and an accelerated loop closure mechanism combining visual template matching with bag-of-words. Experimental results show that MS-NeuroSLAM achieves superior accuracy and robustness compared to state-of-the-art methods (ORB-SLAM3, OpenVINS, and NeuroSLAM) on both the KAIST dataset and our self-collected campus dataset with dynamic environments, while simultaneously reducing loop closure detection time by 94.8%.