Efficient Positioning for Unmanned Systems: A Mamba-Based CSI Fingerprinting Framework

Reliable localization is essential for autonomous operation in unmanned industrial environments, where visionbased methods often degrade due to poor illumination, occlusion, and privacy constraints. WiFi Channel State Information (CSI) enables fingerprinting based localization by capturing fine-grained signal characteristics. However, existing approaches either rely on limited handcrafted features or employ deep models such as CNNs, LSTMs, and Transformers, which suffer from restricted global modeling capability or high computational complexity. To address these challenges, this paper proposes a novel Mamba-based CSI fingerprinting framework built upon a selective state space model. The proposed method efficiently captures long-range dependencies in high-dimensional CSI sequences with linear computational complexity, inherently bypassing the sequential and computational bottlenecks of traditional architectures. Experimental results in a real-world indoor environment demonstrate that the proposed approach achieves high localization accuracy with a root mean square error (RMSE) of 1.04 m and a mean absolute error (MAE) of 1.08 m. Furthermore, complexity analysis on a standard CPU platform reveals that the proposed framework achieves an exceptionally low inference latency of 1.51 ms, outperforming conventional and deep learning baselines in efficiency. This optimal trade-off between localization accuracy and real-time execution speed confirms the immense potential of the proposed method for resource-constrained devices in dynamic unmanned systems.