Kolmogorov-Arnold Network-Based Denoising for LiDAR Odometry in High-Reflectivity Regions

Reliable odometry is important for robots operating in environments with highly reflective surfaces. Yet its high reliability is difficult to achieve since intense reflection-induced noise gives rise to spurious points and feature mismatches. For purely LiDAR-based systems, this motion distortion further exacerbates due to the absence of auxiliary sensors. To solve these problems, this work proposes a denoising framework based on Kolmogorov–Arnold Network (KAN) by integrating the strengths of multi-layer perceptrons and spline functions for robust point cloud denoising. The method can learn to model complex noise distributions and adaptively tunes its parameters to local noise intensity, ensuring the retention of critical geometric features. To enable real-time compensation for motion distortion and adaptive process noise adjustment according to motion confidence, this work proposes a multi-iteration optimization scheme that embeds denoising directly into the state estimation pipeline is designed. Extensive experiments on public benchmarks and real-world experiments validate the superiority of our method in achieving accurate and stable LiDAR localization in high-reflectivity scenarios over the state of the art. The source code is available at https://github.com/T21512/KD-LO.