FAR-LIO: Enabling High-Speed Autonomy through Fast, Accurate, and Robust LiDAR-Inertial Odometry

Maximilian Leitenstern, Marcel Weinmann, Patrick Haft, Tobias Lasser, Dominik Kulmer, Markus Lienkamp · 2026

Read FAR-LIO: Enabling High-Speed Autonomy through Fast, Accurate, and Robust LiDAR-Inertial Odometry by Maximilian Leitenstern, Marcel Weinmann, Patrick Haft, Tobias Lasser, Dominik Kulmer, Markus Lienkamp (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2606-26010.txt.

Introduction

Maximilian Leitenstern and colleagues at the Technical University of Munich build odometry for racing robots. Odometry tracks a vehicle's motion from its sensors. At racing speeds, estimates must be fast, accurate, and robust at once. FAR-LIO merges lidar scans with inertial measurements on graphics hardware. A CUDA voxel hash map speeds nearest-neighbour search and map updates.

Sparsity-aware point matching adds the accuracy. An adaptive threshold tunes the Generalized Iterative Closest Point method on the fly. An extended Kalman filter forms the back end. Upsampling and delay compensation blend lidar output with fast inertial data. Output stays smooth for closed-loop control. Tests span four sensor setups and two real race cars. The cars reached speeds near two hundred fifty kilometres per hour. Positional error drops by about seven percent against top baselines. Runtime drops by about thirty-eight percent, with one shared parameter set.

Catalog record

CREATORMaximilian Leitenstern, Marcel Weinmann, Patrick Haft, Tobias Lasser, Dominik Kulmer, Markus Lienkamp
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE40 KB
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2606.26010

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

FAR-LIO: Enabling High-Speed Autonomy through Fast, Accurate, and Robust LiDAR-Inertial Odometry Maximilian Leitenstern, Marcel Weinmann, Patrick Haft, Tobias Lasser, Dominik Kulmer, Markus Lienkamp Abstract Robust and accurate odometry estimation is essential in modern robotics. In environments characterized by highly dynamic motion and sensor noise, odometry estimation becomes increasingly challenging. Autonomous racing combines both factors in an unstructured setting, where minimizing odometry latency is essential for stable closed-loop control. This paper introduces FAR-LIO , a highly optimized CUDA-accelerated LiDAR-inertial odometry framework developed for Fast , Accurate , and Robust performance. Our system leverages a novel CUDA-based voxel hashmap to enable parallelized nearest-neighbor search and efficient map updates. We employ a sparsity-aware Generalized Iterative Closest Point algorithm with adaptive thresholding on top of the CUDA-based voxel hashmap with adaptive density to achieve low-latency without compromising accuracy. An Extended Kalman Filter serves as a robust backend. It utilizes an upsampling and delay compensation strategy to fuse the LiDAR odometry with high-frequency IMU data, thereby ensuring a robust and smooth odometry output. We evaluate FAR-LIO across four different sensor setups, using both public datasets and data from two autonomous racecars driving at speeds of up to 250 km/h. FAR-LIO achieves an average 6.9% reduction in the positional error and 38.4% lower runtime compared to state-of-the-art baselines on target hardware using a single parameter set. This demonstrates its computational efficiency and broad applicability. To build upon our work, our code is available open-source on https://github.com/TUMFTM/FAR-LIO . ©©2026 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all other uses, in any current or future media, including reprinting/republishing this material for advertising or promotional purposes, creating new collective works, for resale or redistribution to servers or lists, or reuse of any copyrighted component of this work in other works. I INTRODUCTION LiDAR odometry is a core component of many mobile robotics systems, enabling safe navigation under real-time constraints. Most state-of-the-art approaches perform fine registration to align consecutive LiDAR frames and incrementally build a 3D map of their environment for localization. This process is referred to as Simultaneous Localization and Mapping (SLAM). Integrating acceleration and angular velocity measurements from an Inertial Measurement Unit (IMU) can further enhance the accuracy and robustness of these algorithms. Many LiDAR–Inertial Odometry (LIO) frameworks, however, require extensive parameter tuning and compute resources to meet the real-time and robustness constraints necessary for real-world deployment. This is particularly challenging in domains with limited testing opportunities, extreme environmental conditions, and speeds of up to 250\text{\,}\mathrm{k}\mathrm{m}\mathrm{/}\mathrm{h} , such as autonomous racing. This work demonstrates that safe, autonomous, high-speed operation requires minimizing odometry latency and strictly bounding it below critical thresholds. Our experiments show that our highly optimized, CUDA-accelerated framework specifically addresses the limitations of many CPU-based methods, which often fail to meet real-world requirements on target hardware. Our framework integrates a CUDA-accelerated, sparsity-aware Generalized Iterative Closest Point (GICP) front-end with an Extended Kalman Filter (EKF) backend for IMU fusion. The registration utilizes a CUDA-based voxel hashmap with adaptive density, incorporating a robust kernel and the adaptive thresholding scheme from Vizzo et al. [ 31 ] . This way, we demonstrate F ast, A ccurate, and R obust performance across three datasets using four different LiDAR and IMU sensors in residential areas, on highways, and racetracks (Fig. 1 ). To ensure transferability, we evaluate all algorithms on the target hardware of an autonomous racecar and using the Robot Operating System (ROS) [ 25 , 21 ] to account for runtime constraints. Same Parameter Set Fig. 1: Point cloud maps of the Yas Marina Circuit (Abu Dhabi Autonomous Racing League, left) and the KITTI 00 sequence (right) [ 11 , 10 ] . Both were generated with FAR-LIO using identical parameters, which demonstrates its transferability from racetrack to urban environments. Our main contributions are: • We introduce and open-source FAR-LIO , a highly optimized, CUDA-accelerated LiDAR-inertial odometry method with F ast, A ccurate, and R obust performance. • We detail and ablate the core contributions of FAR-LIO , including its CUDA-based voxel hashmap with adaptive density, the sparsity-aware GICP, and the EKF backend. • We validate the real-world capability of our framework by demonstrating the critical role of localization latency on system robustness. • We show that FAR-LIO outperforms state-of-the-art methods with an average 6.9% reduction in the positional error and 38.4% lower runtime, using a single parameter set across public autonomous driving and high-speed autonomous racing datasets up to 250\text{\,}\mathrm{k}\mathrm{m}\mathrm{/}\mathrm{h} . II RELATED WORK Robust localization is crucial for mobile autonomy. LIO has emerged as a leading methodology, combining the geometric data from a LiDAR with the high-frequency motion updates from an IMU. In this review, we outline key LiDAR-only techniques before focusing on the more dominant LIO systems, including SLAM frameworks with loop closure. We categorize these systems by their underlying backend, either filtering or optimization, thereby deriving the research gap. II-A LiDAR-Only Odometry LiDAR odometry algorithms employ scan registration to find the transformation between two frames, thereby incrementally deriving a trajectory. This process usually consists of two steps: data association to find correspondences and the computation of the transformation. One can distinguish between feature-based strategies, deriving the correspondences from geometric features, and direct strategies, estimating temporal correspondences from a heuristic assumption. Leveraging feature-based scan registration, LOAM [ 36 ] can be seen as a fundamental method for LiDAR odometry and SLAM in general, building the foundation for many modern frameworks. Therefore, several approaches, such as LeGO-LOAM [ 30 ] , or F-LOAM [ 32 ] build upon it. In terms of direct scan registration, a pioneering work is the Iterative Closest Point (ICP) algorithm [ 3 ] , using point-to-point distances for temporal correspondence estimation. Since its presentation in 1992, many variants have evolved to enhance its robustness and accuracy [ 28 , 7 , 17 , 8 , 5 , 31 , 37 , 20 , 9 ] . These approaches mainly differ in how they establish correspondences (e.g., point-to-point, point-to-plane, or distribution-based), how they reject outliers, and how they optimize the transformation. The GICP [ 28 ] algorithm, for instance, which is employed in this work, combines the point-to-point and point-to-plane metric within a probabilistic framework, thereby enhancing alignment robustness. Among recent advancements in robust registration, the KISS-ICP [ 31 ] introduces an adaptive thresholding method for outlier rejection, based on the deviation between the initial guess and the registered pose. Furthermore, they perform a simple scan-to-submap ICP registration using a CPU-based voxel hashmap. II-B LiDAR-Interial Odometry LiDAR-inertial odometry frameworks often extend the scan registration techniques, as used in LiDAR-only approaches, with a backend to allow for the integration of IMU data. Existing backends may be categorized into filter-based approaches, such as Kalman Filters (KF), and optimization-based approaches, such as factor graphs [ 33 ] . In the latter case, LIO-SAM [ 29 ] combines the frontend of LOAM [ 36 ] with a factor graph optimization to directly integrate IMU pre-integration factors. LOCUS 2.0 [ 26 ] , and its predecessor [ 23 ] employ a GICP-based frontend with a sliding-window map optimization approach. A more recent work by Koide et al. [ 18 ] introduces GLIM , a full SLAM framework using a factor graph

Read the complete text