Pushing Radar Odometry Beyond the Pavement: Current Capabilities and Challenges
Shaunak Kolhe, Peng Jiang, Maggie Wigness, Philip Osteen, Timothy Overbye, Christian Ellis, Srikanth Saripalli · 2026
Read Pushing Radar Odometry Beyond the Pavement: Current Capabilities and Challenges by Shaunak Kolhe, Peng Jiang, Maggie Wigness, Philip Osteen, Timothy Overbye, Christian Ellis, Srikanth Saripalli (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-2604-24674.txt.
Introduction
Kolhe and six colleagues test radar odometry, the estimation of a vehicle's path from radar, beyond paved urban roads. Off-road motion in full three dimensions, ground returns from terrain, and sparse or unstable features pose fresh challenges. They offer two simple baselines: Radar-KISSICP, which compensates motion to build three-dimensional radar clouds, and Radar-IMU, which steadies scan matching with inertial data. On the Great Outdoors dataset both improve trajectory estimates on hard routes.
Catalog record
| CREATOR | Shaunak Kolhe, Peng Jiang, Maggie Wigness, Philip Osteen, Timothy Overbye, Christian Ellis, Srikanth Saripalli |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 26960 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2604.24674
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Pushing Radar Odometry Beyond the Pavement: Current Capabilities and Challenges Shaunak Kolhe, Peng Jiang, Maggie Wigness, Philip Osteen, Timothy Overbye, Christian Ellis, Srikanth Saripalli Abstract Radar offers unique advantages for localization in unstructured environments, including robustness to weather, lighting, and airborne particulates. While most prior work has studied radar odometry in urban, largely planar settings, its performance in off-road environments remains less understood. In this paper, we investigate the potential of radar for off-road odometry estimation and identify key challenges that arise from full S E ( 3 ) SE(3) vehicle motion, terrain-induced ground returns, and sparse or unstable features. To address these issues, we introduce two simple baselines: Radar-KISSICP, which applies motion compensation to generate 3D-aware radar pointclouds, and Radar-IMU, which leverages IMU preintegration to stabilize scan matching. Experiments on the Great Outdoors (GO) dataset demonstrate that these baselines improve trajectory estimation in challenging routes and provide a reference point for future development of radar odometry in off-road robotics. I INTRODUCTION Robust localization is a fundamental capability for autonomous robots operating in diverse and unstructured environments. While cameras and LiDARs have long been the dominant sensing modalities for odometry, their performance degrades severely under adverse weather or visually degraded conditions such as rain, snow, fog, dust, or smoke. Millimeter-wave radar, with its longer wavelengths and Doppler measurement capability, has recently re-emerged as a compelling alternative [ 13 ] . Unlike light-based sensors, radar can penetrate particulates and provide reliable returns at long range, enabling continued operation when other modalities degrade or fail. Encouraged by these properties, a growing body of work has explored radar odometry and mapping in automotive and urban driving scenarios. Recent methods, such as CFEAR [ 1 ] and ORORA [ 19 ] , achieve near LiDAR-level localization performance on benchmark datasets like Oxford RobotCar [ 3 ] , MulRan [ 17 ] and Boreas [ 6 ] . However, these datasets primarily constrain vehicle motion to planar trajectories in SE(2) . As a result, many state-of-the-art radar odometry systems implicitly rely on the assumption of flat terrain and limited pitch or roll excitation, which does not generalize to off-road conditions. In contrast, off-road environments demand full SO(3) motion estimation due to rough terrain, elevation changes, and frequent loss of stable landmarks. These conditions expose radar-specific failure modes: ground returns become confounding rather than outliers, false correspondences accumulate in ICP pipelines, and sparse radar pointclouds lead to ill-conditioned registration. To address these challenges, we evaluate existing radar odometry engines in off-road settings and propose two extensions: (i) Radar-KISSICP, which is based on KISS-ICP [ 26 ] and incorporates a lightweight SO(3) motion compensation step to produce sparse but 3D-aware radar pointclouds, and (ii) Radar-IMU, which leverages IMU preintegration to provide a robust initial guess for radar scan-matching in SE(3) . Through experiments on the Great Outdoors (GO) dataset [ 15 ] , we show that our localization methods can match or outperform existing approaches in challenging off-road environments, including scenarios with severe pitch/roll and terrain discontinuities. We further analyze failure cases, such as ravine traversals, where radar-only pipelines accumulate significant drift, and demonstrate how IMU integration helps recover accurate trajectories. Our evaluation and results emphasize both the promise and the limitations of radar odometry in unstructured terrain, laying the groundwork for future robust off-road localization pipelines. Fig. 1 : Motivating scenario for the use of radar in an off-road, degraded environment. A Clearpath Warthog equipped with a Navtech radar operates in a snow covered dense forest. II Related Work Early radar odometry approaches focused on adapting classical registration techniques to handle measurement noise and sparsity. Kung et al. [ 18 ] proposed a normal distributions transform (NDT) based RADAR odometry that also accounts for uncertainties through a probabilistic submap, outperforming state-of-the-art automotive and learning-based scanning methods by better handling outliers. Rapp et al. [ 23 ] also applied joint doppler-clustering based NDT, weighting each RADAR point by similarity between measured and estimated velocities to reduce outlier effects. To address the sparsity issue in automotive RADAR, Holder et al. [ 14 ] proposed submap matching, combining several RADAR scans to construct more information, though this method did not consider the uncertainty of ego-motion or RADAR measurements. More recently, Adolfsson et al. presented the CFEAR approach [ 1 ] , which uses a feature extraction algorithm to generate filtered surface points that can be reliably scan-matched from frame to frame, demonstrating LIDAR-level localization with RADAR. Lim et al. developed ORORA [ 19 ] , an outlier-robust RADAR odometry method that demonstrated robust performance even with numerous outliers and was less sensitive to the specific feature extraction method used, particularly in featureless scenes. Direct methods, on the other hand, directly compare or use raw RADAR measurements without explicit feature extraction. Kellner et al. [ 16 ] used the angular position and radial velocity of automotive RADAR sensors to estimate ego-motion, classifying stationary targets with RANSAC to fit velocity estimates. Le Gentil et al. [ 11 ] introduced a Direct RADAR Odometry (DRO) method that is notable for being the first direct RADAR registration method to account for continuous motion and Doppler-based distortion in a gradient-based optimization, performing well in both structured and unstructured environments, including featureless tunnels. Recent advancements have leveraged deep learning to extract robust features and improve RADAR odometry performance, often outperforming hand-crafted methods. Barnes et al. [ 4 ] demonstrated a fully differentiable, correlation-based RADAR odometry pipeline that learns a binary mask to remove distractor features, such as objects in motion and RADAR noise. Other approaches use unsupervised learning, such as the HERO method by Burnett et al. [ 5 ] , which combines probabilistic trajectory estimation and deep learned features without requiring ground truth pose information for training. This unsupervised method approaches the method in [ 1 ] , outperforms the supervised method in [ 4 ] and demonstrates effectiveness under adverse weather conditions. Lisus et al. [ 21 ] further advanced RADAR-LIDAR localization by presenting a deep-learning-based approach that learns to weight RADAR points for ICP, effectively filtering out harmful RADAR artifacts, noise, and even other vehicles, thus improving convergence and reducing localization errors. Lisus et al. [ 20 ] investigated whether Doppler velocity measurements are useful for spinning RADAR odometry and presented a novel method to analytically extract these velocities while retaining full angular resolution. Their experiments demonstrated that using Doppler measurements improves performance in easy geometric environments and maintains good performance in geometrically degenerate situations like tunnels or skyways where other methods might fail. While the above demonstrate the effectiveness of RADAR as a robust odometry source, rarely are these methods tested in offroad environments. It is hence in our interest to verify if this performance is applicable in more aggressive terrain, and identify the challenges therein. Fig. 2 : Overhead view of routes from the GO dataset [ 15 ] (left) and the Warthog sensor setup used for the dataset (right). III Benchmarking Framework To systematically study radar odometry in off-road environments, we design a benchmarking framework that evaluates state-of-the-art radar odometry engines as well as two diagnostic baselines. Our goal is to reveal the strengths and limitations of existing methods when vehicle motion extends beyond planar SE(2) into