Pushing Radar Odometry Beyond the Pavement: Current Capabilities and Challenges

TL;DR

Radar-KISSICP and Radar-IMU improve trajectory estimation in off-road environments.

cs.RO 🔴 Advanced 2026-04-28 27 views
Shaunak Kolhe Peng Jiang Maggie Wigness Philip Osteen Timothy Overbye Chrisitan Ellis Srikanth Saripalli
Radar Odometry Off-road Navigation Motion Compensation IMU Integration Robotic Localization

Key Findings

Methodology

This study introduces two baseline methods: Radar-KISSICP and Radar-IMU, to address the challenges of radar odometry in off-road environments. Radar-KISSICP applies motion compensation to generate 3D-aware radar pointclouds, while Radar-IMU leverages IMU preintegration to stabilize scan matching. These methods were validated on the Great Outdoors (GO) dataset, showing improved trajectory estimation in complex terrains.

Key Results

  • In experiments on the GO dataset, Radar-KISSICP and Radar-IMU significantly improved trajectory estimation accuracy in challenging off-road routes. For instance, on Route 3, Radar-IMU achieved an absolute trajectory error of 15.12%, compared to 61.08% for CFEAR.
  • Radar-IMU demonstrated excellent performance in handling terrain discontinuities and severe pitch/roll scenarios, particularly recovering vertical displacement in the ravine traversal of Route 3.
  • Radar-KISSICP reduced absolute trajectory error by mitigating planar assumptions, but the sparse nature of radar pointclouds led to unstable nearest-neighbor associations.

Significance

This research highlights the potential and challenges of radar odometry in off-road environments, particularly in complex terrains. By introducing Radar-KISSICP and Radar-IMU, the study provides a reference for future off-road robotic localization systems. These methods not only improve trajectory estimation accuracy but also demonstrate the importance of inertial priors in complex terrains, laying the groundwork for future multi-sensor fusion and more robust off-road localization pipelines.

Technical Contribution

Technical contributions include introducing motion compensation and IMU preintegration to address radar odometry challenges in off-road environments. Radar-KISSICP generates 3D-aware radar pointclouds, reducing planar assumptions, while Radar-IMU improves trajectory recovery by providing vertical displacement awareness. These methods demonstrate how to enhance radar odometry robustness in uneven terrains.

Novelty

This study is the first to systematically evaluate radar odometry capabilities in off-road environments, proposing two novel methods: Radar-KISSICP and Radar-IMU. These methods address the limitations of traditional radar odometry in complex terrains through motion compensation and IMU integration, showcasing new pathways for robust localization in unstructured environments.

Limitations

  • Radar-KISSICP suffers from unstable nearest-neighbor associations due to sparse radar pointclouds, potentially affecting accuracy in some scenarios.
  • Radar-IMU, while improving vertical displacement awareness, may still face inertial drift issues without bias correction.
  • Radar odometry may still face challenges in extremely complex terrains, requiring further sensor fusion.

Future Work

Future work could explore more complex multi-sensor fusion techniques, such as integrating 3D radar and visual/LWIR cameras, to enhance localization accuracy in off-road environments. Additionally, research could further optimize IMU integration methods to reduce inertial drift and develop smarter feature selection algorithms to address sparse radar pointcloud issues.

AI Executive Summary

In the fields of autonomous driving and robotics, precise localization systems are crucial for achieving autonomous navigation. However, traditional cameras and LiDAR sensors perform poorly in adverse weather and complex terrains. Millimeter-wave radar, with its penetration and long-range measurement capabilities, has emerged as an attractive alternative.

This study explores the potential of radar odometry in off-road environments, introducing two baseline methods: Radar-KISSICP and Radar-IMU. Radar-KISSICP applies motion compensation to generate 3D-aware radar pointclouds, while Radar-IMU leverages IMU preintegration to stabilize scan matching. These methods were validated on the Great Outdoors (GO) dataset, showing improved trajectory estimation in complex terrains.

Experimental results demonstrate that Radar-KISSICP and Radar-IMU significantly improved trajectory estimation accuracy in challenging off-road routes. For instance, on Route 3, Radar-IMU achieved an absolute trajectory error of 15.12%, compared to 61.08% for CFEAR. These results highlight the importance of inertial priors in complex terrains.

However, radar odometry in off-road environments still faces challenges, such as sparse radar pointclouds and inertial drift issues. Future work could explore more complex multi-sensor fusion techniques to enhance localization accuracy in off-road environments.

This study provides a reference for future off-road robotic localization systems, demonstrating how to achieve robust localization in unstructured environments. By introducing motion compensation and IMU integration, these methods lay the groundwork for future multi-sensor fusion and more robust off-road localization pipelines.

Deep Analysis

Background

In the fields of autonomous driving and robotics, localization systems are the core technology for achieving autonomous navigation. Traditionally, cameras and LiDAR sensors have been the mainstream ranging tools, but they perform poorly in adverse weather and complex terrains. Millimeter-wave radar, with its longer wavelengths and Doppler measurement capability, has emerged as a compelling alternative. In recent years, radar odometry technology has made significant progress in urban and automotive driving scenarios, such as the CFEAR and ORORA methods achieving near LiDAR-level localization performance on benchmark datasets like Oxford RobotCar and MulRan. However, these datasets primarily constrain vehicle motion to planar trajectories, failing to fully reflect the complexity of off-road environments.

Core Problem

The performance of radar odometry in off-road environments remains unclear, especially in scenarios with rugged terrain, sparse or unstable landmarks. Off-road environments demand full SE(3) motion estimation due to rough terrain and frequent loss of stable landmarks, exposing 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. Addressing these issues is crucial for achieving robust off-road localization.

Innovation

This study introduces two innovative methods:


  • �� Radar-KISSICP: Based on KISS-ICP, incorporates a lightweight motion compensation step to produce sparse but 3D-aware radar pointclouds, reducing planar assumptions.

  • �� Radar-IMU: Leverages IMU preintegration to provide a robust initial guess for radar scan-matching, enhancing vertical displacement awareness and improving recovery from drift. These methods were validated on the GO dataset, demonstrating improved trajectory estimation in complex terrains.

Methodology

  • �� Radar-KISSICP: Based on KISS-ICP, incorporates a lightweight motion compensation step to produce sparse but 3D-aware radar pointclouds. Input: radar feature points, Process: rotation compensation, Output: 3D-aware pointclouds.

  • �� Radar-IMU: Leverages IMU preintegration to provide a robust initial guess for radar scan-matching. Input: radar scans and IMU data, Process: IMU data integration, Output: stabilized trajectory estimation.

  • �� Experiments were conducted on the GO dataset to evaluate the performance of these methods in complex terrains.

Experiments

Experiments were conducted on the GO dataset to evaluate the performance of Radar-KISSICP and Radar-IMU in complex terrains. The dataset includes forest trails, ravines, and snow-covered terrain with frequent pitch/roll excitation. The experimental design includes comparisons with CFEAR and ORORA methods, using KITTI and EVO standard odometry metrics for evaluation. Key hyperparameters include settings for motion compensation and IMU preintegration.

Results

Experimental results demonstrate that Radar-KISSICP and Radar-IMU significantly improved trajectory estimation accuracy in challenging off-road routes. For instance, on Route 3, Radar-IMU achieved an absolute trajectory error of 15.12%, compared to 61.08% for CFEAR. Radar-IMU demonstrated excellent performance in handling terrain discontinuities and severe pitch/roll scenarios, particularly recovering vertical displacement in the ravine traversal of Route 3.

Applications

These methods can be directly applied to off-road robotic navigation and autonomous vehicle localization systems, particularly in complex terrains and adverse weather conditions. Prerequisites include equipping with millimeter-wave radar and IMU sensors. The application of these technologies will significantly enhance autonomous navigation capabilities in off-road environments, driving the development of related industries.

Limitations & Outlook

Despite the excellent performance of Radar-KISSICP and Radar-IMU in off-road environments, there are still some limitations. Radar-KISSICP suffers from unstable nearest-neighbor associations due to sparse radar pointclouds, potentially affecting accuracy in some scenarios. Radar-IMU, while improving vertical displacement awareness, may still face inertial drift issues without bias correction. Additionally, radar odometry may still face challenges in extremely complex terrains, requiring further sensor fusion. Future research could explore more complex multi-sensor fusion techniques to enhance localization accuracy in off-road environments.

Plain Language Accessible to non-experts

Imagine you're driving in a large forest, surrounded by tall trees and uneven ground. Traditional GPS might fail due to poor signal, and cameras and LiDAR don't work well in fog or heavy rain. That's where radar comes in, like a super sensor that can see through fog and rain, helping you find the right direction in the forest.

Radar-KISSICP and Radar-IMU are like two powerful assistants for your super sensor. Radar-KISSICP is like a smart navigator that adjusts your route based on terrain changes, keeping you steady on bumpy roads. Radar-IMU is like a keen scout that senses ground undulations, helping you stay on course in complex terrains.

Together, these technologies ensure you won't get lost or stuck while driving in the forest. Even when encountering ravines or steep slopes, they help you find the best route, making your journey safer and smoother.

ELI14 Explained like you're 14

Hey there! Imagine you're playing a super cool off-road racing game. You're driving through a big forest with lots of tall trees and bumpy ground. Traditional GPS might fail due to poor signal, and cameras and LiDAR don't work well in fog or heavy rain. That's where radar comes in, like a super sensor that can see through fog and rain, helping you find the right direction in the forest.

Radar-KISSICP and Radar-IMU are like two powerful assistants for your super sensor. Radar-KISSICP is like a smart navigator that adjusts your route based on terrain changes, keeping you steady on bumpy roads. Radar-IMU is like a keen scout that senses ground undulations, helping you stay on course in complex terrains.

Together, these technologies ensure you won't get lost or stuck while driving in the forest. Even when encountering ravines or steep slopes, they help you find the best route, making your journey safer and smoother. Isn't that cool?

Glossary

Radar Odometry

Radar odometry is a technique that uses radar sensors to measure relative motion, particularly effective in adverse weather and complex terrains.

Used in the paper to evaluate radar's localization capabilities in off-road environments.

Motion Compensation

Motion compensation is a technique that adjusts sensor data to offset motion effects, ensuring data accuracy.

Used in Radar-KISSICP to generate 3D-aware radar pointclouds.

IMU Preintegration

IMU preintegration is a technique that uses inertial measurement unit data for trajectory estimation, providing initial motion guesses.

Used in Radar-IMU to stabilize scan matching.

Absolute Trajectory Error (ATE)

Absolute trajectory error is a metric for evaluating trajectory estimation accuracy, indicating the deviation between estimated and true trajectories.

Used to evaluate the performance of Radar-KISSICP and Radar-IMU.

CFEAR

CFEAR is a feature-based radar odometry method suitable for localization in urban environments.

Used as one of the baseline comparison methods.

ORORA

ORORA is a robust radar odometry method capable of maintaining performance in the presence of numerous outliers.

Used as one of the baseline comparison methods.

Sparse Pointcloud

A sparse pointcloud refers to pointcloud data with low point density, potentially leading to poor registration.

Causes unstable nearest-neighbor associations in Radar-KISSICP.

Inertial Drift

Inertial drift is a deviation in trajectory estimation caused by accumulated inertial measurement errors.

May affect accuracy in Radar-IMU.

Unstructured Environment

An unstructured environment refers to a terrain that is complex and variable, lacking obvious landmarks.

Presents challenges for radar odometry.

Multi-Sensor Fusion

Multi-sensor fusion is a technique that combines data from multiple sensors to enhance system performance.

Explored as a future direction in the paper.

Open Questions Unanswered questions from this research

  • 1 The performance of radar odometry in extremely complex terrains remains to be further studied, especially in scenarios with severe terrain undulations and sparse landmarks. Current methods may face challenges in these scenarios, requiring more complex multi-sensor fusion techniques.
  • 2 How to effectively combine 3D radar and visual/LWIR cameras to enhance localization accuracy in off-road environments remains an open question. Existing methods have limitations in addressing sparse radar pointclouds and inertial drift.
  • 3 Radar-KISSICP and Radar-IMU perform well in handling terrain discontinuities and severe pitch/roll scenarios, but IMU integration may face inertial drift issues without bias correction.
  • 4 In off-road environments, optimizing IMU integration methods to reduce inertial drift is an important direction for future research. Current methods still have room for improvement in this regard.
  • 5 The performance of radar odometry under adverse weather conditions, especially in extreme weather conditions such as fog and heavy rain, needs further validation.

Applications

Immediate Applications

Off-road Robotic Navigation

Radar-KISSICP and Radar-IMU can be directly applied to off-road robotic navigation systems, particularly in complex terrains and adverse weather conditions.

Autonomous Vehicle Localization

These technologies can be used in autonomous vehicle localization systems to enhance autonomous navigation capabilities in complex terrains.

Drone Navigation

The combination of radar and IMU technologies can enhance drone navigation capabilities in complex environments, particularly in areas with significant terrain undulations.

Long-term Vision

Fully Autonomous Off-road Vehicles

By integrating multi-sensor fusion technologies, achieve fully autonomous off-road vehicle navigation, overcoming challenges of complex terrains and adverse weather.

Smart City Infrastructure

Utilize radar odometry technology to enhance the sensing capabilities of smart city infrastructure, improving the safety and efficiency of urban transportation systems.

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 $SE(3)$ 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.

cs.RO

References (20)

Lidar-Level Localization With Radar? The CFEAR Approach to Accurate, Fast, and Robust Large-Scale Radar Odometry in Diverse Environments

Daniel Adolfsson, Martin Magnusson, Anas W. Alhashimi et al.

2022 54 citations ⭐ Influential View Analysis →

KISS-ICP: In Defense of Point-to-Point ICP – Simple, Accurate, and Robust Registration If Done the Right Way

Ignacio Vizzo, Tiziano Guadagnino, Benedikt Mersch et al.

2022 496 citations ⭐ Influential View Analysis →

Vision meets robotics: The KITTI dataset

Andreas Geiger, Philip Lenz, C. Stiller et al.

2013 10243 citations

ORORA: Outlier-Robust Radar Odometry

Hyungtae Lim, Kawon Han, Gunhee Shin et al.

2023 24 citations View Analysis →

Object recognition from local scale-invariant features

D. Lowe

1999 18718 citations

DRO: Doppler-Aware Direct Radar Odometry

C. Gentil, Leonardo Brizi, Daniil Lisus et al.

2025 7 citations View Analysis →

Pointing the Way: Refining Radar-Lidar Localization Using Learned ICP Weights

Daniil Lisus, Johann Laconte, Keenan Burnett et al.

2023 10 citations View Analysis →

A New Wave in Robotics: Survey on Recent MmWave Radar Applications in Robotics

Kyle Harlow, Hyesu Jang, T. Barfoot et al.

2023 75 citations View Analysis →

Under the Radar: Learning to Predict Robust Keypoints for Odometry Estimation and Metric Localisation in Radar

Dan Barnes, I. Posner

2020 133 citations View Analysis →

Boreas: A multi-season autonomous driving dataset

Keenan Burnett, David J. Yoon, Yuchen Wu et al.

2022 186 citations View Analysis →

ORB: An efficient alternative to SIFT or SURF

Ethan Rublee, V. Rabaud, K. Konolige et al.

2011 10661 citations

The Oxford Radar RobotCar Dataset: A Radar Extension to the Oxford RobotCar Dataset

Dan Barnes, Matthew Gadd, P. Murcutt et al.

2019 445 citations View Analysis →

Radar Odometry Combining Probabilistic Estimation and Unsupervised Feature Learning

Keenan Burnett, David J. Yoon, Angela P. Schoellig et al.

2021 68 citations View Analysis →

Precise Ego-Motion Estimation with Millimeter-Wave Radar Under Diverse and Challenging Conditions

Sarah H. Cen, P. Newman

2018 192 citations

A Normal Distribution Transform-Based Radar Odometry Designed For Scanning and Automotive Radars

Pou-Chun Kung, C. Wang, Wen-Chieh Lin

2021 68 citations View Analysis →

GO: The Great Outdoors Multimodal Dataset

Peng Jiang, Kasi Viswanath, Akhil Nagariya et al.

2025 5 citations View Analysis →

Instantaneous ego-motion estimation using Doppler radar

Dominik Kellner, M. Barjenbruch, J. Klappstein et al.

2013 170 citations

MulRan: Multimodal Range Dataset for Urban Place Recognition

Giseop Kim, Yeong-Sang Park, Younghun Cho et al.

2020 336 citations

Are Doppler Velocity Measurements Useful for Spinning Radar Odometry?

Daniil Lisus, Keenan Burnett, David J. Yoon et al.

2024 15 citations View Analysis →

Parallel Maximum Clique Algorithms with Applications to Network Analysis

Ryan A. Rossi, D. Gleich, A. Gebremedhin

2013 120 citations View Analysis →