Radar Odometry Subject to High Tilt Dynamics of Subarctic Environments

Abstract: Rotating FMCW radar odometry methods often assume flat ground conditions. While this assumption is sufficient in many scenarios, including urban environments or flat mining setups, the highly dynamic terrain of subarctic environments poses a challenge to standard feature extraction and state estimation techniques. This paper benchmarks three existing radar odometry methods under demanding conditions, exhibiting up to 13° in pitch and 4° in roll difference between consecutive scans, with absolute pitch and roll reaching 30° and 8°, respectively. Furthermore, we propose a novel radar-inertial odometry method utilizing tilt-proximity submap search and a hard threshold for vertical displacement between scan points and the estimated axis of rotation. Experimental results demonstrate a state-of-the-art performance of our method on an urban baseline and a 0.3% improvement over the second-best comparative method on a 2-kilometer-long dynamic trajectory. Finally, we analyze the performance of the four evaluated methods on a complex radar sequence characterized by high lateral slip and a steep ditch traversal.

• The paper presents a radar-inertial odometry technique that compensates for extreme pitch and roll dynamics in subarctic environments.
• It integrates high-frequency IMU data with tilt-aware submap retrieval and vertical filtering to enhance motion compensation and point cloud registration.
• Experimental results on the FoMo dataset show a median RTE of 2.8%, outperforming other state-of-the-art radar odometry methods under challenging terrain conditions.

Radar Odometry Under High Tilt Dynamics in Subarctic Terrains

Introduction and Motivation

The paper "Radar Odometry Subject to High Tilt Dynamics of Subarctic Environments" (2604.19962) systematically investigates the robustness and adaptability of rotating Frequency Modulated Continuous Wave (FMCW) radar odometry under the severe tilt dynamics encountered in subarctic environments. Traditional radar odometry pipelines are generally evaluated in milder, urban or industrial settings where pitch and roll remain limited. The authors address the gap by exposing contemporary radar odometry methods to abrupt terrain-induced motion, including pitch changes of up to $30^\circ$ and roll up to $8^\circ$, as well as substantial inter-scan tilt discontinuities. The core claim is the introduction of a radar-inertial odometry technique that, via tilt-proximity-based submap association and vertical displacement filtering, delivers superior robustness and accuracy under such non-ideal, dynamic conditions.

Figure 1: Typical rotating FMCW radar solutions are contrasted between flat and high-dynamic terrains, highlighting the severity of tilt changes in the latter.

Methodology

The proposed approach extends classical point-based radar odometry with two key innovations, leveraging inertial measurements for tilt compensation at multiple stages of the pipeline. The methodology comprises:

• IMU Integration and Tilt-Aware Submap Retrieval: High-frequency IMU data (200 Hz) is preprocessed, orientation estimates are computed using the Madgwick filter, and used to select submaps with minimal spatial and tilt (pitch/roll) discrepancy relative to the current vehicle attitude. If no suitable submap matches within given tilt/distance thresholds, velocity estimates are reset, ensuring conservative state propagation in extreme cases.
• Feature Extraction and Motion Compensation: Radar intensity images are processed to extract strong scattering features, with noise and irrelevant returns excluded through spatial limits and intensity-based thresholds. Each feature’s position is deskewed by interpolating the instantaneous vehicle orientation, ensuring accurate point cloud reconstruction during motion.
• Tilt-Based Scan Filtering: After motion compensation, each radar point is assessed for its vertical deviation from the estimated rotation axis corresponding to the current tilt. Points exhibiting excessive deviations, measured against a hard threshold after Cauchy weighting, are culled.
• ICP Registration with Subsampling: Both live and reference submap clouds are voxelized, centroids are retained, and ICP correspondence is executed with a small neighborhood parameter.
• Submap Update Strategy: New submaps are initiated based on spatial or tilt increments, maintaining global consistency despite dramatic changes in vehicle attitude.

  Figure 2: Schematic of the radar-inertial odometry pipeline, with bottom images detailing sequential processing of radar data from raw input to filtered point clouds.

Experimental Setup

The evaluation leverages the FoMo dataset, collected in the subarctic region of Quebec, using a Navtech CIR-304 radar and VectorNav VN-100 IMU mounted on a Clearpath Warthog UGV. The dataset is uniquely challenging, covering a 300 m urban loop (Red) and a 2 km dynamic trajectory (Orange) that includes a stone quarry, steep ditches, and lateral slip in conditions with snow banks up to 1 m high.

The proposed method is benchmarked against three state-of-the-art radar odometry algorithms:

• CFEAR (k-strongest feature extraction with direct matching)
• ORORA (decoupled rotation/translation and anisotropic radar noise handling)
• RTR (radar-gyro teach and repeat, with heading from IMU z-axis angular velocity)

  Figure 3: UGV equipped with rotating radar and IMU, depicted traversing a steep ditch, indicative of the acute orientation changes encountered.

Results and Analysis

The main evaluation metrics include Relative Translation Error (RTE) and qualitative trajectory plots. The results demonstrate:

• Urban Baseline: All methods perform adequately on the flat Red loop, but orientation errors and drift escalate sharply in the Orange sequence.
• Dynamic Trajectories: The new method achieves a median RTE of 2.8% on the dynamic loop, outperforming CFEAR (3.1%), ORORA (substantially worse due to heading loss), and RTR (catastrophic failures in 2/9 runs due to excessive tilt).
• Heading Robustness: By leveraging full 3D orientation from IMU, as opposed to single-axis priors, the method consistently constrains heading drift even under extreme pitch/roll variations.

  Figure 4: RTE distribution for the four methods across urban and dynamic experiments, highlighting the proposed method’s lower median errors and reduced outliers in complex terrain.

Visual inspection of estimated trajectories in the most demanding runs shows:

• Mode Failures: RTR and ORORA suffer from rapid divergence in high-slip and extreme pitch/roll segments, sometimes crossing the ground-truth path or diverging by tens of meters.
• Corridor Effects: CFEAR, while more robust than the preceding methods, is sensitive to geometric aliasing and artifacts from snowbanks, leading to bias in some segments.
• Persistence in Adversity: The proposed method completes all loops, suffers no catastrophic failures, and exhibits superior recovery following large inclination events, albeit with some residual drift.

  Figure 5: Output trajectory comparison in a winter scenario, demonstrating severe deviations in RTR and ORORA versus contained error in the proposed radar-inertial method.

A close-up analysis in a quarry section exposes the limits of FMCW radar odometry when compounded by high lateral slip and abrupt slope transitions; all methods display increased high-frequency pose noise and occasional discontinuous jumps, emphasizing an inherent limitation in discrete radar feature-based registration under extreme terrain-induced motions.

Figure 6: Detailed trajectory view in a stone quarry segment; while all methods show artifacts, the proposed approach displays less gross deviation and better heading stability through the ditch and incline.

Implications and Future Directions

The findings underscore the essential value of tightly fusing inertial and radar data for odometry in non-planar, non-urban domains, such as mining, forestry, and remote inspection. The tilt-aware submap selection and per-point vertical filtering are shown to be effective countermeasures against feature misassociation and orientation drift in the presence of severe pitch/roll. These techniques may generalize to other forms of sparse sensor registration in robotics.

Key practical implications include improved navigation reliability for UGVs operating in northern and off-road zones, where vision and lidar experience degraded performance. However, the residual high-frequency noise and pose discontinuities necessitate further research, possibly involving smoothing, learned priors, or hybrid dense-sparse strategies. The authors propose future integration of translational priors and benchmarking against dense radar methods; validation on large-scale urban datasets could further test the limits of the architecture.

Conclusion

This study delivers a comprehensive, well-controlled comparative analysis of radar odometry under the extreme dynamics of subarctic terrains and introduces a radar-inertial odometry framework that defines a new robustness frontier for FMCW radar-based localization. The results have practical relevance for autonomous navigation in highly dynamic and unstructured environments, with fertile potential for extensions into denser radar registration paradigms and robust smoothing for high-frequency motion artifacts.