Pushing Radar Odometry Beyond the Pavement: Current Capabilities and Challenges

1. Problem Statement

While millimeter-wave radar has proven robust for localization in urban, planar environments (SE(2)) due to its immunity to weather, lighting, and particulates, its performance in unstructured off-road environments remains poorly understood.

• The Gap: State-of-the-art (SOTA) radar odometry methods (e.g., CFEAR, ORORA) are designed for automotive scenarios where vehicle motion is largely constrained to a flat plane. They implicitly assume limited pitch and roll.
• The Challenge: Off-road terrain introduces full SE(3) motion (including significant pitch, roll, and vertical displacement). In these conditions:
  • Ground returns become confounding signals rather than removable outliers.
  • Sparse radar point clouds lead to ill-conditioned registration problems.
  • False correspondences accumulate in Iterative Closest Point (ICP) pipelines, causing catastrophic trajectory drift.
  • Existing methods fail to generalize, often breaking down in scenarios like ravine traversals or steep terrain.

2. Methodology

The authors propose a benchmarking framework to evaluate SOTA methods and introduce two diagnostic baselines to address off-road specific failures.

A. Benchmarking Framework

The study evaluates four methods on both automotive and off-road datasets:

1. CFEAR: A surfel-based feature extraction method using point-to-line error optimization. It relies on keyframing and assumes planar motion.
2. ORORA: An outlier-robust method using decoupled matching, Doppler compensation, and maximum clique inlier selection.
3. Radar-KISSICP (Proposed Baseline): Adapts the KISS-ICP algorithm to radar. It applies SO(3) motion compensation to generate sparse, 3D-aware point clouds. It removes the planar assumption but suffers from sparsity-induced instability.
4. Radar-IMU (Proposed Baseline): Extends Radar-KISSICP by integrating IMU preintegration (via GTSAM). This provides a robust initial SE(3) pose guess, introducing vertical displacement awareness that radar alone cannot observe.

B. Datasets

• Automotive (Control Group): Oxford RobotCar, MulRan, and Boreas. These represent structured, planar SE(2) environments.
• Off-Road (Test Group): The Great Outdoors (GO) dataset, collected using a Clearpath Warthog UGV with a Navtech radar. It includes three routes with increasing difficulty:
  • Route 1: Paved area (low difficulty).
  • Route 3: Gravel/forest with ravines (high pitch/roll).
  • Route 4: Mixed off-road deviations (complex geometry).

C. Evaluation Metrics

• Relative Pose Error (RPE): Measures local drift over fixed segments.
• Absolute Trajectory Error (ATE): Measures accumulated global drift.
• SE(3) Analysis: All methods are evaluated in full 6-DoF space, analyzing the deviation of ground-truth transforms from the XY-plane.

3. Key Contributions

1. Systematic Evaluation: The first comprehensive benchmark of radar odometry in off-road, SE(3) environments, demonstrating that SOTA urban methods fail significantly when pitch/roll excitation increases.
2. Diagnostic Baselines:
   • Radar-KISSICP: Demonstrates that applying SO(3) motion compensation reduces planar bias but highlights the instability caused by sparse radar features.
   • Radar-IMU: Proves that even simple IMU preintegration (without bias correction) significantly stabilizes trajectory estimation by recovering vertical displacement and rejecting false ground returns.
3. Failure Mode Analysis: Identifies that in off-road settings, ground returns are not outliers but integral parts of the signal. The paper analyzes how ravine traversals cause "catastrophic drift" due to the loss of stable features and the dominance of wall reflections.

4. Results

• Automotive Datasets: Both CFEAR and ORORA perform competitively, confirming their suitability for planar SE(2) motion.
• Off-Road Datasets (GO):
  • CFEAR & ORORA: Show substantial drift. ORORA fails completely on Route 4. CFEAR produces plausible local estimates but accumulates large global errors due to false matches in ravines.
  • Radar-KISSICP: Improves performance over CFEAR/ORORA on Route 1 and 3 by reducing planar bias, but struggles with convergence due to point cloud sparsity.
  • Radar-IMU: Achieves the best performance across all off-road routes.
    • On Route 3 (ravine traversal), Radar-IMU successfully recovers the trajectory where Radar-KISSICP diverges.
    • It significantly reduces ATE (e.g., Route 3 ATE translation drops from ~61m for CFEAR to ~15m for Radar-IMU).
• Key Observation: Local relative pose errors can be small, but isolated severe errors (e.g., at the start of a ravine) propagate to cause massive global drift.

5. Significance and Future Directions

• Paradigm Shift: The paper argues that off-road radar odometry cannot rely on treating ground returns as outliers. Instead, systems must treat them as integral signal components and rely on inertial cues to disambiguate vertical motion.
• Limitations of Current SOTA: Urban success does not transfer to off-road; the assumption of flat terrain is a critical bottleneck.
• Future Work: The authors suggest that robust off-road localization requires:
  • Tighter Multi-sensor Fusion: Specifically combining Radar with IMU, LiDAR, and cameras.
  • 3D Radar: Moving beyond 2D scanning radars to capture full volumetric data.
  • Learned Approaches: Using deep learning for denoising and feature selection to handle sparse and unstable radar returns.

In conclusion, this paper establishes that while radar is a promising sensor for all-weather off-road robotics, current algorithms are insufficient. The integration of inertial data (IMU) and motion compensation is essential to bridge the gap between radar's weather robustness and the geometric complexity of natural terrain.