SPARO Lab

RadLoc: Radar-based 3-DoF Global Localization
via Fast, Robust, and Lightweight Spatial Descriptor
Across Diverse Environmental Scenarios

IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2026
Hogyun Kim1, Jiwon Choi1, Jungwoo Lee1, and Younggun Cho1†
† Corresponding author
1Spatial AI and Robotics (SPARO) Lab, Inha University, South Korea
Paper (TBD) arXiv Video Code
RadLoc Overview
TL;DR: Fast, Robust, and Lightweight Radar-based 3-DoF Global Localization

Abstract

While global localization using spinning radar has gained attention for its robustness to adverse weather and challenging environments, many studies have focused on individual components such as place recognition or pose estimation. In this paper, we take a holistic view of radar sensor-based global localization and present RadLoc, a fast, robust, and lightweight end-to-end pipeline from place recognition to 3-DoF pose estimation. RadLoc accelerates pre-processing using 1D CA-CFAR filtering and leverages the near-range dominance in spinning radar images to design a compact descriptor and an efficient hierarchical coarse-to-fine retrieval strategy. Moreover, coupled with phase correlation-based 3-DoF pose estimation, it forms a versatile global localization module applicable to SLAM and multi-session SLAM systems. Extensive experiments on 15 sequences across 5 datasets demonstrate that RadLoc achieves robust performance while maintaining the smallest descriptor size and fastest retrieval time among state-of-the-art approaches.

Motivation

Radar-based global localization has advanced rapidly, but most prior work optimizes a single component in isolation. As a result, each line of research is tuned for its own sub-problem rather than for a complete localization system.

Place Recognition

Many radar studies focus solely on recognizing revisited places, designing descriptors and retrieval pipelines while leaving metric pose estimation out of scope.

Pose Estimation

Others concentrate on estimating relative or absolute pose from radar scans, assuming the matching place is already known or coarsely provided.

SLAM

Radar SLAM does cover the full localization-and-mapping loop, but higher-level tasks such as multi-session SLAM remain underexplored and insufficiently supported.

Our RadLoc takes a holistic view, covering the full stack from place recognition and 3-DoF pose estimation to SLAM and multi-session SLAM in one fast, robust, and lightweight pipeline.

Demo

RadLoc runs end-to-end across the full localization stack: place recognition, 3-DoF pose estimation, and SLAM, including multi-session settings under long-term and cross-weather variations.

Demo #1: Place Recognition
Demo #2: Multi-Session Place Recognition
Demo #3: SLAM
Demo #4: Multi-Session SLAM

Pipeline

An overview of the RadLoc pipeline is summarized below. Given a polar radar image from a spinning radar sensor, we convert it into a Cartesian representation for pose estimation while retaining the original polar image for descriptor construction. We then apply lightweight pre-processing to suppress radar-specific noise, and the processed image is used to construct a compact descriptor, which is stored in a database for efficient retrieval. At query time, RadLoc performs hierarchical coarse-to-fine place retrieval against the database, and retrieves the corresponding polar and Cartesian images of the matched place for 3-DoF pose estimation via phase correlation.

RadLoc pipeline

Evaluation

We evaluate RadLoc on five public radar datasets (Oxford Radar Robotcar, OORD, MulRan, Boreas, and Hercules) spanning structured urban and unstructured off-road scenes, multiple weather conditions, and three Navtech radar sensors. We compare against learning-based (SHeRLoc, Kidnapped Radar) and handcrafted (RSC, RaPlace, RadVLAD, RadFFTVLAD, ReFeree) descriptors, and a feature + GICP pipeline for pose estimation. In every table, the top three results are highlighted as 1st, 2nd, and 3rd.

Robust

Single-Session Place Recognition

Mean R@1 over 15 sequences from 5 datasets. Higher is better; best highlighted.

R@1 @10 m ↑

Kidnapped0.349
SHeRLoc0.528
RSC0.655
RaPlace0.489
RadVLAD0.466
RadFFTVLAD0.624
ReFeree0.695
RadLoc0.731

R@1 @20 m ↑

Kidnapped0.361
SHeRLoc0.529
RSC0.674
RaPlace0.470
RadVLAD0.458
RadFFTVLAD0.645
ReFeree0.716
RadLoc0.750
Higher is better (↑).Best value highlighted.
Robust

Multi-Session Place Recognition

Mean over 3 OORD sequences (Bellmouth, Hydro, Maree). Higher is better; best highlighted.

R@1 ↑

Kidnapped0.006
SHeRLoc0.966
RSC0.787
RaPlace0.539
RadVLAD0.007
RadFFTVLAD0.900
ReFeree0.967
RadLoc0.986

AUC ↑

Kidnapped0.006
SHeRLoc0.891
RSC0.830
RaPlace0.233
RadVLAD0.008
RadFFTVLAD0.623
ReFeree0.840
RadLoc0.926

F1 ↑

Kidnapped0.013
SHeRLoc0.881
RSC0.756
RaPlace0.471
RadVLAD0.016
RadFFTVLAD0.744
ReFeree0.851
RadLoc0.887
Higher is better (↑).Best value highlighted.
Accurate & Fast

3-DoF Pose Estimation

Feat. + GICP RadLoc (Proposed)

Mean over the 7 Hercules sequences (Mountain 01-03, Parking 01-04) of the per-sequence values in the paper.

RTE (m) ↓

Feat. + GICP 0.78
RadLoc 0.68

RRE (°) ↓

Feat. + GICP 3.24
RadLoc 1.91

Success Rate (%) ↑

Feat. + GICP 96.2
RadLoc 99.8

Runtime (s) ↓

Feat. + GICP 0.296
RadLoc 0.057
↓ lower is better, ↑ higher is better. Best value highlighted.
Lightweight

Descriptor Size

Total descriptor storage across two sessions (sunny + snowy) on Hercules. Smaller is better for scalable multi-session map management.

Descriptor Size (MB) ↓

Kidnapped 499.31
SHeRLoc 4.88
RSC 4.57
RaPlace 10.98
RadVLAD 998.62
ReFeree 1.30
RadLoc 1.22
Lower is better (↓). Bars are log-scaled so MB-scale bars stay visible against the GB-scale baselines.

Ablation Study

We ablate the two core design choices of RadLoc: the range-dependent weighting and the hierarchical coarse-to-fine retrieval, and analyze the sensitivity to the radial partition index.

Mountain 03

I. Radial Partition K

Larger K improves R@1 by using more structure, while AUC and F1 decrease as far-range bins enter the coarse stage. Higher is better; K = 20 (selected default) highlighted.

R@1 ↑

K=80.934
K=100.946
K=200.960
K=400.972

AUC ↑

K=80.897
K=100.901
K=200.893
K=400.875

F1 ↑

K=80.803
K=100.800
K=200.790
K=400.767
Boreas 21-01-19

Radial Partition K

Higher is better for all metrics. K = 20 (selected default) highlighted.

R@1 ↑

K=80.354
K=100.387
K=200.499
K=400.551

AUC ↑

K=80.654
K=100.650
K=200.582
K=400.546

F1 ↑

K=80.603
K=100.582
K=200.503
K=400.459
Discriminability

II. Range-dependent Weighting

The proposed weighting aligns descriptor distances with ground-truth spatial separation: it yields lower Kolmogorov-Smirnov statistics than the unweighted variant for both positive and negative pairs, mitigating the near-range bias.

KS plot: range-dependent weighting
Recall preserved

III. Hierarchical Retrieval

Hierarchical coarse-to-fine retrieval preserves recall comparable to exhaustive flat search while substantially cutting runtime, with the advantage growing as the database size increases.

PR curves: hierarchical vs flat retrieval
Scalable

IV. Runtime

The retrieval runtime of the hierarchical search scales far better than exhaustive flat search, with the gap widening as the database grows.

Retrieval runtime: hierarchical vs flat

BibTeX

@inproceedings{kim2026radloc,
  author    = {Kim, Hogyun and Choi, Jiwon and Lee, Jungwoo and Cho, Younggun},
  title     = {RadLoc: Radar-based 3-DoF Global Localization via Fast, Robust, and Lightweight Spatial Descriptor Across Diverse Environmental Scenarios},
  booktitle = {IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)},
  year      = {2026},
}

@article{kim2026radloc,
  title={RadLoc: Radar-based 3-DoF Global Localization via Fast, Robust, and Lightweight Spatial Descriptor Across Diverse Environmental Scenarios},
  author={Kim, Hogyun and Choi, Jiwon and Lee, Jungwoo and Cho, Younggun},
  journal={arXiv preprint arXiv:2607.08115},
  year={2026}
}