RadGeoLoc: Radar Geolocalization via Particle Filtering

Shaunak Kolhe; Peng Jiang; Kasi Viswanath; Srikanth Saripalli

cs.ROOct 19, 2026

RadGeoLoc: Radar Geolocalization via Particle Filtering

Authors: Shaunak Kolhe, Peng Jiang, Kasi Viswanath, Srikanth Saripalli

Organizations: J. Mike Walker ’66 Department of Mechanical Engineering, Texas A&M University, College Station, TX 77840, USA

Abstract

In real-time autonomous field applications, GNSS is often unreliable and inaccurate. A possible solution is localizing against a prebuilt map like a satellite image, as it is georeferenced and already surveyed. In this work, we localize a spinning radar measurement against overhead RGB imagery with a learned observation model. We use a radar representation pretrained by masked reconstruction and align it to features from a satellite image. A particle filter propagated by radar odometry proposes candidate poses, and our observation model provides a likelihood over these poses. We evaluate our method over 31 distinct routes, over 8 datasets spanning both urban and offroad conditions using only ground radar and a coarse local prior. We also examine the robustness over changing weather using repeated traversals over the same route over 13 months, in both urban and offroad conditions.

Figures & tables

Fig. 1: The premise, on Boreas. (a) the satellite image crop the particles are scored against, with the radar footprint boxed. (b) the Cartesian radar sweep at the same instant. (c) The two overlaid at the filter’s selected estimate from the likelihood. The structure the radar returns coincides with the structure in the image, which is what makes them comparable. The 128 particles are drawn on top, the one nearest ground truth in green, and the rest in yellow.

Fig. 2: Network architecture of the observation model.

Method	ATE RMSE				Decomposition
Method	01	02	03	04	Lat.	Long.	Yaw
Hong [ 7 ]	$4.70$	$4.90$	$4.90$	$3.80$	–	–	–
Li [ 37 ]	$3.00$	$3.00$	$3.40$	$3.70$	–	–	–
RaSCL [ 11 ] ∗	1.30	1.50	1.60	1.60	1.08	$1.30$	$3.13$
Ours	$1.80$	$1.82$	$1.83$	$1.73$	$1.57$	0.88	1.11

TABLE I: Comparison on the Boreas sequences of [ 11 ] .

Dataset	Sequence	ATE	Med.	Lat.	Long.	Yaw
GOd (train)	morning_run_1	$269.75$	$4.53$	$2.43$	$2.85$	$9.99$
	route1_morning	$20.82$	$2.59$	$3.15$	$4.52$	$6.15$
	route_1_run_2	$2.51$	$1.93$	$1.18$	$2.21$	$2.47$
HeRCULES (test)	Bridge01	$978.61$	$794.89$	$4.49$	$8.44$	$2.90$
	Library01	$6.98$	$3.72$	$3.75$	$5.89$	$4.25$
	Mountain01	$5.96$	$4.53$	$4.86$	$3.45$	$2.17$

TABLE II: Per-route results.

Fig. 3: Cumulative distribution of per-frame position error, by dataset. The dashed line is the blind radius; a dataset whose curve reaches 1.0 to the left of it never lost lock.

Fig. 4: Estimated trajectories on four routes spanning four datasets, plotted with ground truth and the scan-matching odometry the filter propagates on. While the odometry drifts on every route, the filter estimate stays on ground truth, showing that the observation model provides meaningful corrections.

Date	Conditions	ATE	Median	Lat.	Long.	Yaw
		[ $\mathrm{m}$ ]	[ $\mathrm{m}$ ]	[ $\mathrm{m}$ ]	[ $\mathrm{m}$ ]	[ $\mathrm{\SIUnitSymbolDegree}$ ]
2020-12-18	Sun, Snow	$2.68$	$1.34$	$2.51$	$0.93$	$1.34$
2021-01-15	Sun, Clouds, Snow	$1.98$	$1.34$	$1.77$	$0.88$	$1.25$
2021-01-19	Clouds, Snow	$1.97$	$1.25$	$1.78$	$0.84$	$1.34$
2021-01-26	Overcast, Snow, Snowing	$2.27$	$1.58$	$1.92$	$1.21$	$1.35$
2021-02-02	Overcast, Snow	$2.02$	$1.35$	$1.81$	$0.90$	$1.31$

TABLE III: Boreas: one route, thirteen months. One traversal for every unique weather condition the dataset labels (14 of them).

yellow route
Date	Snow	Rain	RH	Temp.	ATE	Median
	[ $\mathrm{mm}$ ]	[ $\mathrm{mm}$ ]	[ $\mathrm{\%}$ ]	[ $\mathrm{\SIUnitSymbolCelsius}$ ]	[ $\mathrm{m}$ ]	[ $\mathrm{m}$ ]
2024-11-28	–	$0.0$	$89$	$-2.5$	$13.20$	$5.96$
2025-03-10	$1060$	$0.0$	$79$	$-7.1$	$84.70$	$7.60$
2025-08-20	$0$	$0.0$	$45$	$18.3$	$10.47$	$4.90$
2025-09-24	$0$	$0.0$	$50$	$16.2$	$9.00$	$5.32$

TABLE IV: FoMo: two off-road routes, six traversals each over a year.

		MAE init		Random init
Dataset	$n$	ATE	Lost	ATE	Lost
		[ $\mathrm{m}$ ]		[ $\mathrm{m}$ ]
Boreas (Glen Shields)	$14$	2.06	0/14	$1442.66$	14/14
FoMo	$12$	$13.14$	4/12	12.43	1/12
HeRCULES	$7$	6.98	1/7	$259.36$	6/7
Oxford	$1$	7.39	0/1	$656.32$	1/1

TABLE V: Encoder initialisation. Median ATE over each dataset’s traversals, and the number that lose lock. Held-out datasets above the rule, training datasets below.

Stage	Time [ $\mathrm{ms}$ ]	Share [ $\mathrm{\%}$ ]
Aerial feature crop	$24.73$	$81.2$
Host-to-device transfer	$2.12$	$7.0$
Radar encoder	$2.38$	$7.8$
Aerial projection	$0.26$	$0.9$
Scoring, 128 particles	$0.86$	$2.8$
Device-to-host, filter update	$0.12$	$0.4$

TABLE VI: Per-frame runtime of the localization stage.

Fig. 5: HeRCULES Street01, with the estimate coloured by per-frame error and the dashed box marking the zoom. Ground truth and the estimate run parallel, separated by a sustained offset toward the buildings.

Map-Based Localization

Map-Based Localization estimates a device or robot’s position and orientation by matching sensor observations to stored maps.

See all 106 papers

CFEAR-Teach-and-Repeat: Fast and Accurate Radar-only Localization
Mar 6, 2026Maximilian Hilger, Daniel Adolfsson, Ralf Becker +2Robot Localization Map-Based Localization
Chair of Perception for Intelligent Systems, Munich Institute of Robotics and Machine Intelligence, Technical University of Munich, Munich, Germany · Robot Navigation and Perception Lab of the AASS Research Center, Örebro University, Örebro, Sweden · Bosch Rexroth, Germany
RadLoc: Radar-based 3-DoF Global Localization via Fast, Robust, and Lightweight Spatial Descriptor Across Diverse Environmental Scenarios
Jul 9, 2026Hogyun Kim, Jiwon Choi, Jungwoo Lee +1Map-Based Localization Robot Localization
Electrical Engineering, Inha University, Incheon, South Korea
DRT&R: Direct Radar Teach & Repeat
Sep 15, 2026Alexander Krawciw, Daniil Lisus, Cedric Le Gentil +1Off-Road Robot Navigation Embodied Navigation
Robotics Institute, University of Toronto, Canada · Mobile Robotics Lab, ETH Zürich, Switzerland

Mar 6, 2026J/K move · Enter open · S save

CFEAR-Teach-and-Repeat: Fast and Accurate Radar-only Localization

Maximilian Hilger, Daniel Adolfsson, Ralf Becker, Henrik Andreasson+1

Chair of Perception for Intelligent Systems, Munich Institute of Robotics and Machine Intelligence, Technical University of Munich, Munich, Germany · Robot Navigation and Perception Lab of the AASS Research Center, Örebro University, Örebro, Sweden · Bosch Rexroth, Germany

Robot Localization Map-Based Localization

Open paper AskPDF arXiv

Reliable localization in prior maps is essential for autonomous navigation, particularly under adverse weather, where optical sensors may fail. We present CFEAR-TR, a teach-and-repeat localization pipeline using a single spinning radar, which is designed for easily deployable, lightweight, and robust navigation in adverse conditions. Our method localizes by jointly aligning live scans to both stored scans from the teach mapping pass, and to a sliding window of recent live keyframes. This ensures accurate and robust pose estimation across different seasons and weather phenomena. Radar scans are represented using a sparse set of oriented surface points, computed from Doppler-compensated measurements. The map is stored in a pose graph that is traversed during localization. Experiments on the held-out test sequences from the Boreas dataset show that CFEAR-TR can localize with an accuracy as low as 0.117 m and 0.096°, corresponding to improvements of up to 63% over the previous state of the art, while running efficiently at 29 Hz. These results substantially narrow the gap to lidar-level localization, particularly in heading estimation. We make the C++ implementation of our work available to the community.

Abstract

Map-Based Localization

Explore similar work

CFEAR-Teach-and-Repeat: Fast and Accurate Radar-only Localization

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

DRT&R: Direct Radar Teach & Repeat