RAGNAROK: Radar-Aided Gravity-Normalized Alignment for Robust Open Keyframe-based Radar-Visual-Kinematic-Inertial SLAM

Hanjun Kim; Chiyun Noh; Sangwoo Jung; Jaehyung Jung; Simon Boche; Cedric Le Gentil; Stefan Leutenegger; Ayoung Kim

cs.ROOct 8, 2026

RAGNAROK: Radar-Aided Gravity-Normalized Alignment for Robust Open Keyframe-based Radar-Visual-Kinematic-Inertial SLAM

Authors: Hanjun Kim, Chiyun Noh, Sangwoo Jung, Jaehyung Jung, Simon Boche, Cedric Le Gentil, Stefan Leutenegger, Ayoung Kim

Organizations: Dept. of Mechanical Engineering, SNU, Seoul, S. Korea · Mobile Robotics Lab, School of Computation, Information and Technology (CIT) at the Technical University of Munich (TUM), 80333 Munich, Germany, and with the Munich Institute of Robotics and Machine Intelligence (MIRMI) · Mobile Robotics Lab, ETH Zurich, 80892 Zurich, Switzerland

Abstract

Legged robots offer superior mobility in unstructured environments, but reliable operation in such conditions requires robust state estimation. To address the vulnerability of proprioceptive estimators in rough terrain, recent methods have incorporated radar to provide velocity measurements. However, their limited yaw observability still leads to drift, and failure-aware fusion for adverse environments remains underexplored. In this letter, we present RAGNAROK, the first radar-visual-kinematic-inertial SLAM designed for robust operation in challenging environments. It integrates slip- and rolling-contact-aware leg velocity estimation, a kinematics-aware radar factor, and degradation-aware image enhancement. We further incorporate a B-spline-based radar-aided proprioceptive backbone, adaptive weighting, and online extrinsic calibration. Extensive experiments on public and self-collected datasets demonstrate that RAGNAROK achieves robust performance under challenging conditions and outperforms state-of-the-art baselines. The source code and dataset are available at https://github.com/hanjun815/RAGNAROK.

Figures & tables

Fig. 1 : Top: Performance comparison with SOTA baselines, with scores computed as $\mathrm{ATE}_{\mathrm{mean}}/(\mathrm{ATE}_{\mathrm{mean}}+\mathrm{ATE})$ . Middle: Representative challenging scenarios from the self-collected dataset. Bottom: Overview of the proposed RAGNAROK pipeline.

Fig. 2 : Overview of RAGNAROK: The radar-aided proprioceptive estimator optimizes the splines of velocity, rotation, and local gravity. The exteroceptive estimator performs image enhancement, keypoint matching, landmark triangulation, place recognition, and submapping.

Fig. 3 : Kinematics-aware multi-radar factor. (a) For the horizontal radar, wall points aligned with the azimuth are weighted higher. (b) For the vertical radar, floor and ceiling points aligned with the azimuth are weighted higher. (c) Under harsh motion, increased positional uncertainty lowers the radar Doppler residual weights.

Fig. 4 : Left : Plot of $\pi(H_{i})M_{i}(H_{i})$ . Middle : Before enhancement, features are non-uniformly distributed. Right : After enhancement, more features are extracted and more uniformly distributed.

Fig. 5 : Qualitative comparison of image enhancement results.

Dataset	Platform	Radar	Camera	LiDAR	IMU	GT	Scenarios
RRxIO [ 4 ]	UAV	1 $\times$ SoC	1 $\times$ Mono	-	✓	Vicon/ VISLAM	Indoor, Outdoor, Dark
HeRCULES [ 13 ]	Car	1 $\times$ 4D, 1 $\times$ Spinning	2 $\times$ Mono	✓	✓	RTK	Outdoor, Dark, Rain, Snow
Co-RaL [ 11 ]	Legged Robot	1 $\times$ SoC	-	✓	✓	LIO	Indoor, Outdoor, Hybrid
GaRLILEO [ 20 ]	Legged Robot	1 $\times$ SoC	-	✓	✓	TLS	Indoor, Outdoor, Hybrid
GrandTour [ 23 ]	Legged Robot	-	7 $\times$ Stereo, 8 $\times$ Mono	✓	✓	RTS/ GNSS	Indoor, Outdoor, Dim, Dark, Snow
RAGNAROK	Legged Robot	2 $\times$ SoC	1 $\times$ Stereo, 2 $\times$ Mono	✓	✓	TLS	Indoor, Outdoor, Hybrid, Dark, Dim, Glare

TABLE I : Comparison with Existing Datasets

Sequence	Outdoor	Indoor	Length (m)	Elevation	Self-similarity	# of Loops
Terrace	Dark	-	68.38	-	-	1
TerraceLoop	Dark	-	134.47	-	-	2
Garden	Dark	-	162.54	✓	-	2
Quad	Dim	-	438.28	✓	-	1
Overpass	Glare	-	160.10	✓	-	1
Mountain-Long	✓	-	207.59	✓	-	1

TABLE II : Overview of Sequences

Fig. 6 : System overview of the platform and coordinates of sensors.

Modality				Methods	Terrace		TerraceLoop		Garden		Quad		Overpass		Mountain-Long		Mountain-Short		Average
R	V	K	I	Methods	APE t	APE r	APE t	APE r	APE t	APE r	APE t	APE r	APE t	APE r	APE t	APE r	APE t	APE r	APE t	APE r
✓			✓	x-RIO [ 5 ]	1.655	7.384	2.829	8.555	4.316	18.500	16.595	21.122	3.222	8.068	10.096	10.139	9.691	9.421	6.915	11.884
	✓		✓	OpenVINS [ 7 ]	-	-	-	-	-	-	8.275	2.970	0.591	2.045	0.406	0.754	0.564	0.889	-	-
	✓		✓	OKVIS2-X [ 2 ]	0.474	2.596	0.568	3.073	6.626	10.046	-	-	1.166	3.447	0.311	1.196	0.465	1.014	-	-
		✓	✓	DRIFT [ 17 ]	1.767	7.466	2.942	6.116	4.676	18.563	13.150	8.879	4.691	3.569	7.661	4.805	6.473	5.964	5.909	7.909
		✓	✓	Holistic [ 21 ]	2.077	13.082	6.152	15.696	10.225	32.851	27.420	25.328	6.937	17.755	2.101	8.072	13.689	21.850	9.800	19.233

TABLE III : Evaluation Results on Outdoor Sequences (R: Radar, V: Visual, K: Kinematic, I: Inertial, 1st , 2nd , 3rd )

Fig. 7 : Reconstructed map of Mountain-Long with trajectories.

Fig. 8 : Trajectory comparison with SOTA methods on Overpass .

Fig. 9 : Trajectory comparison with SOTA methods on Downstair .

Methods	Building	Garage	Stair	Trail	Average
Co-RaL [ 11 ]	4.922	1.494	1.562	2.586	2.641
GaRLILEO [ 20 ]	4.017	1.367	3.040	2.888	2.828
RAGNAROK-RKI (w/o R.C.)	3.565	0.890	0.856	0.993	1.576
RAGNAROK-RKI (w/o K.A.C.)	3.610	0.879	0.856	0.951	1.574
RAGNAROK-RKI	3.484	0.890	0.826	0.962	1.541

TABLE IV : Evaluation Results on Co-RaL Dataset ( APE t )

Methods	BridgeLoop	SlopeStair	MoCap-E	MoCap-H	Average
Co-RaL [ 11 ]	2.182	4.696	-	1.549	-
GaRLILEO [ 20 ]	0.345	1.132	0.356	0.467	0.575
RAGNAROK-RKI (w/o R.C.)	0.408	0.946	0.346	0.365	0.516
RAGNAROK-RKI (w/o K.A.C.)	0.326	1.089	0.346	0.365	0.532
RAGNAROK-RKI	0.338	0.905	0.345	0.365	0.488

TABLE V : Evaluation Results on GaRLILEO Dataset ( APE t )

S	I	G	M	A	D	O	TerraceLoop		Overpass		Atrium		Average
S	I	G	M	A	D	O	APE t	APE r	APE t	APE r	APE t	APE r	APE t	APE r
$\times$	✓	✓	✓	✓	✓	✓	0.223	1.619	0.538	2.317	0.084	1.656	0.282	1.864
✓	$\times$	✓	✓	✓	✓	✓	0.753	1.952	0.806	3.557	0.466	2.314	0.675	2.608
✓	✓	$\times$	✓	✓	✓	✓	0.254	1.598	0.566	2.513	0.087	1.582	0.302	1.898
✓	✓	✓	$\times$	✓	✓	✓	0.223	1.588	0.679	2.729	0.085	1.623	0.329	1.980
✓	✓	✓	✓	$\times$	✓	✓	0.785	3.549	1.657	5.142	4.217	16.370	2.220	8.354

TABLE VI : Ablation Study on RAGNAROK Components

Fig. 10 : Precision–recall curves for sequences with multiple loops.

Fig. 11 : Average runtime of RAGNAROK-SLAM in ms per frame.

Modality				Methods	Upstair		Downstair		FloorGyre		CorriLoop		Atrium		SlopeStair		Tunnel		Average
R	V	K	I	Methods	APE t	APE r	APE t	APE r	APE t	APE r	APE t	APE r	APE t	APE r	APE t	APE r	APE t	APE r	APE t	APE r
✓			✓	x-RIO [ 5 ]	2.020	7.873	1.755	6.715	1.123	6.587	2.535	7.361	1.562	6.934	1.497	7.277	8.181	29.785	2.668	10.362
	✓		✓	OpenVINS [ 7 ]	0.351	1.433	0.701	1.873	-	-	0.631	1.293	-	-	0.410	1.352	3.611	30.145	-	-
	✓		✓	OKVIS2-X [ 2 ]	1.008	2.957	0.539	1.661	0.872	3.536	0.427	1.500	-	-	0.261	1.252	2.550	5.260	-	-
		✓	✓	DRIFT [ 17 ]	2.139	6.010	3.870	10.425	2.016	8.602	5.656	9.392	2.204	8.220	3.099	6.396	6.454	42.345	3.634	13.056
		✓	✓	Holistic [ 21 ]	2.443	13.713	10.834	27.907	5.110	27.343	10.834	27.907	3.327	25.615	6.170	25.278	14.001	23.840	7.531	24.515

TABLE VII : Evaluation Results on Indoor and Hybrid Sequences (R: Radar, V: Visual, K: Kinematic, I: Inertial, 1st , 2nd , 3rd )

Explore similar work

GaRLILEO: Gravity-aligned Radar-Leg-Inertial Enhanced Odometry
Nov 17, 2025Chiyun Noh, Sangwoo Jung, Hanjun Kim +3Legged Robotics Inertial Odometry
Dept. of Mechanical Engineering, SNU, Seoul, S. Korea · Robotics and AI Institute, Cambridge, USA
DiffRadar: Differentiable Physics-Aware Radar SLAM with Gaussian Fields
Jul 14, 2026Gaurav Bagwe, Xiaoyong Yuan, Yongji Wu +1Simultaneous Localization and Mapping Millimeter-Wave Radar
Clemson University Clemson, South Carolina, USA
Dr-PoGO: Direct Radar Pose-Graph Optimization
May 6, 2026Cedric Le Gentil, Weican Li, Leonardo Brizi +1Simultaneous Localization and Mapping Relative Pose Estimation
Robotics Institute University of Toronto. · Department of Computer, Control, and Management Engineering “Antonio Ruberti”, Sapienza University of Rome.

Nov 17, 2025, cs.ROJ/K move · Enter open · S save

GaRLILEO: Gravity-aligned Radar-Leg-Inertial Enhanced Odometry

Chiyun Noh, Sangwoo Jung, Hanjun Kim, Yafei Hu+2

Dept. of Mechanical Engineering, SNU, Seoul, S. Korea · Robotics and AI Institute, Cambridge, USA

Legged Robotics Inertial Odometry Radar-Inertial Odometry Robot State Estimation

Open paper AskPDF arXiv

Deployment of legged robots for navigating challenging terrains (e.g., stairs, slopes, and unstructured environments) has gained increasing preference over wheel-based platforms. In such scenarios, accurate odometry estimation is a preliminary requirement for stable locomotion, localization, and mapping. Traditional proprioceptive approaches, which rely on leg kinematics sensor modalities and inertial sensing, suffer from irrepressible vertical drift caused by frequent contact impacts, foot slippage, and vibrations, particularly affected by inaccurate roll and pitch estimation. Existing methods incorporate exteroceptive sensors such as LiDAR or cameras. Further enhancement has been introduced by leveraging gravity vector estimation to add additional observations on roll and pitch, thereby increasing the accuracy of vertical pose estimation. However, these approaches tend to degrade in feature-sparse or repetitive scenes and are prone to errors from double-integrated IMU acceleration. To address these challenges, we propose GaRLILEO, a novel gravity-aligned continuous-time radar-leg-inertial odometry framework. GaRLILEO decouples velocity from the IMU by building a continuous-time ego-velocity spline from SoC radar Doppler and leg kinematics information, enabling seamless sensor fusion which mitigates odometry distortion. In addition, GaRLILEO can reliably capture accurate gravity vectors leveraging a novel soft S2-constrained gravity factor, improving vertical pose accuracy without relying on LiDAR or cameras. Evaluated on a self-collected real-world dataset with diverse indoor-outdoor trajectories, GaRLILEO demonstrates state-of-the-art accuracy, particularly in vertical odometry estimation on stairs and slopes. We open-source both our dataset and algorithm to foster further research in legged robot odometry and SLAM. https://garlileo.github.io/GaRLILEO

Abstract

Explore similar work

GaRLILEO: Gravity-aligned Radar-Leg-Inertial Enhanced Odometry

DiffRadar: Differentiable Physics-Aware Radar SLAM with Gaussian Fields

Dr-PoGO: Direct Radar Pose-Graph Optimization