4DRaL: Bridging 4D Radar with LiDAR for Place Recognition using Knowledge Distillation

Ningyuan Huang; Zhiheng Li; Zheng Fang

cs.CVMar 27, 2026

4DRaL: Bridging 4D Radar with LiDAR for Place Recognition using Knowledge Distillation

Authors: Ningyuan Huang, Zhiheng Li, Zheng Fang

Organizations: The authors are all with the Faculty of Robot Science and Engineering, Northeastern University, Shenyang, China.

Abstract

Place recognition is crucial for loop closure detection and global localization in robotics. Although mainstream algorithms typically rely on cameras and LiDAR, these sensors are susceptible to adverse weather conditions. Fortunately, the recently developed 4D millimeter-wave radar (4D radar) offers a promising solution for all-weather place recognition. However, the inherent noise and sparsity in 4D radar data significantly limit its performance. Thus, in this paper, we propose a novel framework called 4DRaL that leverages knowledge distillation (KD) to enhance the place recognition performance of 4D radar. Its core is to adopt a high-performance LiDAR-to-LiDAR (L2L) place recognition model as a teacher to guide the training of a 4D radar-to-4D radar (R2R) place recognition model. 4DRaL comprises three key KD modules: a local image enhancement module to handle the sparsity of raw 4D radar points, a feature distribution distillation module that ensures the student model generates more discriminative features, and a response distillation module to maintain consistency in feature space between the teacher and student models. More importantly, 4DRaL can also be trained for 4D radar-to-LiDAR (R2L) place recognition through different module configurations. Experimental results prove that 4DRaL achieves state-of-the-art performance in both R2R and R2L tasks regardless of normal or adverse weather.

Figures & tables

Fig. 1: (a) Comparison of runtime and performance across several methods on SJTU4D dataset [ 8 ] . (b) The framework of 4D radar place recognition based on knowledge distillation. The query scan is encoded into a descriptor, and the database is searched for the closest matching descriptors. The R2R (radar-to-radar) and L2L (LiDAR-to-LiDAR) databases consist of 4D radar descriptors and LiDAR descriptors, respectively.

Fig. 2: Comparison of the two paradigms of R2L (radar-to-LiDAR). The joint training paradigm employs one or more encoders to directly align the features of LiDAR and 4D radar, while the knowledge distillation paradigm leverages a teacher model to guide the feature alignment of two modalities.

Fig. 3: Overview of 4DRaL. Given LiDAR and 4D radar point clouds of the same scene, we generate BEV images. The teacher model extracts features from LiDAR BEV to guide the student model, which consists of R2R and R2L. R2R applies LIE for image enhancement, FDD for feature learning, and RD for refining output relationships. R2L enhances radar data with LIE and aligns LiDAR features via RD. The teacher model is used only during training.

Dataset	Splits	Frame Num.	Dataset	Splits	Frame Num.
NTU4D	train_query	7620	SJTU4D	test_a_query	7634
	train_database	10000		test_a_database	7500
	validation_query	7002		test_b_query	6501
	validation_database	10839		test_b_database	2500

TABLE I: The details of the benchmark datasets.

Methods	Modality	NTU4D			SJTU4D-TestA			SJTU4D-TestB			Runtime	Param
Methods	Modality	R@1	R@5	R@10	R@1	R@5	R@10	R@1	R@5	R@10	ms	$\cdot 10^{6}$
4DRaL (Teacher)	L2L	97.2	97.5	98.0	93.8	94.4	94.5	89.0	90.2	92.4	1.2	2.78
In. Scan Context [ 27 ]	R2R	84.1	86.2	90.2	67.9	79.8	83.7	78.0	86.5	90.2	0.2	$10^{-5}$
TransLoc3D [ 28 ]	R2R	87.0	90.2	92.6	75.6	89.2	92.2	79.3	86.0	87.5	11.2	6.63
MinkLoc3Dv2 [ 29 ]	R2R	91.5	93.8	95.2	88.6	93.2	94.0	85.8	87.9	88.9	8.8	2.66
PTC-Net-L [ 3 ]	R2R	95.1	95.9	96.5	79.8	89.6	91.4	80.3	85.3	86.8	4.2	0.08

TABLE II: Comparison with state-of-the-art of R2R in benchmark datasets.

Methods	Modality	NTU4D			SJTU4D-TestA			SJTU4D-TestB			Runtime	Param
Methods	Modality	R@1	R@5	R@10	R@1	R@5	R@10	R@1	R@5	R@10	ms	$\cdot 10^{6}$
4DRaL (Teacher)	L2L	97.2	97.5	98.0	93.8	94.4	94.5	89.0	90.2	92.4	1.2	2.78
Radar-to-LiDAR [ 12 ]	R2L	14.3	21.0	27.4	4.0	7.1	13.0	19.8	36.3	39.6	2.0	4.28
RaLF [ 13 ]	R2L	22.4	27.5	34.3	27.4	40.8	46.8	49.6	64.8	68.1	2.6	2.97
4DRaL (w/o KD)	R2L	16.5	22.5	28.9	13.3	27.2	36.3	32.3	46.1	53.5	1.2	2.78
4DRaL (w/ KD)	R2L	34.8	43.5	48.7	38.2	49.8	58.1	75.5	80.3	83.2	2.5	3.46

TABLE III: Comparison with state-of-the-art of R2L in benchmark datasets.

Methods	Modality	$\mathcal{L}_{E}$	$\mathcal{L}_{FDD}$	$\mathcal{L}_{RD}$	NTU4D
Methods	Modality	$\mathcal{L}_{E}$	$\mathcal{L}_{FDD}$	$\mathcal{L}_{RD}$	R@1	R@5	R@10
a0	R2R				94.5	95.1	95.5
a1	R2R	$\checkmark$			95.2	95.6	96.0
a2	R2R	$\checkmark$	$\checkmark$		95.7	96.1	96.3
a3	R2R	$\checkmark$	$\checkmark$	$\checkmark$	96.1	96.3	96.5
b0	R2L				16.5	22.5	28.9

TABLE IV: Ablation studies to evaluate the KD module composition.

Methods	Modality	NTU4D
	Modality	R@1	R@5	R@10
Raw	R2R	94.5	95.1	95.5
Global	R2R	94.9	95.6	96.2
Local	R2R	96.1	96.3	96.5
Raw	R2L	16.5	22.5	28.9
Global	R2L	25.7	30.9	38.2

TABLE V: Ablation studies on Local Image Enhancement.

Method		NTU4D
	R@1	R@5	R@10
Raw		94.5	95.1	95.5
Single	KL	95.3	95.9	96.2
Single	MSE	95.0	95.5	96.2
Dual	KL	96.1	96.3	96.5
Dual	MSE	95.8	96.1	96.4

TABLE VI: Ablation studies on Feature Distribution Distillation.

Fig. 4: The influence of margin on Response Distillation (RD) module.

Fig. 5: Visualization of Local Image Enhancement. In the area highlighted by the blue ellipse, the enhanced 4D Radar BEV image presents a clearer outline compared to the original image.

Fig. 6: Visualization of LiDAR and 4D radar data across different weather conditions in the SNAIL radar dataset.

Methods	Mod.	SS-night			SS-moderate-rain			SS-heavy-rain
Methods	Mod.	R@1	R@5	R@10	R@1	R@5	R@10	R@1	R@5	R@10
RaLF	R2R	82.9	90.5	93.8	84.8	91.5	95.7	81.2	89.6	94.3
TransLoc4D	R2R	84.5	88.1	91.5	81.9	91.1	93.3	79.8	84.3	90.4
4DRaL (Ours)	R2R	88.4	93.3	95.9	90.1	95.7	97.4	88.1	93.5	96.0
Radar-to-LiDAR	R2L	17.8	30.8	38.1	16.2	28.4	33.4	20.8	34.8	45.6
RaLF	R2L	25.2	36.7	42.8	28.9	42.8	57.2	30.5	44.8	60.4

TABLE VII: Comparison of R2R and R2L performance on the SNAIL.

Sensor	SS-night			SS-moderate-rain			SS-heavy-rain
	R@1	R@5	R@10	R@1	R@5	R@10	R@1	R@5	R@10
LiDAR	94.4	96.9	97.1	93.7	94.5	96.8	87.7	90.3	94.2
4D Radar	88.4	93.3	95.9	90.1	95.7	97.4	88.1	93.5	96.0

TABLE VIII: Comparison of LiDAR-to-LiDAR and 4D radar-to-4D radar place recognition performance under adverse weather.

Millimeter-Wave Radar

Millimeter-wave radar uses reflected high-frequency radio waves to sense motion, distance, and scene structure in settings where cameras may fail.

See all 51 papers

RLPR: Radar-to-LiDAR Place Recognition via Two-Stage Asymmetric Cross-Modal Alignment for Autonomous Driving
Mar 9, 2026Zhangshuo Qi, Jingyi Xu, Luqi Cheng +2Autonomous Driving Cross-Modal Representation Learning
Beijing Institute of Technology, Beijing, 100081, China. · Shanghai Jiaotong University, Shanghai, 200240, China.
Spatially Stratified Distillation for Heterogeneous Radar Place Recognition
Jun 17, 2026Sagun Singh Shrestha, Samuel Harding, Abdelwahed Khamis +2Cross-Modal Knowledge Distillation Knowledge Distillation
CSIRO Robotics, Brisbane, Australia · University of Queensland, Brisbane, Australia
Registering the 4D Millimeter Wave Radar Point Clouds Via Generalized Method of Moments
Aug 4, 2025Xingyi Li, Han Zhang, Ziliang Wang +2Millimeter-Wave Radar mmWave Radar Sensing
School of Automation and Intelligent Sensing, Shanghai Jiao Tong University, Shanghai, China · Institute of Medical Robotics, Shanghai Jiao Tong University, Shanghai, China · Department of Statistics, Uppsala University, Sweden

Mar 9, 2026J/K move · Enter open · S save

RLPR: Radar-to-LiDAR Place Recognition via Two-Stage Asymmetric Cross-Modal Alignment for Autonomous Driving

Zhangshuo Qi, Jingyi Xu, Luqi Cheng, Shichen Wen+1

Beijing Institute of Technology, Beijing, 100081, China. · Shanghai Jiaotong University, Shanghai, 200240, China.

Autonomous Driving Cross-Modal Representation Learning LiDAR Place Recognition

Open paper AskPDF arXiv

All-weather autonomy is critical for autonomous driving, which necessitates reliable localization across diverse scenarios. While LiDAR place recognition is widely deployed for this task, its performance degrades in adverse weather. Conversely, radar-based methods, though weather-resilient, are hindered by the general unavailability of radar maps. To bridge this gap, radar-to-LiDAR place recognition, which localizes radar scans within existing LiDAR maps, has garnered increasing interest. However, extracting discriminative and generalizable features shared between modalities remains challenging, compounded by the scarcity of large-scale paired training data and the signal heterogeneity across radar types. In this work, we propose RLPR, a robust radar-to-LiDAR place recognition framework compatible with single-chip, scanning, and 4D radars. We first design a dual-stream network to extract structural features that abstract away from sensor-specific signal properties (e.g., Doppler or RCS). Subsequently, motivated by our task-specific asymmetry observation between radar and LiDAR, we introduce a two-stage asymmetric cross-modal alignment (TACMA) strategy, which leverages the pre-trained radar branch as a discriminative anchor to guide the alignment process. Experiments on four datasets demonstrate that RLPR achieves state-of-the-art recognition accuracy with strong zero-shot generalization capabilities.

Abstract

Millimeter-Wave Radar

Explore similar work

RLPR: Radar-to-LiDAR Place Recognition via Two-Stage Asymmetric Cross-Modal Alignment for Autonomous Driving

Spatially Stratified Distillation for Heterogeneous Radar Place Recognition

Registering the 4D Millimeter Wave Radar Point Clouds Via Generalized Method of Moments