Mag4D-SLAM Dataset

Abstract

Simultaneous Localization and Mapping (SLAM) is a fundamental capability for autonomous robotic systems operating in complex and large-scale environments. The geomagnetic field—despite being globally available, infrastructure-free, and inherently absolute in orientation—remains largely absent from robotics-oriented SLAM benchmarks. To address this gap, we introduce the Mag4D Dataset, a multi-modal geomagnetic benchmark specifically designed for magnetic-aware localization and mapping research. The dataset provides time-synchronized measurements from LiDAR, camera, IMU, tri-axis magnetometer, and GPS sensors, together with LiDAR–IMU-based ground truth alignment. Unlike prior magnetic datasets that focus primarily on indoor anomaly localization, Mag4D emphasizes structured geometric trajectories including circular and rectangular paths recorded across seven repeated traversals of identical routes—enabling quantitative evaluation of global yaw drift, cross-session heading consistency, magnetic map stability, and alignment error under identical spatial constraints. The dataset includes both daytime and nighttime recordings on the campus, providing a controlled yet realistic outdoor benchmark for multi-modal SLAM evaluation.

Key Contributions

🧲

4D Geomagnetic Sensing

Continuous tri-axis magnetic field measurements (M_x, M_y, M_z) captured as a function of time along spatial trajectories, enabling joint spatial–temporal magnetic field analysis.

🔁

Repeated-Traversal Design

Seven sequences recorded over identical paths enable systematic investigation of magnetic repeatability, heading-dependent variation at identical spatial locations, and the quantitative impact of geomagnetic cues on yaw drift mitigation.

🌗

Day & Night Conditions

Paired daytime and nighttime sessions allow controlled analysis of heading robustness under illumination transitions—a scenario absent from existing magnetic-aware benchmarks.

📍

High-Accuracy Ground Truth

Precise 6-DoF ground truth poses generated by registering LiDAR scans against a high-accuracy reference map via ICP, with static validation showing standard deviation below 3 cm.

🏗️

Structured Geometric Trajectories

Circular and rectangular paths with mild elevation changes around the campus provide structured outdoor benchmarking suited for global heading and loop-closure evaluation.

📡

Full Multi-Modal Synchronization

Hardware-synchronized LiDAR, camera, IMU, magnetometer, and GPS streams timestamped to a common system clock, with temporal offsets calibrated by minimizing angular velocity residuals.

Sensor Suite

Sensor	Model	Rate	Specifications
📷 Camera	Gemini 336L	30 Hz	Depth 1280×800, H90°/V65°; RGB 1280×800, H94°/V68°
🔵 LiDAR	Mid-360	10 Hz	360°/59° FOV; 200k pts/s; 70 m range
📐 IMU	Mid-360 (built-in)	200 Hz	3-axis Gyroscope; 3-axis Accelerometer
🧲 Magnetometer	3DM-GX5-AHRS	100 Hz	3-axis; ±8 Gauss; Temperature compensated
🛰️ GPS	EVK-F9P	1 Hz	Multi-band GNSS receiver; Operated in standalone mode (~meter-level accuracy)

Sensor Extrinsics

All transformations are defined with LiDAR as the base frame · T_{lidar←sensor}

Sensor extrinsic coordinate frames animated 360° view

RGB axes = XYZ · Dashed lines indicate sensor offsets from LiDAR origin

# ==============================================================
# Sensor Extrinsic Parameters (Base Frame: LiDAR)
# ==============================================================

# 1. Camera → LiDAR
T_lidar_camera: [
   0.0062967,  0.0002136,  0.9999802,  0.0720748,
  -0.9999516,  0.0075562,  0.0062949,  0.0193199,
  -0.0075547, -0.9999714,  0.0002611, -0.0151779,
   0.0,        0.0,        0.0,        1.0
]

# 2. Magnetometer → LiDAR
T_lidar_mag: [
  -1.0,  0.0,  0.0,  0.115,
   0.0,  1.0,  0.0,  0.0,
   0.0,  0.0, -1.0, -0.13,
   0.0,  0.0,  0.0,  1.0
]

# 3. IMU → LiDAR (Livox Mid-360 Factory Spec)
T_lidar_imu: [
   1.0,  0.0,  0.0,  0.011,
   0.0,  1.0,  0.0,  0.02329,
   0.0,  0.0,  1.0, -0.04412,
   0.0,  0.0,  0.0,  1.0
]

📷 Camera

7.6 cm

🧲 Magnetometer

17.4 cm

📐 IMU

5.1 cm

Sequences

14 sequences · 3 environments · day & night · forward & reverse traversals

Sequence	Env.	Light	Direction	Path (m)	Duration (s)
Campus
campus_day_fwd_01	Campus	☀️ Day	→ Fwd	1382.9	1203.8
campus_day_fwd_04	Campus	☀️ Day	→ Fwd	1554.1	1291.9
campus_day_fwd_08	Campus	☀️ Day	→ Fwd	1465.6	1181.2
campus_day_fwd_09	Campus	☀️ Day	→ Fwd	1443.6	1132.3
campus_day_rev_05	Campus	☀️ Day	← Rev	1637.2	1299.1
campus_day_rev_07	Campus	☀️ Day	← Rev	1605.3	1234.9
campus_night_fwd_00	Campus	🌙 Night	→ Fwd	1376.6	1117.3
campus_night_rev_11	Campus	🌙 Night	← Rev	1527.3	1224.4
Circle
circle_day_fwd_rev_02	Circle	☀️ Day	↔ Fwd+Rev	1119.1	822.1
circle_day_fwd_rev_06	Circle	☀️ Day	↔ Fwd+Rev	987.5	763.0
circle_day_fwd_rev_10	Circle	☀️ Day	↔ Fwd+Rev	1402.8	993.2
circle_night_fwd_rev_00	Circle	🌙 Night	↔ Fwd+Rev	1064.4	816.9
Library
library_day_fwd_03	Library	☀️ Day	→ Fwd	189.5	163.3
Campus + Circle
campus_circle_night_rev_12	Campus+Circle	🌙 Night	← Rev	1676.0	1406.2

Comparison with Existing Datasets

Year	Dataset	Seq.	Avg. Length	Mag.	Day/Night	Precise GT
2013	KITTI	22	~2 km	✗	✓	✓
2016	EuRoC	11	~80 m	✗	✗	✓
2017	RobotCar	100+	~10 km	✗	✓	✓
2017	MagPIE	10	~500 m	✓	✗	✓
2018	TUM VI	28	~200 m	✗	✗	✓
2020	Newer College	4	~1.5 km	✗	✓	✓
2021	MagWi	100+	Indoor multi-floor	✓	✓	✗
2023	Boreas	40+	~8 km	✗	✓	✓
2025	GrandTour	49	>10 km	✗	✓	✓
2026	Mag4D (Ours)	7	~1.3 km	✓	✓	✓

Data Format

Directory Structure

Mag4D/
├── campus_day_fwd_01/
│   ├── raw.bag      # All ROS topics
│   ├── imu.csv      # IMU data
│   ├── mag.csv      # Magnetic field
│   ├── gps.csv      # GPS positions
│   ├── gt_pose.csv  # Ground truth poses
│   ├── map.pcd      # Global LiDAR map
│   ├── lidar/       # Point clouds
│   └── images/      # Camera frames
├── campus_day_fwd_04/
│   └── ...
├── campus_night_fwd_00/
│   └── ...
├── circle_day_fwd_rev_02/
│   └── ...
├── library_day_fwd_03/
│   └── ...
└── ...              # 14 sequences total

ROS Topics

/imu/data Linear acceleration and angular velocity

/imu/mag Tri-axis magnetic field (M_x, M_y, M_z) in µT

/lidar/points 3D LiDAR point clouds (PCD format)

/camera/image_raw Monocular RGB images

/gps/fix GPS position measurements

/tf Coordinate frame transformations

Citation

@inproceedings{nayak2026mag4d,
  author    = {Nayak, Bibhutibhusan and Ju, Hyoseok and Kim, Giseop},
  title     = {Mag4D-SLAM Dataset: A Repeated-Traversal Multi-Modal 4D Geomagnetic Dataset for Localization and Mapping},
  booktitle = {Proceedings of the IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)},
  year      = {2026},
  note      = {to appear, arXiv preprint arXiv:2607.21986}
}