[IROS2025]M3DGR: A Multi-sensor, Multi-scenario and Massive-baseline SLAM Dataset for Ground Robots

💎 Projcet Leader: Jie Yin 殷杰  🌐 [Website]  📝 [Paper]   ➡️ [Algorithm Code]   ⭐️ [Pre Video]   🔥 [News]

Our goal is to benchmark "all" cutting-edge SLAM!

News

• 🔥2026/06/14: We are excited to release a preview version of our strongest SLAM system Ultra-Fusion system,which unifies WIO, VIO, LIO, and LVIO in one configurable optimization framework, with optional wheel/GNSS fusion and online calibration.

1. Project Overview 🎯

This repository contains the official implementation of our IROS 2025 paper:

"Towards Robust Sensor-Fusion Ground SLAM: A Comprehensive Benchmark and a Resilient Framework"

In this work, we propose a complete solution for robust SLAM on ground robots operating under degraded conditions. Our key contributions are:

• Ground-Fusion++ (Link): A resilient and modular SLAM framework integrating heterogeneous sensors for robust localization and high-quality mapping.
• M3DGR Benchmark(Link): A challenging multi-sensor, multi-scenario SLAM benchmark dataset with systematiclly induced degradation.
• Comprehensive Evaluation(Link): A comprehensive evaluation of over 40 cutting-edge SLAM methods on M3DGR.

2. Latest Updates 📢

2.1 🔥 Latest works citing M3DGR(Updating!)

We are excited to see that M3DGR continues to inspire follow-up works. To facilitate tracking the latest developments, we provide a curated list of papers citing M3DGR:

• [RAL2026] UAF-LIVO: Uncertainty-Aware Feature-Based LiDAR-Inertial-Visual Odometry Robust to Cross-Modal Geometric Inconsistency, [Paper]
• [IROS2026] FAST-LIVGO: A Degeneracy-Robust LiDAR-Inertial-Visual-GNSS Fusion Odometry, [Paper]
• [TASE2026] Implicit Point-to-Voxel LiDAR-IMU SLAM, [Paper], [code]
• [RAL2026] R-VoxelMap: Accurate Voxel Mapping With Recursive Plane Fitting for Online LiDAR Odometry, [Paper], [code]
• [Arxiv2026] GenZ-LIO: Generalizable LiDAR-Inertial Odometry Beyond Indoor–Outdoor Boundaries, [Paper]
• [Arxiv2026] Surfel-LIO: Fast LiDAR-Inertial Odometry with Pre-computed Surfels and Hierarchical Z-order Voxel Hashing, [paper] [code]
• [TIM2026] Pizza-LIO: Intensity-Enhanced LiDAR-Inertial Odometry with Adaptive Scan Slicing, [paper]
• [RAL2025] LIVOX-CAM: Adaptive Coarse-to-Fine Visual-assisted LiDAR Odometry for Solid-State LiDAR, {paper], [code]

2.2 TODO

• Release camera-ready version of IROS2025 paper.[paper]
• Release 40 SLAM codes adapted for M3DGR dataset.[codes]
• Release Ground-Fusion++ code, with examples on M3DGR on M2DGR-plus. [code]
• Release most sequences in the paper included with GT and calibration files to make sure all results can be reproduced.[data]
• Release long-term sequences upon our journal paper acception.

🔍 For those interested in accessing the unreleased M3DGR sequences in advance, we recommend first thoroughly evaluating your methods on the already released sequences. After that, feel free to contact us at zhangjunjie587@gmail.com to request early access for research purposes.

3. Contribute to M3DGR

The M3DGR project is an open and collaborative effort. We encourage you to adapt and evaluate your SLAM or localization algorithms on top of the M3DGR dataset! Our goal is to build an open and dynamic community, where researchers can not only use the dataset, but also contribute back by:

• Integrating your algorithms as baseline methods, which can enable fair comparison and promote your algorithm.
• Sharing configuration files, evaluation results, and insights.

Let’s make M3DGR a growing hub for robust, reproducible SLAM research! You can

• Submit a Pull Request to contribute new algorithms, configuration files, or improvements via Github Pull Request to post your adapted codes [here]
• Report bugs or request features via GitHub Issues.
• Join discussions or ask questions on GitHub Discussions.

4. SENSOR SETUP

4.1 Acquisition Platform

Physical drawings and schematics of the ground robot. (a) Side view of the robot. (b) Sensor arrangement on the top layer. (c) Sensor arrangement on the middle and bottom layers. All dimensions are provided in centimeters.

Figure 1. The directions of the sensors are marked in different colors,red for X,green for Y and blue for Z.

4.2 Sensor parameters

All the sensors and track devices and their most important parameters are listed as below:

The rostopics of our rosbag sequences are listed as follows:

4.3 Calibration

Sensor extrinsics, camera intrinsics, and IMU noise parameters are provided in calibration.md. IMU noise parameters (gyr_n, gyr_w, acc_n, acc_w) follow the Kalibr convention and can be used directly by most LIO/LVIO systems (e.g., FAST-LIO2, LIO-SAM, R3LIVE).

5. DATASET SEQUENCES

Figure 2. All trajectories are mapped in different colors.

An overview of M3DGR is given in the table below:

5.1 Standard

Figure 3. Outdoor01 Sequences

Sequence Name	Collection Date	Total Size	Duration	Features	Rosbag	GT
Longtime01	2025-01-14	30.2g	1799s	Long Time	[OneDrive]/[Alipan]	OneDrive/Alipan
Longtime02	2025-01-18	36.3g	2118s	Long Time	[OneDrive]/[Alipan]	OneDrive/Alipan
Outdoor01	2025-01-03	6.10g	411s	Outdoor	OneDrive/Alipan	OneDrive/Alipan
Outdoor04	2025-01-03	13.4g	782s	Outdoor	OneDrive/Alipan	OneDrive/Alipan

5.2 Visual Challenge 📷

Figure 4. Light01 Sequences

Indoor:

Outdoor:

5.3 LiDAR Degeneration 🌐

Figure 5. corridor01 Sequences

Sequence Name	Collection Date	Total Size	Duration	Features	Rosbag	GT
Corridor01	2025-01-21	6.39g	403s	Long Corridor	OneDrive/Alipan	OneDrive/Alipan
Corridor02	2025-01-21	4.62g	293s	Long Corridor	OneDrive/Alipan	OneDrive/Alipan
Elevator01	2025-01-21	11.2g	699s	Long Corridor,Elevator	OneDrive/Alipan	OneDrive/Alipan

5.4 Wheel Slippage 🚗

Figure 6. Wheelfloat01 Sequences

Indoor:

Outdoor:

5.5 GNSS Denied 🛰️

Figure 7. GNSS_Denied01 Sequences

Sequence Name	Collection Date	Total Size	Duration	Features	Rosbag	GT
GNSS-denial01	2025-01-19	10.5g	609s	Long time,GNSS Denial	OneDrive/Alipan	OneDrive/Alipan
GNSS-denial02	2025-01-21	12.7g	750s	Long time,GNSS Denial	OneDrive/Alipan	OneDrive/Alipan

⚠️ Known Issues:

• The RGB images collected by the D435i and X4 cameras are rolling shutter, which might affect the performance of some visual SLAM systems which require global shutter.
• The dataset lacks external trigger between sensors, instead, we perform synchronization via software synchronization.

5.6 Download

We offer two download options for each M3DGR sequence: OneDrive (recommended) and Alipan(阿里云盘). If you use Alipan to download sequences as a folder, you need to double-click the 双击合并.bat file in the folder in Win10/Win11 to merge the original data automatically. Download all our public sequences at once through this link.

5.7 Data inspection

We provide the SHA256 value for users to check the integrity of rosbag. The file format is <SHA256><two-spaces><path\**.rosbag>.

You can run on Linux:

sha256sum path/to/your_download_bag.bag

or Windows:

(Get-FileHash -Algorithm SHA256 -Path '\path\to\your_download_bag.bag').Hash

Compare the hash output above with the hash of the corresponding line in checksums.txt. If they match exactly, the comparison is successful.

5.8 Convert ROS 1 Bags to ROS 2 Humble Format

M3DGR sequences are provided as ROS 1 bag files. If you want to use them in ROS 2 Humble, you can convert a ROS 1 bag into a ROS 2 bag directory with rosbags-convert.

Install the rosbags command line tools:

pip install rosbags

Make sure the source bag file is available in your current working directory. For example:

ls Occlusion01.bag

Run the following command to convert a ROS 1 Noetic bag into a ROS 2 Humble bag:

rosbags-convert \
  --dst-version 8 \
  --src-typestore ros1_noetic \
  --dst-typestore ros2_humble \
  --src Occlusion01.bag \
  --dst Occlusion01_humble

Single-line version:

rosbags-convert --dst-version 8 --src-typestore ros1_noetic --dst-typestore ros2_humble --src Occlusion01.bag --dst Occlusion01_humble

Parameter notes:

• --dst-version 8: Writes rosbag2 metadata in a format that can be read directly by ROS 2 Humble tools.
• --src-typestore ros1_noetic: Reads message definitions using the ROS 1 Noetic type store.
• --dst-typestore ros2_humble: Writes message definitions using the ROS 2 Humble type store.
• --src Occlusion01.bag: Specifies the input ROS 1 bag file.
• --dst Occlusion01_humble: Specifies the output ROS 2 bag directory.

After conversion, Occlusion01_humble should contain the ROS 2 bag metadata and storage files.

Note: With rosbags 0.11.3, the default destination version is 9. A version 9 bag may be created successfully, but ros2 bag info from ROS 2 Humble can fail when parsing metadata.yaml. Use --dst-version 8 for Humble compatibility.

If ROS 2 Humble is available in the environment, inspect the converted bag:

ros2 bag info Occlusion01_humble

You can also play the converted bag:

ros2 bag play Occlusion01_humble

Troubleshooting:

• If conversion fails because of missing message definitions, make sure all custom message packages used by the original ROS 1 bag are available to rosbags-convert.
• If ros2 bag play skips Livox topics such as /livox/mid360/lidar or /livox/avia/lidar, install and source the corresponding ROS 2 message packages, for example livox_ros_driver2 and livox_ros_driver.
• If the destination directory already exists, remove it or choose a new output directory before running the conversion again.

6. Evaluation

You can quickly get the trajectory in TUM format through the TF tree method like this.

• If the GT is obtained by RTK/Mocap, you can directly use evo to evaluate：

evo_ape tum GTDynamic01.txt Dynamic01.txt -ap

• If the GT is obtained by ArUco, you can use our script to evaluate：

pip install numpy colorama tabulate evo

python ArUco_evaluate.py <GT_file_path> <evaluation_dir_path> [options] 

<GT_file_path>: GT file path. <evaluation_dir_path>: folder path where the tum format file to be evaluated is stored. [options]: -t: sort the results by translation error, -r: sort by rotation error, -a: sort by root mean square error. Default sorting is based on translation error. For example：

python ArUco_evaluate.py GTCorridor01.txt ./M3DGR

Our paper uses a more stringent translation error. The Tracking Rate is calculated based on the recording time and the actual trajectory change time. If the algorithm crashes prematurely or the trajectory stops updating, the Tracking Rate will be less than 100%.

7. Supported SLAM Algorithm List🔥

We have tested following cutting-edge methods on M3DGR🦄 dataset with well-tuned parameters. We will release all these custom baseline codes upon paper acceptance!. The testing configuration is detailed below:

7.1 40 Evaluated SLAM Methods

⚠️ Known Issues:

• Please note that experimental performance may exhibit variability across runs and hardware platforms; the results reported in the paper represent averaged outcomes under our testing conditions.
• It is possible to further improve performance through careful parameter tuning and repeated evaluation in specific scenarios.

7.2 Open-source Contribution

8. Citation 📄

@article{zhang2025towards,
  title={Towards Robust Sensor-Fusion Ground SLAM: A Comprehensive Benchmark and A Resilient Framework},
  author={Zhang, Deteng and Zhang, Junjie and Sun, Yan and Li, Tao and Yin, Hao and Xie, Hongzhao and Yin, Jie},
  journal={arXiv preprint arXiv:2507.08364},
  year={2025}
}

@inproceedings{yin2024ground,
  title={Ground-fusion: A low-cost ground slam system robust to corner cases},
  author={Yin, Jie and Li, Ang and Xi, Wei and Yu, Wenxian and Zou, Danping},
  booktitle={2024 IEEE International Conference on Robotics and Automation (ICRA)},
  pages={8603--8609},
  year={2024},
  organization={IEEE}
}
@article{yin2021m2dgr,
  title={M2dgr: A multi-sensor and multi-scenario slam dataset for ground robots},
  author={Yin, Jie and Li, Ang and Li, Tao and Yu, Wenxian and Zou, Danping},
  journal={IEEE Robotics and Automation Letters},
  volume={7},
  number={2},
  pages={2266--2273},
  year={2021},
  publisher={IEEE}
}

9. Star History ⭐️

10. Contributing 👷‍♂️

We appreciate all contributions to improving M3DGR.

11. Acknowledgements 🤝

Thanks to everyone for supporting this project.

🔝Back to top