ucu-autonomous-ugv/ucu-slam-dataset-v1

UCU SLAM Dataset v1

Outdoor RGB-D + IMU + wheel odometry + GNSS dataset for evaluating SLAM systems, recorded with a Clearpath Husky A200 UGV on the Ukrainian Catholic University campus and in Stryiskyi Park (Lviv, Ukraine).

This is a gated dataset. To access the files, log in to Hugging Face and accept the conditions on this page. Your contact information will be shared with the dataset authors. Once access is granted, authenticate locally and download as usual:

hf auth login
hf download ucu-autonomous-ugv/ucu-slam-dataset-v1 --repo-type dataset --include "sequence-1/*" --local-dir ucu-slam-dataset-v1

Full documentation (PDF) · Calibration files (GitHub) · UCU Autonomous UGV Club

At a glance

Platform	Clearpath Husky A200, manually driven at approximately constant speed
Environments	UCU campus (structured) · Stryiskyi Park (natural) — outdoor only, never under buildings or arches
Sequences	13 evaluation sequences + 3 calibration sequences (A1–A3)
Variations	day / twilight · forward / reverse · flat / elevation · loop / figure-eight / linear
Formats	ROS 2 bag (MCAP) · TUM RGB-D–compatible
Ground truth	GNSS position in a local ENU frame (meter-level, no orientation) — see Limitations

Sensors

Sensor	Model	Output
RGB-D camera	Intel RealSense D435	RGB 640×480 @ 30 fps (FOV ≈ 69°×42°) · depth 640×480 @ 30 fps (FOV ≈ 75°×62°, range ≈ 0.3 m+), hardware-synchronized
IMU	Phidgets Spatial 1042	accelerometer ±8 g, gyroscope ±2000°/s @ 125 Hz
Wheel encoders	Husky A200, ×4	78,000 ticks/m
GNSS	Holybro H-RTK F9P	position fix, without RTK correction in v1

Calibration was performed with Kalibr on an AprilGrid (10×5 tags) and allan_variance_ros: camera intrinsics (A1, pinhole + radial-tangential), IMU noise model (A2), camera–IMU extrinsics and time offset (A3). Depth-to-color extrinsics come from the RealSense driver. All timestamps in the delivered data are already shifted to base time. Values are available in /tf_static, the camera_info topics and calibration.yaml.

Frames: base_link (body frame; ground truth refers to it) · camera_color, camera_depth (OpenCV convention) · imu · gnss · odom (wheel odometry) · map — ENU with origin at φ = 49.81754, λ = 24.02295, h = 378.3 m (WGS84), in the center of the campus.

Data format

ROS 2 bag (MCAP)

Topic	Message type	Content
/cam_color/image_raw, /cam_color/camera_info	sensor_msgs/Image, sensor_msgs/CameraInfo	RGB
/cam_depth/image_rect_raw, /cam_depth/camera_info	sensor_msgs/Image, sensor_msgs/CameraInfo	depth
/cam_depth_aligned/image_raw, /cam_depth_aligned/camera_info	sensor_msgs/Image, sensor_msgs/CameraInfo	depth aligned to RGB
/imu	sensor_msgs/Imu	IMU
/odom	nav_msgs/Odometry	wheel odometry
/gnss	sensor_msgs/NavSatFix	raw GNSS
/pose	geometry_msgs/PoseWithCovarianceStamped	ground truth of base_link in ENU
/tf, /tf_static	tf2_msgs/TFMessage	ground truth & odometry transforms · calibrated static transforms

TUM RGB-D–compatible: rgb.txt, depth.txt (aligned) / depth_not_aligned.txt, associations.txt / associations_not_aligned.txt, imu.txt, odom.txt, groundtruth.txt — works with the standard TUM RGB-D tools.

Repository layout

sequence-{1..13}/
├── sequence-N.mcap      # ROS 2 bag
└── metadata.yaml        # rosbag2 metadata
sequence-a{1..3}/        # calibration sequences
sequence-a{1..3}.bag
documentations/          # full PDF documentation
└── sequences/           # route previews and trajectory maps

Sequences

Challenge	Sequences
Small loop closure	1
Medium loop closure	2, 3, 6, 7, 8
Figure-eight / self-intersection	4, 5, 10
Reverse traversal	3, 7, 12
Elevation change (6-DoF)	9
Long-term drift (no loop closure)	11, 12, 13
Low-light (twilight)	5, 8, 13
Structured urban environment	1–5
Natural outdoor environment	6–13

Calibration — A1–A3 (UGV Lab, AprilGrid)

ID	Recording	Used for
A1	Slow camera motion in front of the AprilGrid	Camera intrinsics (uses Kalibr)
A2	Platform fully stationary for ~35 min	IMU noise densities and bias random walks (uses allan_variance_ros)
A3	Smooth motion exciting all IMU axes	Camera–IMU extrinsics and time offset (uses Kalibr)

Campus — 1–5

Buildings, sidewalks and corners — favorable for visual SLAM.

Sequence	Camera	Trajectory
1 · Church loop ☀️ day Closed loop around the university church (outer perimeter, archway not entered). Simplest sequence — small loop closure.
2 · Medium loop ☀️ day Loop around the church and the multifunctional building. Medium-scale loop closure.
3 · Medium loop — reverse ☀️ day · ↩️ reverse Exactly the route of seq. 2 in the opposite direction. Directional invariance.
4 · Figure-eight ☀️ day Self-intersecting path around both buildings — several loop closures in one run.
5 · Figure-eight — twilight 🌆 twilight Route of seq. 4 at twilight. Illumination is the only variable.

Stryiskyi Park — 6–10

Vegetation, moving branches, repetitive visual patterns and few planar structures.

Sequence	Camera	Trajectory
6 · Medium loop ☀️ day Closed loop through the park. Loop closure with few man-made structures.
7 · Medium loop — reverse ☀️ day · ↩️ reverse Route of seq. 6 in the opposite direction.
8 · Medium loop — twilight 🌆 twilight Route of seq. 6 at twilight. Low-light in a natural environment.
9 · Loop with elevation ☀️ day · ⛰️ elevation Uphill and downhill segments — full 6-DoF motion. Direction over the elevation segment is fixed.
10 · Figure-eight ☀️ day Self-intersecting path in the park; loop closure relies on natural visual features.

Linear — 11–13

Point-to-point routes for accumulated drift: errors cannot be corrected by revisiting places.

Sequence	Camera	Trajectory
11 · Linear A → B ☀️ day Point-to-point route from the campus into the park. No loop closure — pure drift.
12 · Linear B → A ☀️ day · ↩️ reverse Route of seq. 11 in the opposite direction.
13 · Linear A → B — twilight 🌆 twilight Route of seq. 11 at twilight. Drift under low light.

_{Map data in trajectory figures: © OpenStreetMap contributors.}

Limitations

Depth data is currently broken. Depth streams (/cam_depth/*, /cam_depth_aligned/*, depth/, depth_not_aligned/) should not be used in v1. A fix is in progress and will be released soon. RGB, IMU, wheel odometry and GNSS are not affected.

• Ground truth is position-only and meter-level. It is obtained by converting standalone GNSS fixes (no RTK) to ENU; there is no orientation, and accuracy may degrade further under tree canopy. It is suitable for coarse trajectory comparison, not for centimeter-level ATE/RPE.
• GNSS–base extrinsics are assumed to be identity; GNSS and wheel-odometry time offsets are not estimated.
• No LiDAR in v1.
• Planned: 6-DoF ground truth by fusing visual-inertial odometry, wheel odometry and GNSS; future releases with indoor–outdoor transitions and passages beneath structures.

Citation

@misc{ucu_slam_dataset_v1_2026,
  title     = {{UCU SLAM Dataset v1}},
  author    = {{Andriy Kryvyi, Hordii Yeliseev, Oleksandr Kosovan, Yaroslav Prytula}},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/ucu-autonomous-ugv/ucu-slam-dataset-v1}
}

Contact

UCU Autonomous UGV Club, Faculty of Applied Sciences, Ukrainian Catholic University · apps@ucu.edu.ua