Introduction

We present 8 datasets collected using legged-wheel robots, containing LiDAR data, IMU data, joint sensor data and, where available, ground truth. These datasets cover different challenging scenes. To the best of our knowledge, there are limited public datasets collected from legged-wheel robots. We hope our datasets enable the development of legged-wheel robot SLAM in the community.

Our datasets are now available, you can download by clicking the links below:

Sensor setup

Except for the DeepRobotics Lynx M20 half-stair stress-test sequence, the datasets were collected using the legged-wheel robot shown in Fig. 1. It’s equipped with 2 MID360 LiDAR, a BG-610M RTK GNSS, and a Realsense d435i RGB-D camera. Note that our datasets only include the LiDAR and IMU data from 2 MID360 LiDAR, the GPS-RTK data from RTK GNSS module and joints sensors data from the robot’s API. Data from d435i is not in the datasets.

Fig. 1. Our data collection platform.

LiDAR1(top-mounted, upside-down)：This LiDAR is mounted upside-down on the top of the robot. Its z-axis points downward, x-axis points forward along the robot’s heading direction, and the y-axis is determined by the right-hand rule (pointing to the robot’s left). It outputs 10 Hz LiDAR point cloud( topic: /livox/lidar_10_192_1_141) and 200 Hz IMU data( topic: /livox/imu_10_192_1_141).
The GNSS system has a centimeter-level localization ability and provides the vehicle’s ground truth pose in the Universal Transverse Mercator (UTM) coordinate system, at a frequency of 10 Hz.
The joint sensors output the joint position at a high frequency of 500 Hz( topic: /joint_states).
- For a detailed robot description, please refer to here.
- For a detailed data structure of joint position, please refer to here

Data Collection

We collected the datasets in different challenging scenes, primarily on the campus of Shanghai Jiao Tong University. During the data collection process, the average robot speed is under 1 m/s. In particular:

Staircase Scene 1 & 2: different staircases with step heights 5–15 cm, 1–8 steps, feature-rich railings on one side vs. feature-scarce on the other, including single-pass and multiple back-and-forth traversals.

Fig. 3. Left: staircase Scene 1 Dataset; Right: Staircase Scene 2 Dataset

Artificial Hill: stone-slab path with multiple single-step ledges; walked from the base to the summit and descend via the back trail, exhibiting pronounced elevation changes and varied surface textures.

Fig. 4. Artificial Hill Dataset

Rose Garden: paths interleaving short staircase and gently undulating grass fields, with low obstacles (flower beds, shrubs) providing moderate feature density. In addition, the staircase is flanked by walls, which leads to sparse features.

Fig. 5. Rose Garden Dataset

Botanical Garden: a route that covers a suspension bridge, an arch bridge, and flat paved sections; the dynamic sway on the suspension bridge and curved elevation on the arch pose challenges for SLAM.

Fig. 6. Botanical Garden Dataset

Indoor Staircase Only: This sequence contains a short indoor traversal in which the robot climbs a long staircase. It is used as a qualitative stress test and does not provide RTK ground truth.

Fig. 7. Indoor Staircase Only Dataset

Wet Grass: This sequence is collected after rain on wet grass terrain with an undulating overall elevation profile and local surface irregularities. It serves as a slippery-surface stress test.

Fig. 8. Wet Grass Dataset

DeepRobotics Lynx M20 Half-Stair Stress Test: This sequence uses a DeepRobotics Lynx M20 four-wheel-legged robot to climb a half-floor staircase. It is a qualitative cross-platform stress test with a different robot platform and a forward-facing LiDAR mounting configuration, and does not provide RTK ground truth.

Fig. 9. DeepRobotics Lynx M20 Platform and LiDAR Field of View for the Half-Stair Stress Test

License

This dataset is licensed under the Creative Commons Attribution 4.0 International License (CC BY 4.0).