espejelomar/go2-air-system-id-rgb-lidar-v0

Go2 Air System-ID RGB/LiDAR v0.1

This v0.1 dataset records a stock Unitree Go2 Air over local WebRTC while issuing bounded high-level Move commands and measuring the resulting odometry. It includes RGB snapshots, raw decoded built-in ULIDAR sidecars, robot state, commanded action parameters, and measured command outcomes.

The point of the dataset is system identification: high-level Go2 Air command parameters do not map cleanly to integrated motion. For example, in this capture x=0.10 for 0.6s is mostly deadband, while x=0.25 for 0.6s produces roughly 0.13m planar motion on this surface.

What Is Included

• —LeRobot v3-style data: data/chunk-000/file-000.parquet
• —RGB snapshot video: videos/observation.images.front/chunk-000/file-000.mp4
• —Original RGB snapshots: images/observation.images.front/chunk-000/frame_*.png
• —Raw decoded ULIDAR sidecars: lidar/ulidar/chunk-000/frame_*_ulidar.npz
• —Snapshot metadata: meta/snapshots.parquet
• —System-ID trial table: meta/system_id_trials.parquet
• —Hash manifest: MANIFEST.csv

Aggregate Command Outcomes

Privacy And Release Boundary

The RGB frames were reviewed by the dataset author. People may appear in the far background, but faces are not intentionally featured. Do not use this dataset for identity, surveillance, people tracking, or face recognition. If stricter privacy is required, use the LiDAR/odom sidecars and ignore RGB.

No robot IP address or serial number is included. The local private raw capture should not be redistributed separately without review.

Caveats

• —This is a small v0.1 release: one scene, one indoor floor surface, one Go2 Air, and 14 movement trials in the fixed raw-lidar run.
• —Go2 Air high-level Move parameters are treated as command parameters, not ground-truth velocities.
• —RGB is snapshot-based, not a continuous 30 FPS video stream. The video is an indexed view over captured state snapshots at 2 FPS.
• —Raw ULIDAR arrays are decoded as sidecar .npz files. They are intentionally not embedded into the MP4.
• —Firmware was not read directly from the WebRTC telemetry in this run.

Prior-Art Positioning

This is an early v0.1 system-ID artifact for stock Go2 Air command-outcome behavior using built-in lidar. It is adjacent to Go2 state-estimation and quadruped-lidar work, but the contribution here is the command-to-measured-odom layer plus decision-evidence-ready metadata.

Relevant adjacent work:

• —M-SEVIQ: Go2 event/vision/IMU/joint dataset, not command-outcome lidar system-ID: https://arxiv.org/abs/2601.02777
• —Go2 State Estimation Dataset: Go2 IMU/leg odometry data for state estimation: https://zenodo.org/records/19336009
• —Kine2Go: simulated kinematic Go2 data: https://huggingface.co/datasets/MIMUW-Robotics/kine2go
• —legkilo: Go1 with external Velodyne and leg kinematics: https://github.com/ouguangjun/legkilo-dataset
• —Go2-Dynamic-Inspection: Go2 navigation stack with external Livox Mid-360, not this dataset shape: https://github.com/YasiruDEX/Go2-Dynamic-Inspection

Suggested Citation

@dataset{go2_air_system_id_rgb_lidar_v0_2026,
  title = {Go2 Air System-ID RGB/LiDAR v0.1},
  author = {Omar Espejel},
  year = {2026},
  publisher = {Hugging Face},
  note = {Stock Unitree Go2 Air high-level command to measured odometry outcomes with built-in ULIDAR sidecars}
}