U2UData+
Swarm-UAV autonomous flight dataset for embodied long-horizon tasks: 15 UAVs, 12 scenes, 720 traces, 120 hours of RGB/depth/LiDAR plus environmental sensing.
| Year | 2025 |
|---|---|
| Episodes | 720 |
| Total hours | 120 |
| Embodiments | drone |
| Modalities | depth, point_cloud, proprioception, rgb |
| Task categories | navigation, manipulation |
| License | Apache-2.0 |
| Access | open — commercial use permitted |
| Maintainer | Tsinghua University — THU-EAI Lab |
U2UData+ (published under the name U2UData-2) is a large-scale swarm-UAV dataset for Embodied Long-Horizon tasks, collected entirely in the U2USim UE5.2/AirSim simulator on terrain mapped from Yunnan Province. Fifteen UAVs fly in autonomous formation mode executing a wildlife-conservation task across 12 weather and terrain scenes, producing 720 traces of 600 s each — 12.96M RGB frames, 12.96M depth frames, 4.32M LiDAR sweeps, plus brightness, temperature, humidity, smoke and airflow readings. It ships with an online data-collection and closed-loop verification platform, 3D bounding boxes for 15 object classes, and a collaborative-tracking benchmark over 9 methods. Accepted at AAAI 2026; supersedes the smaller U2UData (ACM MM 2024, 3 UAVs / 8.75 h).
Partial public release: only 1 of the 12 scenes (~62 GB) is on HuggingFace. The full 3.62 TB requires emailing the authors for a Baidu Cloud link. The size shown here is the full dataset.