UAV-Flow
30,692 real expert-piloted UAV flights paired with atomic language instructions and egocentric video, for fine-grained language-conditioned drone control.
| Year | 2025 |
|---|---|
| Episodes | 30,692 |
| Embodiments | drone |
| Modalities | language, proprioception, rgb |
| Task categories | manipulation, navigation |
| License | unknown |
| Access | open |
| Maintainer | Beihang University — Institute of Artificial Intelligence |
UAV-Flow is the first real-world benchmark for language-conditioned, fine-grained UAV trajectory control, framed as the "Flying-on-a-Word" (Flow) task. Certified pilots — each with 800+ hours of experience — manually flew 30,692 short-range trajectories across three university campuses (5.02 km²) on DJI Mavic 3T RTK platforms, flying exclusively from the drone's first-person view. Every trajectory pairs an atomic natural-language instruction ("go through the lamp post from the right side") with egocentric onboard video and RTK-accurate 6-DoF state. Accepted at NeurIPS 2025; a companion simulated split is catalogued separately as UAV-Flow-Sim.
Note: the code is Apache-2.0 but neither HuggingFace dataset repo declares a data license. Total recording hours are not stated in the paper.