UAV-Flow-Sim
Simulated companion to UAV-Flow: 10,109 language-conditioned drone trajectories in an Unreal Engine campus, used as the systematic evaluation suite.
| Year | 2025 |
|---|---|
| Episodes | 10,109 |
| Embodiments | drone |
| Modalities | language, proprioception, rgb |
| Task categories | manipulation, navigation |
| License | unknown |
| Access | open |
| Maintainer | Beihang University — Institute of Artificial Intelligence |
UAV-Flow-Sim is the simulated half of the UAV-Flow benchmark, built in an Unreal Engine campus environment and used for systematic, repeatable evaluation of language-conditioned UAV policies. It contains 10,109 trajectories spanning 10 Flow task types, with the same schema as the real split — atomic language instruction, egocentric RGB observation, and 6-DoF UAV state. Collection was hybrid: human pilots manually flew trajectories by locating landmarks in the simulator, while a rule-based collector exploited the simulator's ground-truth scene structure. The paper gives no breakdown between the two, so both collection modes are tagged.
Kept as its own entry rather than merged into UAV-Flow: different realness, a ~7x size difference, and a different role (evaluation vs training).