CognitiveDrone Dataset
8,062 simulated UAV trajectories pairing egocentric FPV frames with language instructions and 4-DoF velocity actions, for drone tasks that need reasoning, not just flight control.
| Year | 2025 |
|---|---|
| Episodes | 8,062 |
| Embodiments | drone |
| Modalities | language, proprioception, rgb |
| Task categories | navigation, manipulation |
| License | unknown |
| Access | open |
| Maintainer | Skoltech — Intelligent Space Robotics Laboratory |
CognitiveDrone is a simulated UAV vision-language-action dataset from Skoltech's Intelligent Space Robotics Laboratory, built to train drones on tasks that require reasoning rather than pure flight control. Each trajectory spawns the quadrotor before a set of virtual gates and asks it to fly through the correct one by solving an embedded cognitive task, across three categories: Human Recognition, Symbol Understanding, and Reasoning. Episodes are generated in Gazebo with ArduPilot SITL, recording 256x256 first-person RealSense frames plus 4D velocity commands (Vx, Vy, Vz, ω) at each step.
Two caveats: the release declares no license, and only 30 of the intended 128 RLDS shards were actually uploaded (with no TFDS metadata), so roughly a quarter of the 8,062 trajectories are retrievable from HuggingFace. The 436 complete raw validation episodes in the collector repo are fully usable — the sample frames here come from those.