You need to agree to share your Hugging Face information to access this dataset
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
This dataset is the training data of the UAV Perception Testing Track at ICST 2027. Access is granted on request by the organizers. Please tell us who you are and how you intend to use the data.
Log in or Sign Up to review the conditions and access this dataset content.
UAV Perception Testing Track flight data
357 simulated quadrotor flights over 336 procedurally generated forest scenarios. This is the training data of the UAV Perception Testing Track at ICST 2027; the rules, the starter kit and the scenario format are in the entrant repository.
Every flight runs one fixed autonomy stack in closed loop in Isaac Sim 6.0.1. A simulated RGB-D camera feeds nvblox mapping while EGO-Planner plans on the map and commands PX4 in software-in-the-loop. Outcomes are adjudicated against the scene's collision geometry rather than against what the vehicle believed.
| Scenarios | 336 |
| Flights | 357 |
| Outcomes | 183 goal, 147 collision, 23 failed and 4 timeout |
| Admissible flights | 329 |
| Cleared lane | 2.0, 3.0, 4.5, 6.0, 9.0 and 12.0 m |
| Size | 17.8 GB |
The scenarios are the train and validation splits of the track, listed in splits.json. A further private split is withheld and is not part of this dataset.
Most work needs only the manifests
The per-chunk manifest.jsonl files carry the scenario-level record: the parameters that were flown, the route, the placed obstacles, the outcome and the failure mechanisms reached. That is the whole training signal for a scenario generator. The remaining files hold depth and map arrays, autopilot logs and the per-flight scene files, which are needed only for work on the perception or control layers.
from huggingface_hub import snapshot_download
import json, glob
d = snapshot_download("merabro/uav-perception-testing-track-2027", repo_type="dataset",
allow_patterns=["*/manifest.jsonl", "*.md", "*.json"])
rows = [json.loads(l) for f in glob.glob(f"{d}/coverage_*/manifest.jsonl")
for l in open(f) if l.strip()]
Download a subset rather than cloning the repository: the collection has about 34 thousand files.
Structure
coverage_NNN/ chunk directory
manifest.jsonl one row per flight in the chunk
adjudication.jsonl the outcome adjudication of each flight
submission.json the scenarios of the chunk, in the submission format
<scenario>_r<n>/ per-flight directory: arrays, autopilot log, scene file
splits.json the train and validation scenario ids
Sixteen scenarios were flown again after a collection failure, so their scenario and repeat appear in two manifest rows while the flight directory holds the files of one attempt. Match a row to its directory by stack.run_id, which equals the run_id in the directory's f10_trace_*.json; the row without files is always marked invalid.
admissible marks flights that produced usable evidence. A run that failed to collect says nothing about whether its scenario was difficult, so drop those rows before fitting a model.
License and citation
The data are released under CC BY 4.0. Work that uses them should cite the UAV Perception Testing Track (https://ideals-lab.github.io/uav-testing-track-2027/).
- Downloads last month
- 12