Dataset

Yonder — drone-perspective indoor navigation at scale

Yonder pairs a very large drone-view indoor corpus with a clear evaluation story: offline perception metrics are not a substitute for closed-loop flight tests, especially when training and evaluation simulators differ.

Headline facts
  • Millions of drone-perspective frames across many indoor scenes
  • Rich sensing per waypoint (stereo RGB, depth, IR, panoramic LiDAR-style data, semantic segmentation, pose)
  • Public host: https://huggingface.co/datasets/astralhf/yonder
  • License inherits HSSD non-commercial terms (CC-BY-NC-4.0); check the dataset card before commercial use
Who this is for

Robotics and ML teams training open-vocabulary detectors, depth estimators, semantic models, and other perception modules for drone-view indoor navigation.

Researchers studying sim-to-sim transfer: the dataset is explicitly motivated by cross-simulator pitfalls that are easy to miss if you only watch offline mAP move.

Quick start

The Hugging Face README includes a minimal download snippet (single-scene smoke test) and notes on repository layout. For a tiny local download before multi-terabyte transfers, start with astralhf/yonder-sample.

from huggingface_hub import snapshot_download

path = snapshot_download(
    repo_id="astralhf/yonder",
    repo_type="dataset",
    allow_patterns="indoor/drone-data/augmented/hssd-102343992/*.npz",
)