greenery8848/skywander-nyc-sim

SkyWander NYC Sim — Synthetic UAV Flight Corpus

2009 synthetic UAV flight clips rendered in an Unreal Engine New York City environment, with full low-dimensional ground truth. Training/eval corpus of the SkyWander pose-conditioned UAV world model.

Contents (per clip, keyed by uuid)

Directory	Files	What
rgb_mp4/	2009	ego-view RGB video, 1280×704
depth_mp4_gemdepth/	2009 mp4 + 2009 json	GemDepth monocular depth (gray-normalized) + per-clip scale metadata sidecars — keep the json next to the mp4
depth_mp4_vggt_conf30/50/70/	2009 each	VGGT-Omega re-rendered imperfect depth at three confidence thresholds
state_h5/	2009	per-frame UAV state (pose, velocity, …), HDF5
depth_npz.tar	2003 subdirs	raw generator camera poses + intrinsics (camera_poses.txt, intrinsic.txt) — untar after download; 6 clips lack raw txts (see audit_triplets/)
poses_npz/	2009	VGGT-Omega pseudo-pose estimates for every clip
audit_triplets/, triplet_match_manifest.csv	—	integrity manifests; re-run the check scripts from the code repo after download

Notes:

• The environment has no simulator ground-truth depth; all depth channels are model-predicted pseudo-labels by design ("loose geometry hint" conditioning).
• Camera center = UAV body pose translated +1 m along body X (known constant).

License

The RGB videos are rendered output of a commercially licensed Unreal Engine Marketplace environment (NYC Environment Megapack). The source 3D assets are not included and cannot be reconstructed from this data. Released for research use; if you redistribute derivatives, keep this notice.