luckyor/A-S

AerialVLN-S DreamFly Dataset Release v1

Canonical storage path: /root/autodl-tmp/DATA/AerialVLN-S.

This is a reproducible, RGB-only AerialVLN-S training dataset for DreamFly. It preserves the OpenFly-compatible episode TFRecord layout while keeping an independent AerialVLN-S native action contract.

Contents

• raw/episodes.jsonl: auditable train-only raw episode manifest. Each row references current-view RGB images, source actions, native model actions, world poses, instruction, and per-image SHA-256.
• raw/images/<source_episode_id>/frame_*.png: RGB 224x224 current frames. These files are hard-linked to the capture archive, not duplicated pixels.
• projection/vlnvN_clean/1.0.0/*.tfrecord-*: OpenFly-style sharded records; one complete episode per tf.train.Example, 128 episodes per shard maximum.
• projection/provenance/source_annotations/: byte-preserved official AerialVLN-S train, val_seen, val_unseen, and test JSON files.
• projection/source_to_dreamfly_episode_id.json: stable source-to-numeric mapping. Do not reorder after building Memory.
• provenance/: capture configuration, excluded trajectory policy, source annotation hash, action schema, and scene-to-subset mapping.

Dataset Contract

The training projection contains exactly 10,091 eligible AerialVLN-S train episodes and 1,168,659 decision steps. Twenty-two episodes from eight source trajectories are excluded by the documented persistent-uniform-frame policy.

Each record has the OpenFly-compatible fields episode_metadata/* and steps/*, with only steps/observation/image_1. The native source/model action IDs are both AerialVLN-S IDs:

0 Stop, 1 Forward 5m, 2 Turn Left 15deg, 3 Turn Right 15deg, 4 Ascend 2m, 5 Descend 2m, 6 Move Left 5m, 7 Move Right 5m.

Do not load this data with the generic OpenFly action-vector interpretation. Use dreamfly.data.aerialvln_dataset.AerialVLNTFRecordDataset and the aerialvln8 action decode policy. The two OpenFly-only forward-distance slots are not present. steps/history is deliberately blank: Memory-v1 has not been built and no future information is embedded.

episode_metadata/source_annotation_json contains the exact official episode object for each training example. The four original split JSON files are also archived in provenance so that val/test, which are not supervision data here, remain available for evaluation tooling.

Verification

The completed release passed validate_aerialvln_dreamfly.py over all records: 10,091 examples and 1,168,659 steps. The verifier decodes every PNG, checks RGB 224x224 dimensions, sequence lengths, terminal/discount flags, native action vectors, and the single-current-image contract.

Per-shard SHA-256 values are in projection/build_report.json; hashes for official annotations are in projection/provenance/source_annotations_manifest.json.

Publishing

Before uploading to Hugging Face, confirm AerialVLN-S's upstream data and rendering asset redistribution terms. Do not upload simulator binaries, Dream-VLA weights, credentials, or model checkpoints. The training projection contains train RGB only; val_seen, val_unseen, and test JSON files are metadata/evaluation annotations and have no captured RGB in this release.