MolmoMotion Stereo4D index references 70 train clips missing from the public Stereo4D GCS bucket

Hi Dear Authors,

We are reconstructing the `stereo4d` subset from the released MolmoMotion-1M data and found that 70 clips referenced by:

```text
stereo4d/track_index/stereo4d_track_index.json
```

cannot be downloaded from the documented public Stereo4D source:

```text
gs://stereo4d/{split}/{clip_id}.npz
```

## What we tried

The MolmoMotion `stereo4d/README.md` says to download upstream NPZ files from:

```bash
gsutil -m cp -r gs://stereo4d/train gs://stereo4d/test /path/to/stereo4d_npz/
```

We generated the Molmo-needed subset directly from `stereo4d_track_index.json` and attempted to download all 23,011 referenced clips.

## Result

Final manifest check:

```text
expected_manifest_entries: 23011
present_nonzero: 22941
zero_byte: 0
missing_or_zero: 70
missing_with_logged_404: 70
bad/corrupt downloaded npz: 0
```

Split breakdown:

```text
test: 2208 / 2208 complete
train: 20733 / 20803 complete
missing: 70 train clips
```

All 70 missing objects returned:

```text
errorCode=3 Resource not found
```

from URLs of the form:

```text
https://storage.googleapis.com/stereo4d/train/.npz
```

## Missing video groups

```text
30 4qVmP5JDa7U
18 SJ5gU2ECATc
10 4VX6HdoVElA
4 Ywx4gQVWEFk
4 K9qlbNKoGT8
1 ffeURUy7gjI
1 iFG8gCrTlLk
1 D5SamQj9rvY
1 9OxThHHGoKU
```

We also checked the public GCS bucket listing for these video ids under:

```text
train/
test/
demo/
bucket root
```

and found no objects for any of these missing video prefixes.

## Why this matters

`reconstruct_tracks.py` directly loads:

```python
np.load(npz_root / split / f"{clip_id}.npz", allow_pickle=True)
```

so these 70 missing NPZ files prevent strict reconstruction of the full released `stereo4d` subset. The RGB/video mapping exists in `stereo4d_videos_index.json`, but RGB alone is not enough to reproduce the exact Molmo tracks/camera because the source NPZ contains `camera2world`, 3D tracks, and related arrays.

## Request

Could you please clarify one of the following?

1. Are these 70 clips intentionally unused and should be filtered from the released index?
2. Is there an updated `stereo4d_track_index.json` / split file that excludes them?
3. Can the missing 70 NPZ objects be restored to the public `gs://stereo4d` bucket?
4. Does Ai2 have a mirror of these exact 70 NPZ files?

We can provide the full missing URL list if helpful.

Thanks.

## Local validation artifacts

```text
stereo4d_final_summary_20260705_160646.json
stereo4d_missing_all_20260705_160646.csv
stereo4d_missing_urls_20260705_160646.txt
stereo4d_404_from_logs_20260705_160646.txt
stereo4d_npz_open_summary_20260705_160858.json
```

Start by comparing stereo4d/track_index/stereo4d_track_index.json with the documented gs://stereo4d/train and test objects, then inspect reconstruct_tracks.py and the listed validation artifacts. Done means the release has a documented resolution: the 70 clips are restored, or the relevant index or split excludes them, with strict reconstruction succeeding.

由索引模型根據 Issue 內容生成。