---
license: other
license_name: mixed-research-use
viewer: false
tags:
- robotics
- 4d-reconstruction
---
# Evaluation data for an anonymous submission

Everything `scripts/evaluate.py` of the supplementary code reads for the main table: the 256 held-out DROID
clips (recorded and Cosmos-generated inputs), their labels, and the robot model files. For **research use in
reviewing the submission**; see `LICENSES.md` for the terms of each component.

| Folder | Content | `configs/paths.yaml` key |
|---|---|---|
| `data_root/` | per clip: `<clip>.npz` (RGB, VAE latents, joints, gripper, cameras) + `.camera.npz`, `.depth.npz`, `.stereo_right.npz`, `.tracks.npz`, `.pwtracks.npz` | `data_root` |
| `split.json` | the 256 evaluation episodes and their split | `split_manifest` |
| `right_depth/` | right-eye stereo depth labels | `right_depth_labels`, `stereo_depth_labels` |
| `real_window_latents/` | Wan2.2-VAE latents of the evaluated windows (LatentStem arms) | `real_window_latents` |
| `cosmos_rollouts/` | Cosmos 3 rollouts of the evaluated windows (`--inputs cosmos`) | `cosmos_rollouts` |
| `droid_calibration/`, `droid_trajectories/` | DROID calibration and trajectory files of the 256 episodes | `droid_calibration`, `droid_trajectories` |
| `robot/meshes/`, `robot/gripper2wrist_transforms.json` | Franka meshes and gripper-to-wrist transforms | `robot_meshes`, `gripper_to_wrist_transforms` |
| `reference_results/` | per-clip records, `summary.json` and `table.md` produced with the supplementary code (recorded inputs, Cosmos inputs, recorded-input ablations) | — |

Download everything (21.9 GB):

```
hf download anonymous61116/supp-eval-data --repo-type dataset --local-dir eval_data
```

or only some clips (clip names are in `split.json`; indices refer to the code's `configs/data/evaluation_clips.json`):

```python
from huggingface_hub import snapshot_download
clip = 'clip_ILIAD_50aee79f_2023-07-12-20h-55m-08s_0068'            # evaluation index 0
episode = 'ILIAD__success__2023-07-12__Wed_Jul_12_20:55:08_2023'
snapshot_download('anonymous61116/supp-eval-data', repo_type='dataset', local_dir='eval_data', allow_patterns=[
    f'data_root/{clip}*', f'right_depth/depth/{clip}*', 'right_depth/provenance.json', 'split.json',
    'real_window_latents/clip000.npz', 'cosmos_rollouts/clip000_repeat00/*', f'droid_calibration/{episode}*',
    f'droid_trajectories/{episode}*', 'robot/*', 'robot/meshes/*', 'robot/meshes/*/*'])
```

Then point the keys above to the downloaded folders, add the model locations
(MoVieS, VGGT, ReSplat and the three checkpoints), then run `scripts/evaluate.py` and `scripts/summarize.py` as in
the code's README and compare with `reference_results/`. Stage 1 records match bit for bit; Stage 2 records
match up to GPU non-determinism. `--indices` evaluates a subset of the clips.

Local file paths recorded as provenance strings inside the original label files were reduced to file names; the
arrays are unchanged.