# RoMALab/video2vla-robotwin-randomized500-cache

RoboTwin randomized500 latent cache generated by video2vla.

- Tasks: 50
- Splits: train and val
- Camera: head_camera
- Hunyuan target_size: 384
- Latent shape: `(1, 32, 4, 21, 27)`
- Latent dtype: `torch.bfloat16`
- Cache archives are under `cache/`, one archive per task.
- Metadata archives are under `metadata/`.

See `manifest.json` for per-task archive names and paths.