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.