Action-free visual dynamics pretrain: frozen SmolVLA vision features → IDM → FDM. Trained on RGB from nvidia/libero_90 only (no action / state / reward labels).

Control run: trained with shuffle_future=True — the future RGB frame is sampled from a different episode, so temporal dynamics are broken. Same architecture / steps / data volume as the real IDM; use as a negative control for video-latent fine-tuning (VIDEO_CKPT=.../idm_fdm.pt from this repo).

Files:

Used as frozen teacher for video-latent few-shot SmolVLA fine-tuning (--video.checkpoint).