robocasa_mg_gr00t_300_val

RoboCasa-Kitchen (Panda+OmronBase) held-out validation set for the wan22 WAM DiT4DiT video-prediction / probing evaluations — the val counterpart of the robocasa_mg_gr00t_300 training set (episodes drawn from the mg1000 holdout source episodes 300–349).

• LeRobot v2.0 layout (data/ parquet · meta/ · videos/ MP4, 3 views: wrist_view / left_view / right_view), 20 fps
• 1,200 episodes · 344,486 frames · 365 task strings
• meta/val_windows_2048.json — the fixed 2,048-window subsample used by vidmetrics_val2048 (PSNR/SSIM/LPIPS/FVD): per window {episode_index, source_episode_index, task_type, task, t0, label}; deterministic generation seed downstream = hash((source_episode_index, t0))
• meta/val_windows.json — the full window list the 2,048 were sampled from

Provenance: videos re-materialized from the shared preprocessed mg1000 copy (the original symlink targets were cleaned up upstream); content is identical to the set used for all published val2048 numbers (2026-07 vidmetrics runs).