Flow-matching SFT dataset for Wan2.2 TI2V (5B) video generation on RoboTwin bimanual manipulation tasks. Derived from RoboTwin hdf5 demos, rendered as 3-view concatenated images (head + left arm + right arm cameras) at 640×736, uniformly sampled to 121 frames per episode.
robotwin_8tasks_121f.json: 160 entries, each with {prompt, filename_stem, media_path, video_path}.
videos/: 160 MP4 videos (H.264, 640×736, 121… See the full description on the dataset page:
https://huggingface.co/datasets/VincentNi/robotwin_sft_8tasks_121f.