VincentNi/robotwin_sft_8tasks_121f

RoboTwin SFT Dataset — 8 Tasks × 121 Frames

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.

Contents

• robotwin_8tasks_121f.json: 160 entries, each with {prompt, filename_stem, media_path, video_path}.
• videos/: 160 MP4 videos (H.264, 640×736, 121 frames, 10 fps).
• frames/: 160 PNG reference images (lossless, first frame of each video).

Tasks (20 demos each)

• dump_bin_bigbin
• handover_block
• handover_mic
• open_laptop
• open_microwave
• place_burger_fries
• place_cans_plasticbox
• place_dual_shoes

JSON entry format

{
  "prompt": "The whole scene is in a realistic, industrial art style ...",
  "filename_stem": "dump_bin_bigbin__demo_clean_ep20_vidar__episode0",
  "media_path": "frames/dump_bin_bigbin__demo_clean_ep20_vidar__episode0.png",
  "video_path": "videos/dump_bin_bigbin__demo_clean_ep20_vidar__episode0.mp4"
}

Source

Derived from RoboTwin demonstration episodes.

Download

from huggingface_hub import snapshot_download
local = snapshot_download("VincentNi/robotwin_sft_8tasks_121f", repo_type="dataset")