RoVid-20K-10s

RoVid-20K-10s is a curated, fixed-length subset of DAGroup-PKU/RoVid-X for robotic image-to-video training. Each sample is a single 10.5417-second shot normalized to 1280x704, 24 FPS, and exactly 253 decoded frames.

The intended teacher-forcing task is:

video frame 0 + task instruction -> future video frames

Frame 0 is decoded directly from the MP4; it is not stored as a duplicate image file. The model-facing caption field is exactly the original RoVid-X task_caption. Original short_caption and detailed_caption values are retained as metadata for analysis, but they are not the primary conditioning text. Automatically generated visual captions used during quality auditing are not published as training conditions.

Dataset statistics

Split	Samples
train	20,311
validation	201
test	208
total	20,720

• Total video duration: 60.67 hours
• Resolution: 1280x704
• Frame rate: 24 FPS
• Decoded frames per video: 253
• Duration per video: 10.5417 seconds
• Video codec: H.264 in MP4
• Conditioning frame: frame index 0

Format

The dataset is distributed as deterministic WebDataset-style tar shards. Each sample has two members with the same 32-character hexadecimal key:

<sample_id>.mp4
<sample_id>.json

The JSON record contains at least:

{
  "id": "...",
  "split": "train",
  "caption": "pick up the object",
  "task_instruction": "pick up the object",
  "source_short_caption": "...",
  "source_detailed_caption": "...",
  "conditioning": {"mode": "i2v", "frame_index": 0},
  "width": 1280,
  "height": 704,
  "fps": 24,
  "frame_count": 253,
  "duration_seconds": 10.541667,
  "source_dataset": "DAGroup-PKU/RoVid-X",
  "source_revision": "32faf783ba532e1107881b29c612be0aec16b888",
  "source_shard": "00",
  "source_filename": "<sample_id>.mp4",
  "training_latent_frames": 64,
  "video_size_bytes": 1234567
}

source_short_caption and source_detailed_caption are provenance fields only. Training readers should use caption (identical to task_instruction) as the text condition.

Metadata JSONL manifests and SHA-256 checksums are included for reproducibility.

Curation and changes from RoVid-X

This is a derived subset, not an official DAGroup-PKU release. The source data was pinned to revision 32faf783ba532e1107881b29c612be0aec16b888. We selected videos that could supply a complete fixed training window, removed hard scene cuts and invalid media, filtered for single-shot robot visibility and usable action content, removed exact temporal duplicates, and normalized accepted clips with aspect-preserving scale-to-cover plus center crop.

All 20,720 published videos were decoded and verified to satisfy the stated spatial and temporal contract. The LongLive raw-video reader was also tested on the materialized dataset.

Intended use

• Robotic image-to-video domain adaptation
• Autoregressive video teacher forcing
• Video/world-model pretraining before adding action or robot-state modules
• Evaluation on held-out robotic video-generation benchmarks such as RBench and EWMBench

Limitations

• This dataset contains RGB video and language metadata, not synchronized robot actions, proprioception, calibration, or force signals. It does not by itself train an action-conditioned world model or executable robot policy.
• The subset is manipulation-heavy and should not be assumed to balance all embodiments.
• Source task captions can be concise or generic. The reference frame supplies additional scene and object context.
• The source pool used 16 uniformly selected RoVid-X shard prefixes rather than the entire RoVid-X release.
• CC BY 4.0 covers the licensor's copyright and applicable database rights; it does not by itself grant privacy, personality, publicity, trademark, or other third-party rights. Downstream users are responsible for assessing those rights for their use case.

License and attribution

The source RoVid-X dataset is released under CC BY 4.0. This derived subset is distributed under the same license. Users must retain attribution to the original dataset and paper and should clearly indicate any further modifications. This derived release is not an official DAGroup-PKU release and does not imply endorsement by the original authors.

@article{deng2026rethinking,
  title={Rethinking Video Generation Model for the Embodied World},
  author={Deng, Yufan and Pan, Zilin and Zhang, Hongyu and Li, Xiaojie and Hu, Ruoqing and Ding, Yufei and Zou, Yiming and Zeng, Yan and Zhou, Daquan},
  journal={arXiv preprint arXiv:2601.15282},
  year={2026}
}