RoMoDataset/RoMo-Hands

RoMo Hands

RoMo Hands: A Large Scale Richly Organized Text to Hand Motion Dataset (NeurIPS 2026)

Project page · Paper · Motion Toolbox · RoMo (full body)

RoMo Hands is an in-the-wild, text-paired dataset of bimanual hand motion: 541,929 sequences, 884 hours, and 95.5M frames at 30 fps. Every sequence is paired with five captions of increasing detail, from a short action tag to a paragraph describing finger articulation, temporal dynamics, and spatial relations.

Summary

Property	Value
Sequences	541,929
Duration	884 h (95.5M frames at 30 fps; mean clip 5.9 s)
Splits (train / val / test)	464,278 / 50,985 / 26,666 (85% / 10% / 5%, segment level)
Motion	393-D bimanual representation per frame (layout below)
Captions	5 per sequence: tag, short, medium, detailed, paragraph
License	CC BY-NC 4.0, non-commercial research use only

Clips range from 30 to 600 frames. The paper's benchmark scores only the 17,348 test clips of 40–196 frames (listed in splits/benchmark_test.txt); evaluating on the full test split will not reproduce the paper's numbers. During training, clips longer than 196 frames are randomly cropped to 196 frames and clips shorter than 40 frames are zero-padded to 40.

Files

The data is stored in LitData streaming format:

train/train/   index.json + chunk-*.bin
val/val/       index.json + chunk-*.bin
test/test/     index.json + chunk-*.bin
Mean.npy       per-dimension mean of the 393-D representation
Std.npy        per-dimension standard deviation (constant dimensions set to 1)

Each sample is a dict:

{
    "motion": np.ndarray,       # (T, 393) float32, unnormalized
    "texts": list[str],         # [tag, short, medium, detailed, paragraph]
    "sample_path": str,         # source segment identifier
}

Video-disjoint splits

The released split is assigned per segment, so some source videos appear in more than one split (15,152 videos are shared by train and test). For leakage-free use, splits/ provides:

• video_disjoint_test.txt: the 7,679 released test segments (7,596 videos) whose source video appears in neither train nor val. Use it to evaluate models trained on the released train split.
• video_level/{train,val,test}.txt: a strict per-source-video split (461,203 / 53,935 / 26,791 segments) for training new models; each video is assigned by sha256(video_id) mod 100 (<85 train, <95 val, else test).
• benchmark_test.txt: the 17,348 test segments of 40–196 frames used by the paper's benchmark.
• split_stats.json: segment and video counts, and the video-id rule.

Segment ids match each sample's sample_path; the source video id is the prefix before _segment_.

Loading

from huggingface_hub import snapshot_download
from litdata import StreamingDataset

root = snapshot_download("RoMoDataset/RoMo-Hands", repo_type="dataset", allow_patterns="test/*")
dataset = StreamingDataset(f"{root}/test/test")   # 26,666 clips; use train/train or val/val likewise
sample = dataset[0]
motion = sample["motion"]        # (T, 393)
tag, short, medium, detailed, paragraph = sample["texts"]

Motion representation

Each frame is a 393-D vector over the 32 MANO joints of both hands (joints 0–15: left hand, 16–31: right hand; wrists at 0 and 16). The root is the midpoint of the two wrists, and the heading is a yaw-only rotation that aligns the inter-wrist segment with the world x-axis.

Dims	Channel	Width
0–1	Root XZ velocity, heading-aligned	2
2	Root height (absolute world y)	1
3–8	Heading rotation, 6D	6
9–104	Joint positions, root-relative and heading-aligned (32 × 3)	96
105–200	Joint velocities, finite differences (32 × 3)	96
201–392	Joint rotations, 6D (32 × 6); wrists in the heading frame, fingers parent-relative	192

See Section 3.3 and Appendix B of the paper for the full definition and the type-aware normalization used for training.

Captions

Level	Mean words
L0 tag	3.5
L1 short	13.6
L2 medium	49.4
L3 detailed	92.5
L4 paragraph	149.3

Captions were generated with Qwen3.5-27B from the source video. A 250-clip human study (12 raters, 370 ratings) rated overall action match at 4.16 / 5 on average.

How it was built

RoMo Hands extends the RoMo collection pipeline. Hands are recovered per frame with WiLoR and anchored to RoMo's temporally consistent body wrists, short gaps are interpolated, and clips are filtered by per-hand detection ratio and a dexterity score before captioning.

License

CC BY-NC 4.0: free for non-commercial research use with attribution; commercial use is not permitted. Reconstructing hand meshes requires the MANO model, which is distributed separately under its own license at https://mano.is.tue.mpg.de/.

Citation

@inproceedings{BenShabat2026RoMoHands,
  author    = {Ben-Shabat, Yizhak and Zhang, Jiahao and Liu, Joseph and Moon, Seonghyeon and Lee, Young-Yoon and Jiang, Haomiao and Jacob, Oren and Kapadia, Mubbasir},
  title     = {{RoMo Hands}: A Large Scale Richly Organized Text to Hand Motion Dataset},
  booktitle = {Advances in Neural Information Processing Systems (NeurIPS)},
  year      = {2026},
}