OmniHM-Intent

An omni-modal benchmark for evaluating intent-to-humanoid motion generation

Evaluation-only split. OmniHM-Intent contains fixed benchmark samples selected from the OmniHM training-data pool. It is intended to evaluate models that were not trained on these exact samples. Any overlap with a model's training data must be disclosed.

Overview

OmniHM-Intent evaluates whether a humanoid motion model can follow intent expressed through four input modalities: natural-language text, speech, music, and RGB video. Every track contains a fixed test split of 2,048 samples and a paired Unitree G1 motion preview, for 8,192 benchmark records in total.

Dataset at a glance

Config	Samples	Intent condition	Paired reference	Duration
text	2,048	English motion prompt	G1 motion render	2–15 s
speech	2,048	English speech + transcript	G1 motion render	2.03–15 s
music	2,048	Full music waveform	G1 motion render	2.05–12 s
video	2,048	RGB human-motion clip	G1 motion render	2–8 s

Benchmark preview

The complete media are available in the Hugging Face Dataset Viewer:

• Text →
• Speech →
• Music →
• Video →

Data format

Each config uses self-contained Parquet shards. Audio and video payloads are embedded directly in the corresponding Hugging Face Audio and Video features, following the layout used by OmniHM-5M.

Common fields include:

Field	Description
sample_id	Stable, zero-padded benchmark identity
duration_sec, fps, motion_frames	Valid reference-motion timeline
motion_preview	Unitree G1 reference render

Track-specific condition fields are:

• Text: prompt
• Speech: condition_audio, transcript
• Music: condition_audio
• Video: condition_video

Aggregate sample and shard counts are published under metadata/.

Loading

Install a recent version of 🤗 Datasets. Video/audio decoding additionally requires TorchCodec and FFmpeg:

pip install -U datasets torchcodec

Load one config in streaming mode:

from datasets import load_dataset

dataset = load_dataset(
    "wendell0218/OmniHM-Intent",
    "text",
    split="test",
    streaming=True,
)
sample = next(iter(dataset))

print(sample["sample_id"])
print(sample["prompt"])
print(sample["motion_preview"])

Choose another modality by replacing text with speech, music, or video.

Numerical ground truth

The Dataset Viewer is limited to the four media-oriented Parquet configs above. Numerical Unitree G1 reference trajectories are stored separately under eval_assets/v1/; they are downloadable repository files and do not appear as Viewer columns.

Download GT only

The following downloads the four numerical references and their manifest without downloading the approximately 16 GB of Viewer Parquet data:

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="wendell0218/OmniHM-Intent",
    repo_type="dataset",
    allow_patterns=[
        "eval_assets/v1/manifest.json",
        "eval_assets/v1/*/reference_qpos.npz",
        "README.md",
        "SHA256SUMS",
    ],
    local_dir="OmniHM-Intent",
)

GT format

Each track provides one compressed reference_qpos.npz with the following arrays:

Array	Type and shape	Description
qpos_36	float32 [2048, T, 36]	root_xyz[3] + root_quaternion_wxyz[4] + G1_joint_angles[29]
fps	float32 [2048]	Reference frame rate
valid	bool [2048, T]	Valid-frame mask
lengths	int32 [2048]	Number of valid frames per sample
sample_id	string [2048]	Canonical sample order matching the corresponding Viewer config
joint_names	string [29]	Unitree G1 joint order used by qpos_36

The padded horizon T is 450 for Text and Speech, 359 for Music, and 240 for Video. Padding is excluded by valid and lengths. Root quaternions use normalized wxyz order on valid frames. File sizes and SHA-256 hashes are recorded in eval_assets/v1/manifest.json and SHA256SUMS.

Dataset integrity

The public release is checked for structural integrity:

1. All four configs contain exactly 2,048 unique, continuously numbered samples.
2. Every condition is paired with one G1 motion preview.
3. Published audio and video payloads are embedded and readable.
4. Parquet schemas, numerical GT contracts, and repository file hashes are verified before publication.

Aggregate counts are in metadata/dataset_summary.json; repository-wide file hashes are in SHA256SUMS.

Intended use

OmniHM-Intent is designed for research on:

• text-, speech-, music-, and video-conditioned humanoid motion generation
• multimodal intent following and cross-modal comparison
• motion quality and intent-alignment evaluation
• benchmark development for Unitree G1 motion

It is not a training split, an identity-recognition dataset, or a safety certification for deployment on physical robots.

Limitations and responsible use

• The benchmark is sampled from a larger training-data pool and may overlap with independently collected web data. Evaluate and disclose train/test contamination where applicable.
• Internet video and audio can carry cultural, geographic, activity, appearance, and recording-style biases.
• Captions and transcripts are partly machine-generated and may contain errors.
• Music and video can contain third-party content. Users are responsible for complying with applicable copyright, privacy, and data-use requirements.

License

This dataset card declares license: other, matching the related OmniHM-5M release. No CC, Apache, commercial-use, or redistribution permission should be inferred. Users must review the applicable source-content rights and final project release terms before use or redistribution.

Citation

OmniHM-Intent accompanies the OmniHM project. An official BibTeX entry will be added when the corresponding report is publicly available. Until then, cite the OmniHM project page and this dataset repository.