Episodes Preview Unitree G1 Visualizer
798k episodes · 50 fps

OMG-Lite-Data: 50 Hz multimodal motion

A deterministic 50 Hz derivative of THU-MARS/OMG-Data at revision 6e0dfbc1c5298bff14d4e2b1459ad678af0a38e7. All source episodes, captions, condition availability, source identities, and train/val/test membership are retained. Consult the upstream dataset for source dataset licenses and attribution.

Representation and time

LeRobot v3 public schema retains 36D qpos (XYZ, WXYZ quaternion, 29 joint angles), next-state action, 35D audio, 66D human reference, and modality masks. OMG-Lite encodes qpos into 38D (29 joints + XYZ + 6D rotation); 38D is not raw qpos. Uniform timestamps are j/50 within [0,(N-1)/30]. No extrapolation or time scaling. Translation, bounded joint angles, and human joint positions use linear interpolation; root quaternions use shortest-arc SLERP. Action is rebuilt from next target state, with terminal repetition. Source frame offsets remain in original 30 Hz units. The last target timestamp can precede the original endpoint by at most 13.334 ms. This resampling does not create new measured motion detail.

Conditions

The upstream audio feature definitions differ by source family. AIST++/FineDance channels 33 and 34 are binary peak/beat events: events are assigned to the nearest in-domain 50 Hz timestamp. Other channels are linearly interpolated. For AIOZ-GDance, ChoreoMaster, COMPAS3D, and OpenDance all 35 channels are continuous and linearly interpolated. These are resampled released features, not features re-extracted from raw audio. They must not be assumed to share one audio encoder definition. Human reference is flattened 22x3 Cartesian joint positions in the source convention. Missing conditions and availability flags are preserved; no conditions are fabricated.

Reproducibility

provenance/build_report.json records algorithms, counts, source/output hashes and versions. Episode metadata retains original provenance plus resampling lengths and source FPS. Official splits are unchanged, including any upstream overlap: this release does not claim recording-level or motion-content deduplication. meta/stats.json contains frame statistics, not 38D model normalization. Compute model normalization from training windows only. Existing checkpoints retain their own stats.

Read splits through meta/info.json and episode metadata. Parquet files can cross split boundaries; do not treat an automatically inferred Hugging Face train split as the official training split.

Use OMG-Lite's pinned dataset config and compact episode-cache materializer. Compact cache stores qpos, conditions, masks, captions and episode provenance, without redundant FK body positions/quaternions. See the OMG-Lite data guide for commands.

Release size and validation

Split Episodes Frames
train 636,865 160,786,490
val 79,975 20,036,877
test 81,341 20,330,291

All 61 frame shards (201,153,658 frames) passed independent validation. All 74 upstream file fingerprints were checked against the pinned official release. See provenance/validation_report.json, source_file_sha256.json, and the executed validator.

38D training normalization

normalization/g1_38d_multimodal_50hz_hist10_stats.json covers every training window: 102,352,253 windows and 10,230,609,902 valid frame observations. The stats use 100-frame windows, 10 history frames, canonical index 9 and stride 1. Their source content revision is the SHA256 of SHA256SUMS (data and metadata files). Do not replace an existing checkpoint's own normalization with these statistics.

The complete compact train/val/test cache passed index, provenance, dtype and shape validation. The direct reader and compact cache agree on 22 representative source/condition combinations. A 50M denoiser completed three real multimodal optimizer steps with frozen T5, the new stats, and the full compact train cache. See provenance/smoke_report.json.

Downloads last month
303