---
language:
- en
- zh
tags:
- robotics
- motion
- unitree-g1
- motion-retargeting
pretty_name: MotionMamba2.0 G1 Mixed Motion CSV
size_categories:
- 100K<n<1M
configs:
- config_name: default
  data_files:
  - split: train
    path: metadata/train.jsonl
  - split: validation
    path: metadata/val.jsonl
  - split: test
    path: metadata/test.jsonl
---

# MotionMamba2.0 G1 Mixed Motion CSV

Accepted G1 motion exports from BONES Seed, HiPHI and HumanML3D. BONES uses
official G1 CSV input; HiPHI and HumanML3D are retargeted to G1. All packaged
motions are sampled at **25 FPS**. The package contains **179,648 motions**,
**660.98 hours**, and their available text descriptions.

| Split | Motions | Hours | Uncompressed CSV bytes |
|---|---:|---:|---:|
| Train | 156,819 | 588.922356 | 24,775,677,931 |
| Validation | 9,998 | 41.175389 | 1,766,247,317 |
| Test | 12,831 | 30.887122 | 1,274,131,579 |
| Total | 179,648 | 660.984867 | 27,816,056,827 |

## Files

Files are stored directly in the repository root:

- `data/{train,val,test}/{source}-NNNN.tar.gz`: sharded CSVs and original motion
  JSON sidecars. Shards contain only one source and one split. Approximately
  512 MiB of uncompressed CSV data is assigned to each full shard.
- `metadata/{train,val,test}.jsonl`: portable sample metadata, captions, splits,
  archive locations, extracted file locations, and original file SHA-256s.
- `dataset_info.json`: measured counts, sizes, source-manifest hashes and shard
  checksums.

Archive members retain readable source sample names, for example
`motions/train/bones/Idle_Left_001__A017.csv`. Characters that cannot be used
safely in file paths are percent-encoded. Each metadata row gives the corresponding `sample_id`,
`archive_path`, `csv_path` and `sidecar_path`. Paths are relative to the release
directory; metadata does not depend on the original server's directory layout.
The original sidecars retain their historical provenance fields.

The package does not include intermediate solver caches, videos, original
BVH/SMPL inputs, `qpos.npy`, or the precomputed 414-dimensional `features.npy`.

## Download and extract

This repository is public; no login is required for downloading.

```python
from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="AIGeeksGroup/MotionMamba2.0",
    repo_type="dataset",
    local_dir="MotionMamba2.0",
)
```

To download through HF-Mirror, set `HF_ENDPOINT=https://hf-mirror.com` before
running the downloader. Uploads are made to the official Hugging Face endpoint.

Extract shards inside the release directory so that their paths agree with the
metadata:

```bash
cd MotionMamba2.0
for shard in data/*/*.tar.gz; do tar -xzf "$shard"; done
```

## CSV representation

Each CSV has **36 columns**: frame index, three root translations, three root
Euler angles, and 29 joint angles. Each CSV contains its ordered column names
in the header row; units and rotation conventions are described below.

- Frame indices start at zero; time is `Frame / 25` seconds.
- Root translation is in **centimetres**; multiply by `0.01` for metres.
- Root Euler angles and articulated joint angles are in **degrees**; convert
  with `numpy.deg2rad` when a consumer expects radians.
- Root Euler convention: SciPy lowercase extrinsic active `xyz`.
- CSV root Euler coordinates are distinct from quaternion-based MuJoCo qpos.

```python
import json
from pathlib import Path
import numpy as np

release = Path("MotionMamba2.0")
with (release / "metadata/train.jsonl").open() as f:
    sample = json.loads(next(f))

motion = np.loadtxt(release / sample["csv_path"], delimiter=",", skiprows=1,
                    ndmin=2)
root_translation_m = motion[:, 1:4] * 0.01
root_euler_xyz_rad = np.deg2rad(motion[:, 4:7])
joint_angles_rad = np.deg2rad(motion[:, 7:])
captions = sample["captions"]
```