Document mixed G1 CSV release; data upload follows
Browse files
README.md
ADDED
|
@@ -0,0 +1,139 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
language:
|
| 3 |
+
- en
|
| 4 |
+
- zh
|
| 5 |
+
tags:
|
| 6 |
+
- robotics
|
| 7 |
+
- motion
|
| 8 |
+
- unitree-g1
|
| 9 |
+
- motion-retargeting
|
| 10 |
+
pretty_name: MotionMamba2.0 G1 Mixed Motion CSV
|
| 11 |
+
size_categories:
|
| 12 |
+
- 100K<n<1M
|
| 13 |
+
configs:
|
| 14 |
+
- config_name: default
|
| 15 |
+
data_files:
|
| 16 |
+
- split: train
|
| 17 |
+
path: metadata/train.jsonl
|
| 18 |
+
- split: validation
|
| 19 |
+
path: metadata/val.jsonl
|
| 20 |
+
- split: test
|
| 21 |
+
path: metadata/test.jsonl
|
| 22 |
+
---
|
| 23 |
+
|
| 24 |
+
# MotionMamba2.0 G1 Mixed Motion CSV
|
| 25 |
+
|
| 26 |
+
Accepted G1 motion exports from BONES Seed, HiPHI and HumanML3D. BONES uses
|
| 27 |
+
official G1 CSV input; HiPHI and HumanML3D are retargeted to G1. All packaged
|
| 28 |
+
motions are sampled at **25 FPS**. The package contains **179,648 motions**,
|
| 29 |
+
**660.98 hours**, and their available text descriptions.
|
| 30 |
+
|
| 31 |
+
| Split | Motions | Hours | Uncompressed CSV bytes |
|
| 32 |
+
|---|---:|---:|---:|
|
| 33 |
+
| Train | 156,819 | 588.922356 | 24,775,677,931 |
|
| 34 |
+
| Validation | 9,998 | 41.175389 | 1,766,247,317 |
|
| 35 |
+
| Test | 12,831 | 30.887122 | 1,274,131,579 |
|
| 36 |
+
| Total | 179,648 | 660.984867 | 27,816,056,827 |
|
| 37 |
+
|
| 38 |
+
## Files
|
| 39 |
+
|
| 40 |
+
Files are stored directly in the repository root:
|
| 41 |
+
|
| 42 |
+
- `data/{train,val,test}/{source}-NNNN.tar.gz`: sharded CSVs and original motion
|
| 43 |
+
JSON sidecars. Shards contain only one source and one split. Approximately
|
| 44 |
+
512 MiB of uncompressed CSV data is assigned to each full shard.
|
| 45 |
+
- `metadata/{train,val,test}.jsonl`: portable sample metadata, captions, splits,
|
| 46 |
+
archive locations, extracted file locations, and original file SHA-256s.
|
| 47 |
+
- `contracts/`: original CSV, robot, skeleton and feature schema references.
|
| 48 |
+
- `dataset_info.json`: measured counts, sizes, source-manifest hashes and shard
|
| 49 |
+
checksums.
|
| 50 |
+
|
| 51 |
+
Archive members retain readable source sample names, for example
|
| 52 |
+
`motions/train/bones/Idle_Left_001__A017.csv`. Characters that cannot be used
|
| 53 |
+
safely in file paths are percent-encoded. Each metadata row gives the corresponding `sample_id`,
|
| 54 |
+
`archive_path`, `csv_path` and `sidecar_path`. Paths are relative to the release
|
| 55 |
+
directory; metadata does not depend on the original server's directory layout.
|
| 56 |
+
The original sidecars retain their historical provenance fields.
|
| 57 |
+
|
| 58 |
+
The package does not include intermediate solver caches, videos, original
|
| 59 |
+
BVH/SMPL inputs, `qpos.npy`, or the precomputed 414-dimensional `features.npy`.
|
| 60 |
+
|
| 61 |
+
## Download and extract
|
| 62 |
+
|
| 63 |
+
For this private repository, use an account/token with read permission.
|
| 64 |
+
|
| 65 |
+
```python
|
| 66 |
+
from huggingface_hub import snapshot_download
|
| 67 |
+
|
| 68 |
+
snapshot_download(
|
| 69 |
+
repo_id="AIGeeksGroup/MotionMamba2.0",
|
| 70 |
+
repo_type="dataset",
|
| 71 |
+
local_dir="MotionMamba2.0",
|
| 72 |
+
)
|
| 73 |
+
```
|
| 74 |
+
|
| 75 |
+
To download through HF-Mirror, set `HF_ENDPOINT=https://hf-mirror.com` before
|
| 76 |
+
running the downloader. Uploads are made to the official Hugging Face endpoint.
|
| 77 |
+
|
| 78 |
+
Extract shards inside the release directory so that their paths agree with the
|
| 79 |
+
metadata:
|
| 80 |
+
|
| 81 |
+
```bash
|
| 82 |
+
cd MotionMamba2.0
|
| 83 |
+
for shard in data/*/*.tar.gz; do tar -xzf "$shard"; done
|
| 84 |
+
```
|
| 85 |
+
|
| 86 |
+
## CSV representation
|
| 87 |
+
|
| 88 |
+
Each CSV has **36 columns**: frame index, three root translations, three root
|
| 89 |
+
Euler angles, and 29 joint angles. The ordered column names and conventions are
|
| 90 |
+
specified in `contracts/bones_g1_csv.json`.
|
| 91 |
+
|
| 92 |
+
- Frame indices start at zero; time is `Frame / 25` seconds.
|
| 93 |
+
- Root translation is in **centimetres**; multiply by `0.01` for metres.
|
| 94 |
+
- Root Euler angles and articulated joint angles are in **degrees**; convert
|
| 95 |
+
with `numpy.deg2rad` when a consumer expects radians.
|
| 96 |
+
- Root Euler convention: SciPy lowercase extrinsic active `xyz`.
|
| 97 |
+
- CSV root Euler coordinates are distinct from quaternion-based MuJoCo qpos.
|
| 98 |
+
|
| 99 |
+
```python
|
| 100 |
+
import json
|
| 101 |
+
from pathlib import Path
|
| 102 |
+
import numpy as np
|
| 103 |
+
|
| 104 |
+
release = Path("MotionMamba2.0")
|
| 105 |
+
with (release / "metadata/train.jsonl").open() as f:
|
| 106 |
+
sample = json.loads(next(f))
|
| 107 |
+
|
| 108 |
+
motion = np.loadtxt(release / sample["csv_path"], delimiter=",", skiprows=1,
|
| 109 |
+
ndmin=2)
|
| 110 |
+
root_translation_m = motion[:, 1:4] * 0.01
|
| 111 |
+
root_euler_xyz_rad = np.deg2rad(motion[:, 4:7])
|
| 112 |
+
joint_angles_rad = np.deg2rad(motion[:, 7:])
|
| 113 |
+
captions = sample["captions"]
|
| 114 |
+
```
|
| 115 |
+
|
| 116 |
+
## Splits and provenance
|
| 117 |
+
|
| 118 |
+
These are the existing accepted release splits, not a new random split of the
|
| 119 |
+
training subset. HumanML3D retains official split assignments. BONES uses the
|
| 120 |
+
project's actor-disjoint protocol; HiPHI uses the existing project grouping.
|
| 121 |
+
Mirror/crop/source-family constraints and existing exact-output duplicate
|
| 122 |
+
quarantines are retained. Frozen calibration-heldout BONES entries are excluded
|
| 123 |
+
from these default manifests.
|
| 124 |
+
|
| 125 |
+
Every packaged CSV and original motion sidecar is checked against its hash in
|
| 126 |
+
the accepted source manifest during packaging. The CSV bytes are not modified.
|
| 127 |
+
Captions and source labels are copied from existing metadata without new
|
| 128 |
+
annotation or inferred measurements.
|
| 129 |
+
|
| 130 |
+
## Scope and source terms
|
| 131 |
+
|
| 132 |
+
This is an **offline kinematic export**. Dynamics validation is recorded as
|
| 133 |
+
`NOT_RUN`; this upload does not certify controller tracking or change any
|
| 134 |
+
historical frozen acceptance result. The BONES official-pool geometry metrics
|
| 135 |
+
remain diagnostic under the user-approved admission policy.
|
| 136 |
+
|
| 137 |
+
The original BONES Seed, HiPHI and HumanML3D terms continue to apply to their
|
| 138 |
+
respective contributions. This packaging does not assign a new license to
|
| 139 |
+
third-party source data.
|