src_root string | out_root string | total int64 | ok int64 | fail int64 | clips list |
|---|---|---|---|---|---|
/home/amdin/syj/motions/dec_release | /home/amdin/syj/motions/parc_ms | 26,577 | 26,396 | 181 | [{"src":"/home/amdin/syj/motions/dec_release/dec2024_iter_1_boxes/mdm_batch_000/BOXES_0_0_opt_dm_dm.(...TRUNCATED) |
hhtools PARC MS — terrain-aware humanoid motion clips
Motion clips in PARC MS layout for human-humanoid-tools (hhtools): per-clip folders with a PARC MSFileData pickle and a static terrain mesh. Suitable for meshmimic / interaction-mesh retargeting (parkour, climbing, box traversal, etc.).
| Clips | 26,396 |
| Total size | ~10.8 GB |
| Skeleton | 15-bone PARC humanoid (humanoid.xml topology) |
| Terrain | Heightfield + Wavefront OBJ per clip |
| Source | Converted from PARC dec_release MS pickles |
Directory layout
Each clip is a self-contained folder (OMOMO-style):
<clip_name>/
<clip_name>.pkl # PARC MSFileData: motion_data + terrain_data
<clip_name>_terrain.obj # static terrain mesh (Wavefront OBJ)
Example:
beyond_dash_vault_001_aug001_dm/
beyond_dash_vault_001_aug001_dm.pkl
beyond_dash_vault_001_aug001_dm_terrain.obj
A manifest.json at the repo root records the batch conversion from dec_release (26,577 source files → 26,396 successful clips).
Pickle format
Each .pkl is a PARC MS container with pickled inner blobs:
| Key | Content |
|---|---|
motion_data |
Root translation/rotation + per-joint rotations (root_pos, root_rot, joint_rot, …) |
terrain_data |
Terrain heightfield used for grounding / interaction mesh |
misc_data |
Optional metadata (may be absent) |
Poses are expressed on the 15-bone PARC humanoid (pelvis → torso/head/arms/hands/legs/feet). Joint order matches PARC humanoid.xml DFS body order; joint_rot has shape (T, 14, 4).
Download
# Hugging Face CLI
hf download YaojieShen/hhtools_parc_ms --repo-type dataset --local-dir ./parc_ms
Or with Python:
from huggingface_hub import snapshot_download
snapshot_download(
repo_id="YaojieShen/hhtools_parc_ms",
repo_type="dataset",
local_dir="./parc_ms",
)
Usage with hhtools
hhtools loads this layout via the built-in parc_ms dataset adapter:
git clone https://github.com/jaggerShen/human-humanoid-tools.git
cd human-humanoid-tools
uv sync --extra all
uv run hhtools web
In the Web UI, point the motion library at the downloaded folder. Clips appear under dataset parc_ms; retarget with interaction mesh (MPC-SQP) to preserve foot–terrain contact.
Programmatic load:
from pathlib import Path
from hhtools.io.datasets.parc_ms import ParcMsAdapter
adapter = ParcMsAdapter(root=Path("./parc_ms"))
clip_id = next(adapter.list_sequences())
motion = adapter.load_motion(clip_id)
print(motion.positions.shape, motion.terrain is not None)
Motion categories (approximate)
Clip names reflect the upstream PARC dec_release batches, including:
- Parkour / traversal —
beyond_*,climbing_*,running_*,strides_* - Box / obstacle interaction —
boxes_*,BOXES_* - Platform / terrain variants —
platform_*,*_dm,*_flipped_*,*_aug* - Teaser / demo scenes —
teaser_*,dec2024_*
Many clips are augmented or mirrored variants of base motions.
Provenance & license
- Upstream method & data: PARC — Physics-based Augmentation with Reinforcement learning for Character control · GitHub
- This release: MS pickles from PARC
dec_release, reorganized into hhtools-compatibleparc_msclip folders with terrain OBJ sidecars. - License: Apache-2.0 (same as human-humanoid-tools).
Please cite PARC when using the motions in research, and hhtools if you use this packaging or the retargeting toolchain.
@article{shoe2025parc,
title = {PARC: Physics-based Augmentation with Reinforcement learning for Character control},
author = {Shoe, Michael and others},
journal = {arXiv preprint arXiv:2505.04002},
year = {2025}
}
@software{human_humanoid_tools2026,
title = {human-humanoid-tools (hhtools): humanoid motion retargeting and dataset analysis},
author = {jaggerShen and hhtools contributors},
year = {2026},
url = {https://github.com/jaggerShen/human-humanoid-tools},
license = {Apache-2.0}
}
中文说明
本仓库提供 PARC MS 格式的地形交互人体动作片段,可直接用于 human-humanoid-tools (hhtools) 的 meshmimic / 交互网格重定向。
数据概览
| 项目 | 数值 |
|---|---|
| 片段数 | 26,396 |
| 总大小 | 约 10.8 GB |
| 骨骼 | 15 骨 PARC 人形(与 PARC humanoid.xml 一致) |
| 地形 | 每个 clip 含 heightfield(pkl 内)+ OBJ 网格 |
目录结构
<clip_name>/
<clip_name>.pkl # 动作 + 地形(PARC MSFileData)
<clip_name>_terrain.obj # 静态地形网格
下载
hf download YaojieShen/hhtools_parc_ms --repo-type dataset --local-dir ./parc_ms
在 hhtools 中使用
下载后启动 Web UI,将动作库路径指向 parc_ms 文件夹即可;数据集类型选择 **parc_ms**,使用交互网格重定向保留脚–地形接触。
来源与引用
数据源自 PARC 项目的 dec_release MS pickle,经批量转换为本仓库的 per-clip 文件夹布局。使用请注明 PARC 论文及 hhtools 工具链(见上方 BibTeX)。
Contact
Issues about loading / retargeting: human-humanoid-tools Issues
Dataset hosting: YaojieShen/hhtools_parc_ms
- Downloads last month
- 10,190