---
license: other
tags:
  - robotics
  - urdf
  - retargeting
---

# robokit assets

Robot descriptions and datasets for robokit. They live in the
Hugging Face dataset [`openrobokit/robokit`](https://huggingface.co/datasets/openrobokit/robokit).

## Layout

```
robots/
  robot_description/   # URDFs + meshes: arms/ end_effectors/ humanoids/ assembly/
  robot_mj_description/ # MJCFs + meshes
  collision_spheres/   # per-robot collision-sphere YAMLs
  contact_points/      # per-hand contact-point JSONs
  self_collision/  contact_avoid_points/
datasets/              # one lowercase folder per source
  omomo/               # clips, metadata, objects/
  parahome/            # seq/, smplx_seq/, objects/
  dexycb/              # poses, contacts, object models
  ...
calibration/
```

Robot files are type-major: one robot's URDF, collision spheres, and contact points live under
different top-level `robots/` subdirs keyed by the same robot name.

Each dataset keeps its motions and object meshes together under `datasets/<name>/`.
MANO/SMPL-X body models are not included; download them from their official sites.

## Consuming

robokit reads assets through `robokit.assets`. Each named asset is a small module that exposes
static path symbols or access functions. Requested assets are cached under `~/.robokit/hf_cache/`
and exposed at `~/.robokit/assets/<pinned-commit>/`; the
symbols are then ready-to-use `Path`s:

```python
from robokit.assets.robots.humanoids import unitree_g1
from robokit.assets.datasets import omomo

unitree_g1.URDF_PATH              # a Path
unitree_g1.COLLISION_SPHERE_PATH
omomo.clip_path("sub3_largebox_003")
omomo.mesh_path("largebox")
```

For a path that is not known statically, `fetch(patterns)` downloads a subtree and returns the
asset root: `fetch(["datasets/sfu/**"]) / "datasets/sfu/0005/x.npz"`.

RoboKit pins the dataset commit in `src/robokit/assets/__init__.py`; runtime does not follow `main`.

## Authoring

The dataset is a plain git repo whose binary assets are stored with **git-LFS**, so install
`git-lfs` (`git lfs install`) before cloning or pushing. `robokit.assets.checkout()` clones it into
`~/.robokit/hf_repo` and symlinks `<cwd>/robokit_assets` to it, so `robokit_assets/` is a normal git checkout:

```shell
cd robokit_assets && git add -A && git commit -m "add <asset>" && git push
```

Or clone it directly with `git clone https://huggingface.co/datasets/openrobokit/robokit`.

Every mesh/texture (`.obj .glb .stl .dae .png .npz ...`) is LFS-tracked via `.gitattributes`. When
adding a new binary extension, run `git lfs track "*.<ext>"` first so it is not committed as a raw blob.

Keep the type-major layout above. To expose the new asset as a static symbol, add a module under
`src/robokit/assets/robots/` or `src/robokit/assets/datasets/`, and update RoboKit's pinned revision.

Assets are under their own licenses. See [license files and outstanding gaps](LICENSES.md).