rtgtlab7291/robokit-assets

robokit assets

Robot descriptions, body models, motion datasets, and benchmarks for robokit. They live in the HuggingFace dataset jianglong-org/robokit-internal.

Layout

robots/
  robot_description/   # URDFs + meshes: arms/ end_effectors/ humanoids/ assembly/
  collision_spheres/   # per-robot collision-sphere YAMLs
  contact_points/      # per-hand contact-point JSONs
  self_collision/  contact_avoid_points/
body_models/           # mano/ smplx/  (.npz)
motions/               # AMASS-style datasets (ACCAD, SFU, DanceDB, ...) + dex-retargeting/
benchmarks/            # dexycb/
contact_retarget/      # self-contained contact-aware bundle (g1 model + spheres + demo meshes)
objects/               # grasp/object datasets

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.

Consuming

robokit reads assets through robokit.assets. Each named asset is a small module that exposes static path symbols. Importing the module downloads that asset into ~/.robokit/cache/; the symbols are then ready-to-use Paths:

from robokit.assets.robots import unitree_g1   # downloads g1 on import

unitree_g1.URDF_PATH              # a Path
unitree_g1.COLLISION_SPHERE_PATH

For a path that is not known statically, fetch(patterns) downloads a subtree and returns the snapshot dir: fetch(["motions/SFU/**"]) / "motions/SFU/0005/x.npz".

The default revision is main. Pin a commit with the ROBOKIT_ASSETS_REVISION environment variable or fetch(revision=...).

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.link() clones it into ~/.robokit/hf_repo and symlinks <cwd>/robokit_assets to it, so robokit_assets/ is a normal git checkout:

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

Or clone it directly with git clone https://huggingface.co/datasets/jianglong-org/robokit-internal.

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/<kind>/<name>.py following an existing one (e.g. robots/franka_panda.py).