[codex] Self-host public model assets

README.md

@@ -66,7 +66,7 @@ work can use `skip_vertices=True` to avoid preparing rest vertices.
 - Anatomicals: SKEL, MyoFullBody
 - Heads: FLAME
 - Hands: MANO
-- Robots: BrainCo, G1
+- Robots: BrainCo, G1, SmplHumanoid
 
 See the [model docs](https://abcamiletto.github.io/body-models/#supported-models)
 for setup, supported backends, inputs, and model-specific behavior.

docs/index.md

@@ -56,6 +56,7 @@ uv add "body-models[jax]"
 | --- | --- | --- |
 | [BrainCo](models/brainco.md) | BrainCo Revo 2 robotic hand | auto-download |
 | [G1](models/g1.md) | Unitree G1 rigid links | auto-download |
+| [SmplHumanoid](models/smpl-humanoid.md) | SMPL-compatible humanoid MJCF variants | auto-download |
 
 ## Common Usage
 

docs/models/anny.md

@@ -4,7 +4,9 @@ ANNY is a phenotype-driven body model with configurable rig and topology variant
 
 ## Setup
 
-ANNY downloads automatically on first use. To prefetch and save the path:
+ANNY downloads automatically on first use from `https://huggingface.co/abcamiletto/body-models`.
+The Hugging Face repo records the original ANNY Apache 2.0 and MPFB2 CC0 provenance. To prefetch
+and save the path:
 
 ```bash
 # Download the ANNY assets and store their path in the body-models config.

docs/models/brainco.md

@@ -4,7 +4,7 @@ BrainCo is a rigid articulated model of the BrainCo Revo 2 robotic hand using th
 
 ## Setup
 
-BrainCo downloads automatically on first use. To prefetch and save the path:
+BrainCo downloads from the public [`abcamiletto/body-models`](https://huggingface.co/abcamiletto/body-models) Hugging Face repository on first use. To prefetch and save the path:
 
 ```bash
 # Download the BrainCo MuJoCo XML and STL assets.
@@ -13,6 +13,8 @@ body-models download brainco
 
 When passed manually, `model_path` should contain `left.xml`, `right.xml`, and `meshes/{left,right}/*.STL`.
 
+The original BrainCo Revo2 description license is included with the hosted assets.
+
 ## Usage
 
 ```python

docs/models/g1.md

@@ -4,7 +4,9 @@ G1 is a rigid articulated Unitree G1 model with STL link meshes attached to the
 
 ## Setup
 
-G1 downloads automatically on first use. To prefetch and save the path:
+G1 downloads automatically on first use from `https://huggingface.co/abcamiletto/body-models`.
+The Hugging Face repo records the original GR00T-WholeBodyControl / LeRobot provenance. To prefetch
+and save the path:
 
 ```bash
 # Download the Unitree G1 XML and link meshes.

docs/models/garment-measurements.md

@@ -4,8 +4,9 @@ GarmentMeasurements is a PCA body model with an FBX-derived skeleton and skinnin
 
 ## Setup
 
-GarmentMeasurements downloads its preprocessed asset from
-`https://huggingface.co/datasets/abcamiletto/body-models-assets` on first use. To prefetch and save the path:
+GarmentMeasurements downloads its preprocessed assets from
+`https://huggingface.co/abcamiletto/body-models` on first use. The Hugging Face repo records
+the original SOMA-X Apache 2.0 provenance for the source asset. To prefetch and save the path:
 
 ```bash
 # Download the preprocessed GarmentMeasurements body-model asset.

docs/models/mhr.md

@@ -4,13 +4,17 @@ MHR is an expressive full-body model with neural pose correctives.
 
 ## Setup
 
-MHR downloads automatically on first use. To prefetch and save the path:
+MHR downloads from the public [`abcamiletto/body-models`](https://huggingface.co/abcamiletto/body-models) Hugging Face repository on first use. The hosted package keeps the original MHR checkpoint for the default LOD 1 path and adds preprocessed FBX-derived mesh assets for LODs 0 through 6.
+
+To prefetch and save the path:
 
 ```bash
 # Download the MHR assets and store their path in the body-models config.
 body-models download mhr
 ```
 
+The original MHR license is included with the hosted assets.
+
 ## API
 
 ::: body_models.bodies.mhr.numpy.MHR

docs/models/myofullbody.md

@@ -4,7 +4,8 @@ MyoFullBody is a MuJoCo-derived musculoskeletal full-body model from `amathislab
 
 ## Setup
 
-MyoFullBody downloads automatically on first use. To prefetch and save the path:
+MyoFullBody downloads automatically on first use from `https://huggingface.co/abcamiletto/body-models`.
+The Hugging Face repo records the original MuscleMimic Apache 2.0 provenance. To prefetch and save the path:
 
 ```bash
 # Download the MyoFullBody MJCF and referenced mesh assets.

docs/models/smpl-humanoid.md

@@ -0,0 +1,18 @@
+# SmplHumanoid
+
+SmplHumanoid is a rigid articulated humanoid model loaded from SMPL-compatible MJCF XML variants.
+
+## Setup
+
+SmplHumanoid downloads its XML assets from the public [`abcamiletto/body-models`](https://huggingface.co/abcamiletto/body-models) Hugging Face repository. To prefetch and save the hosted path:
+
+```bash
+# Download the SmplHumanoid MJCF XML assets.
+body-models download smpl-humanoid
+```
+
+The hosted folder includes license/provenance notes for the XML variants.
+
+## API
+
+::: body_models.robots.smpl_humanoid.numpy.SmplHumanoid

docs/models/soma.md

@@ -4,7 +4,8 @@ SOMA provides a native implementation for SOMA-X assets with identity, pose, and
 
 ## Setup
 
-SOMA downloads automatically on first use. To prefetch and save the path:
+SOMA downloads automatically on first use from `https://huggingface.co/abcamiletto/body-models`.
+The Hugging Face repo records the original SOMA-X Apache 2.0 provenance. To prefetch and save the path:
 
 ```bash
 # Download the SOMA-X assets used by the native SOMA implementation.

src/body_models/__init__.py

@@ -67,6 +67,7 @@ def main() -> None:
     from .config import CONFIG_FILE, MODELS, get_model_path, set_model_path, unset_model_path
     from .robots.brainco.io import download_model as download_brainco_model
     from .robots.g1.io import download_model as download_g1_model
+    from .robots.smpl_humanoid.io import download_model as download_smpl_humanoid_model
     from .skeletons.myofullbody.io import download_model as download_myofullbody_model
 
     Model = Literal[
@@ -88,6 +89,7 @@ def main() -> None:
         "flame",
         "brainco",
         "g1",
+        "smpl-humanoid",
         "soma",
         "garment-measurements",
         "myofullbody",
@@ -130,6 +132,7 @@ def download(
                 "brainco",
                 "mhr",
                 "g1",
+                "smpl-humanoid",
                 "soma",
                 "garment-measurements",
                 "myofullbody",
@@ -234,6 +237,11 @@ def download(
             set_model_path("g1", str(path))
             print(f"Set g1 = {path}")
 
+        if model in ("smpl-humanoid", "all"):
+            path = download_smpl_humanoid_model()
+            set_model_path("smpl-humanoid", str(path))
+            print(f"Set smpl-humanoid = {path}")
+
         if model in ("soma", "all"):
             path = download_soma_model()
             set_model_path("soma", str(path))

src/body_models/bodies/anny/io.py

@@ -11,14 +11,14 @@
 from nanomanifold import SO3
 
 from body_models import config
-from body_models.cache import download_and_extract, get_cache_dir
+from body_models.cache import HF_MODEL_BASE_URL, download_and_extract, get_cache_dir
 
 PathLike = Path | str
 Array = Any
 
 Front = tuple[list[int], list[int]]  # One FK depth level: (joint_indices, parent_indices).
 
-ANNY_URL = "https://github.com/naver/anny/archive/refs/heads/main.zip"
+ANNY_URL = f"{HF_MODEL_BASE_URL}/anny/assets.zip"
 
 PHENOTYPE_VARIATIONS = {
     "race": ["african", "asian", "caucasian"],
@@ -83,7 +83,7 @@ def get_model_path(model_path: PathLike | None = None) -> Path:
 def download_model() -> Path:
     cache_dir = get_cache_dir() / "anny"
     print(f"Downloading ANNY model to {cache_dir}...")
-    download_and_extract(url=ANNY_URL, dest=cache_dir, extract_subdir="anny-main/src/anny/")
+    download_and_extract(url=ANNY_URL, dest=cache_dir)
     print("Done")
     return cache_dir
 

src/body_models/bodies/garment_measurements/io.py

@@ -10,15 +10,15 @@
 from jaxtyping import Float, Int
 
 from body_models import config
-from body_models.cache import HF_DATASET_BASE_URL, download_file, get_cache_dir
+from body_models.cache import HF_MODEL_BASE_URL, download_and_extract, get_cache_dir
 
 PathLike = Path | str
 Array = Any
 
 Front = tuple[list[int], list[int]]  # One FK depth level: (joint_indices, parent_indices).
 
 PREPROCESSED_FILENAME = "garment_measurements.npz"
-GARMENT_MEASUREMENTS_URL = f"{HF_DATASET_BASE_URL}/garment_measurements/{PREPROCESSED_FILENAME}"
+GARMENT_MEASUREMENTS_URL = f"{HF_MODEL_BASE_URL}/garment_measurements/assets.zip"
 
 
 @dataclass(frozen=True)
@@ -65,7 +65,7 @@ def download_model() -> Path:
     """Download preprocessed GarmentMeasurements data assets."""
     cache_dir = get_cache_dir() / "garment_measurements"
     print(f"Downloading GarmentMeasurements model to {cache_dir}...")
-    download_file(GARMENT_MEASUREMENTS_URL, cache_dir / PREPROCESSED_FILENAME)
+    download_and_extract(GARMENT_MEASUREMENTS_URL, cache_dir)
     print("Done")
     return cache_dir
 

src/body_models/bodies/mhr/io.py

@@ -11,7 +11,7 @@
 
 from body_models import config
 from body_models.common import simplify_mesh
-from body_models.cache import download_and_extract, get_cache_dir
+from body_models.cache import HF_MODEL_BASE_URL, download_and_extract, get_cache_dir
 
 PathLike = Path | str
 
@@ -28,7 +28,8 @@
     "load_pose_correctives",
 ]
 
-MHR_URL = "https://github.com/facebookresearch/MHR/releases/download/v1.0.0/assets.zip"
+MHR_URL = f"{HF_MODEL_BASE_URL}/mhr/assets.zip"
+SUPPORTED_LODS = tuple(range(7))
 
 
 @dataclass(frozen=True)
@@ -71,35 +72,40 @@ def get_model_path(model_path: PathLike | None = None) -> Path:
         return validate_path(model_path)
 
     cache_path = get_cache_dir() / "mhr"
-    if (cache_path / "mhr_model.pt").exists():
+    if _has_hosted_assets(cache_path):
         return cache_path
 
     return download_model()
 
 
 def download_model() -> Path:
-    """Download MHR model from GitHub releases."""
+    """Download MHR model assets."""
     cache_dir = get_cache_dir() / "mhr"
     print(f"Downloading MHR model to {cache_dir}...")
-    download_and_extract(url=MHR_URL, dest=cache_dir, extract_subdir="assets/")
+    download_and_extract(url=MHR_URL, dest=cache_dir)
     print("Done")
     return cache_dir
 
 
 def load_model_data(asset_dir: Path, *, lod: int = 1, simplify: float = 1.0) -> MhrWeights:
     if simplify < 1.0:
         raise ValueError("simplify must be >= 1.0")
-    if lod != 1:
-        raise ValueError("MHR lod values other than 1 are not supported.")
+    if lod not in SUPPORTED_LODS:
+        raise ValueError(f"MHR lod must be one of {SUPPORTED_LODS}, got {lod}")
 
-    data = _load_raw_model_data(asset_dir)
+    shared_data = _load_raw_model_data(asset_dir)
+    data = shared_data if lod == 1 else _load_preprocessed_lod_data(asset_dir, lod, shared_data)
     base_vertices = data["base_vertices"]
     blendshape_dirs = data["blendshape_dirs"]
     skin_weights = data["skin_weights"]
     skin_indices = data["skin_indices"].astype(np.int64)
     faces = data["faces"].astype(np.int64)
     corrective_weights = load_pose_correctives_weights(asset_dir, lod)
     corrective_W2 = corrective_weights["W2"]
+    if corrective_W2.shape[0] != len(base_vertices) * 3:
+        raise ValueError(
+            f"MHR lod{lod} corrective W2 has {corrective_W2.shape[0]} rows, expected {len(base_vertices) * 3}"
+        )
 
     if simplify > 1.0:
         target_faces = int(len(faces) / simplify)
@@ -169,6 +175,43 @@ def _load_raw_model_data(asset_dir: Path) -> dict[str, Any]:
     }
 
 
+def _load_preprocessed_lod_data(asset_dir: Path, lod: int, shared_data: dict[str, Any]) -> dict[str, Any]:
+    path = asset_dir / f"mhr_lod{lod}.npz"
+    with np.load(path, allow_pickle=False) as asset:
+        joint_names = [str(name) for name in asset["skin_joint_names"].tolist()]
+        checkpoint_joint_index = {name: index for index, name in enumerate(shared_data["joint_names"])}
+        missing = sorted(name for name in joint_names if name not in checkpoint_joint_index)
+        if missing:
+            raise ValueError(f"{path} references joints missing from mhr_model.pt: {missing}")
+
+        skin_joint_indices = np.asarray(asset["skin_joint_indices"], dtype=np.int64)
+        mapped_joint_indices = np.asarray([checkpoint_joint_index[joint_names[index]] for index in skin_joint_indices])
+        base_vertices = np.asarray(asset["base_vertices"], dtype=np.float32)
+        skin_indices, skin_weights = _build_dense_skinning(
+            asset["skin_vertex_indices"],
+            mapped_joint_indices,
+            asset["skin_weights"],
+            len(base_vertices),
+        )
+
+        return shared_data | {
+            "base_vertices": base_vertices,
+            "blendshape_dirs": np.asarray(asset["blendshape_dirs"], dtype=np.float32),
+            "skin_weights": skin_weights,
+            "skin_indices": skin_indices,
+            "faces": np.asarray(asset["faces"], dtype=np.int64),
+        }
+
+
+def _has_hosted_assets(model_path: Path) -> bool:
+    asset_names = [
+        "mhr_model.pt",
+        "corrective_activation.npz",
+        *(f"mhr_lod{lod}.npz" for lod in SUPPORTED_LODS),
+    ]
+    return all((model_path / name).is_file() for name in asset_names)
+
+
 def _get_attr(obj: Any, path: str) -> Any:
     cur = obj
     for part in path.split("."):