---
pretty_name: Placement Assets
license: other
task_categories:
  - robotics
tags:
  - robotics
  - mujoco
  - mjcf
  - assets
  - placement
---

# Asset Library

[中文说明](README_zh.md)

Unified robotics and simulation asset library.

This repository stores normalized reusable assets from RoboCasa, Discoverse, TinyForge, Hunter, and third-party asset sites. It is designed to be hosted as a Hugging Face Dataset.

## Layout

Raw source-preserved asset layout:

```text
assets/<source>/raw/<asset_type>/<category>/<asset_id>/
```

Example:

```text
assets/robocasa/raw/fixtures/dishwashers/Dishwasher031/
```

Derived asset layout:

```text
assets/<source>/derived/<asset_type>/<category>/<asset_id>/
```

Use `derived/` for generated or transformed outputs such as convex decomposition, collision meshes, cleaned meshes, and conversions.

For convex decomposition scaffolding:

```bash
python scripts/scaffold_convex_decomposition.py assets/robocasa/raw/fixtures/dishwashers/Dishwasher031
```

Collection-level imports are allowed when an upstream repository has many internal relative references. They still use the same layout under `raw/`:

```text
assets/<source>/raw/<asset_type>/<category>/<asset_id>/
```

Current example:

```text
assets/discoverse/raw/collections/models/discoverse_models/
```

Top-level repository structure:

```text
.
├── AGENTS.md
├── README.md
├── assets/
│   ├── discoverse/
│   ├── hunter/
│   ├── robocasa/
│   ├── third_party/
│   └── tinyforge/
├── templates/
├── docs/
│   ├── asset_readiness_classification.md
│   ├── convex_decomposition_spec.md
│   ├── external_source_catalog.md
│   ├── naming_spec.md
│   ├── placement_asset_notes.md
│   ├── primitive_collision_proxy_spec.md
│   ├── structure_mapping.md
│   ├── source_notes/
│   └── storage_spec.md
├── manifest/
│   ├── assets.jsonl
│   ├── licenses.yaml
│   └── sources.yaml
└── scripts/
    ├── run_convex_decomposition.py
    ├── run_primitive_collision_proxy.py
    ├── scaffold_convex_decomposition.py
    ├── download_from_hf.sh
    ├── upload_to_hf.sh
    └── validate_asset.py
```

Naming rule:

```text
<source>.<asset_type>.<category>.<upstream_id>
```

Example:

```text
robocasa.fixtures.dishwashers.Dishwasher031
```

## Current seed assets

- RoboCasa dishwasher fixtures: 25 MJCF assets under `assets/robocasa/raw/fixtures/dishwashers/`
- RoboCasa dishwasher support files:
  - fixture registry: `assets/robocasa/raw/fixtures/fixture_registry/dishwasher.yaml`
  - implementation reference: `docs/source_notes/robocasa/dishwasher_fixture.py`
- DISCOVERSE models collection: upstream `models/` tree under `assets/discoverse/raw/collections/models/discoverse_models/`
  - stored as a collection-level asset to preserve internal relative references
  - includes MJCF, URDF, mesh, and reference image files

Placement-related notes and candidate external links are tracked in:

```text
docs/asset_readiness_classification.md
docs/convex_decomposition_spec.md
docs/external_source_catalog.md
docs/placement_asset_notes.md
docs/primitive_collision_proxy_spec.md
docs/structure_mapping.md
docs/source_notes/third_party.md
```

Readiness is tracked separately from storage. A downloaded asset is not automatically production-ready:

- L0: validated for the target task, robot, policy, verifier, and replay pipeline
- L1: complete simulation asset in its original simulator
- L2: articulated structure exists but physics/task information is incomplete
- L3: movable parts are separated but joints are not reliable
- L4: complete static object
- L5: raw or incomplete static mesh

See `docs/asset_readiness_classification.md` for the full standard.

## Index

The global asset index is:

```text
manifest/assets.jsonl
```

Each asset directory also contains a local `metadata.yaml`.

## Validate

```bash
python scripts/validate_asset.py assets/robocasa/raw/fixtures/dishwashers/Dishwasher031
python scripts/validate_asset.py --all
```

Current expected result:

```text
validated_assets=26
```

## Upload to Hugging Face

Login first:

```bash
export HF_ENDPOINT=https://hf-mirror.com
hf auth login
```

Create the dataset repo if needed:

```bash
hf repo create <namespace>/<dataset_repo> --type dataset --private
```

Upload:

```bash
cd /path/to/asset-library
scripts/upload_to_hf.sh <namespace>/<dataset_repo>
```

## Download from Hugging Face

```bash
export HF_ENDPOINT=https://hf-mirror.com

scripts/download_from_hf.sh <namespace>/<dataset_repo> ./asset-library
```