docs/convex_decomposition_spec.md

Convex Decomposition Storage Specification

Convex decomposition outputs are derived assets. They must not overwrite raw meshes or upstream XML files.

Path layout

Store convex decomposition results under:

assets/<source>/derived/convex_decompositions/<category>/<asset_id>/

Example for a RoboCasa dishwasher:

assets/robocasa/derived/convex_decompositions/dishwashers/Dishwasher031_vhacd_v1/
├── metadata.yaml
├── source_refs.yaml
├── meshes/
│   ├── door/
│   │   ├── part_000.obj
│   │   ├── part_001.obj
│   │   └── ...
│   ├── rack0/
│   └── rack1/
├── mjcf/
│   ├── convex_assets_include.xml
│   ├── convex_geoms_include.xml
│   └── convex_collision_include.xml
└── logs/
    └── decomposition.json

Example for a DISCOVERSE object:

assets/discoverse/derived/convex_decompositions/objects/cup_vhacd_v1/
├── metadata.yaml
├── source_refs.yaml
├── meshes/
│   └── cup/
│       ├── part_000.obj
│       └── part_001.obj
├── mjcf/
│   ├── convex_assets_include.xml
│   ├── convex_geoms_include.xml
│   └── convex_collision_include.xml
└── logs/
    └── decomposition.json

Required files

Scaffold command

Use the helper script to create the derived asset skeleton:

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

This creates a new directory under assets/robocasa/derived/convex_decompositions/... with placeholder metadata, source references, logs, and MJCF include files. Fill in the generated meshes and replace placeholders before adding the asset to manifest/assets.jsonl.

Run CoACD decomposition

Use scripts/run_convex_decomposition.py when you want to actually generate convex parts and MuJoCo include files:

pip install trimesh coacd

python scripts/run_convex_decomposition.py \
  assets/robocasa/raw/fixtures/dishwashers/Dishwasher031 \
  --mesh visuals/door.obj \
  --mesh visuals/rack0.obj \
  --variant coacd_v1 \
  --register-manifest

If --mesh is omitted, the script searches common mesh locations under the raw asset directory:

visuals/*.obj
visuals/*.stl
meshes/*.obj
meshes/*.stl

The script writes generated pieces to meshes/<mesh_name>/part_000.obj, records hashes and parameters in logs/decomposition.json, and creates MJCF include files under mjcf/.

For each generated convex part, the script also records size and complexity statistics in both logs/decomposition.json and logs/parts_summary.csv:

bbox_extents        # x/y/z axis-aligned bounding-box dimensions
bbox_center         # axis-aligned bounding-box center
bbox_volume
mesh_volume
surface_area
num_vertices
num_faces

Use logs/parts_summary.csv to find parts that are too small, too large, or unexpectedly complex before using them as collision geometry.

Decomposition granularity controls

CoACD does not directly support a target physical bbox size such as "each part should be about 5 cm". Part size is controlled indirectly through decomposition granularity parameters:

--threshold              Lower value -> stricter convex approximation -> usually more/smaller parts.
--max-convex-hull        Caps the number of convex hulls. Lower cap -> fewer/larger parts.
--no-merge               Disables hull merge. Usually keeps more/smaller parts.
--mcts-max-depth         Higher value allows deeper recursive splitting.
--mcts-nodes
--mcts-iterations        More search can find finer/better splits, with higher runtime.
--preprocess-resolution
--resolution             Higher values preserve more geometric detail before decomposition.
--max-ch-vertex          Controls complexity of each convex hull, not physical size directly.

Practical presets:

# Coarser collision: fewer/larger parts, faster simulation
python scripts/run_convex_decomposition.py <raw_asset_dir> \
  --mesh <mesh.obj> \
  --threshold 0.08 \
  --max-convex-hull 8

# Finer collision: more/smaller parts, more accurate but heavier
python scripts/run_convex_decomposition.py <raw_asset_dir> \
  --mesh <mesh.obj> \
  --threshold 0.02 \
  --max-convex-hull 32 \
  --no-merge \
  --mcts-max-depth 4

After generation, inspect logs/parts_summary.csv to decide whether the resulting part sizes are acceptable for the target task.

Full script usage

Print the current CLI help:

python scripts/run_convex_decomposition.py --help

General form:

python scripts/run_convex_decomposition.py <raw_asset_dir> \
  --mesh <mesh_path_relative_to_raw_asset_dir> \
  --variant <variant_name> \
  [CoACD options] \
  [MJCF options] \
  [repository options]

Required argument

Argument	Meaning
raw_asset_dir	Raw asset directory under assets/<source>/raw/<asset_type>/<category>/<asset_id>/.

Input selection

Option	Default	Meaning
--mesh <path>	none	Mesh to decompose. Path can be relative to raw_asset_dir or absolute. Repeat this option for multiple meshes.
--mesh-glob <glob>	visuals/*.obj, visuals/*.stl, meshes/*.obj, meshes/*.stl	Glob used when --mesh is omitted. Repeatable.

Example:

python scripts/run_convex_decomposition.py \
  assets/robocasa/raw/fixtures/dishwashers/Dishwasher031 \
  --mesh visuals/door.obj \
  --mesh visuals/rack0.obj \
  --mesh visuals/rack1.obj

Repository/output options

Option	Default	Meaning
--variant <name>	coacd_v1	Output suffix. The derived asset directory becomes <source_asset_id>_<variant>.
--overwrite	false	Allow writing into an existing derived output directory. Use cautiously.
--register-manifest	false	Append the generated derived asset to manifest/assets.jsonl. Use only for outputs you intend to keep.

Output path:

assets/<source>/derived/convex_decompositions/<category>/<source_asset_id>_<variant>/

MJCF options

Option	Default	Meaning
--class-name <name>	convex_collision	class attribute for generated <geom> entries. Empty string disables it.
--rgba "r g b a"	0.8 0.2 0.2 0.35	RGBA for generated collision geoms.
--group <int>	0	MuJoCo geom group. Empty string disables it.
--contype <int>	empty	MuJoCo contact type. Empty means use MuJoCo default.
--conaffinity <int>	empty	MuJoCo contact affinity. Empty means use MuJoCo default.

Generated MJCF files:

mjcf/convex_assets_include.xml      # include at MJCF root/top level
mjcf/convex_geoms_include.xml       # include inside the target body
mjcf/convex_collision_include.xml   # standalone wrapper for preview/convenience

CoACD granularity and quality options

Option	Default	Effect
--threshold <float>	0.05	Main concavity threshold. Lower values usually create more/smaller and more accurate parts. Higher values create fewer/coarser parts.
--max-convex-hull <int>	-1	Maximum hull count. -1 means CoACD default/unlimited. Lower cap forces fewer/larger parts.
--no-merge	false	Disable CoACD hull merge. Usually preserves more/smaller parts. Useful for wire/rack-like geometry, but may create many parts.
--mcts-max-depth <int>	3	Maximum recursive split depth. Higher values can produce finer decomposition and longer runtime.
--mcts-nodes <int>	20	Search nodes per MCTS step. Higher values can improve split search and increase runtime.
--mcts-iterations <int>	150	MCTS iterations. Higher values can improve split quality and increase runtime.
--resolution <int>	2000	CoACD sampling/resolution parameter. Higher values preserve more detail and cost more time.
--preprocess-mode {auto,on,off}	auto	Whether CoACD preprocesses the input mesh.
--preprocess-resolution <int>	50	Resolution for preprocessing. Higher values may preserve more detail and cost more time.
--max-ch-vertex <int>	256	Maximum vertices per convex hull. Controls hull complexity, not physical part size directly.
--pca	false	Enable CoACD PCA mode. Can affect orientation/splitting behavior.
--decimate	false	Enable CoACD decimation. May simplify output.
--extrude	false	Enable CoACD extrusion.
--extrude-margin <float>	0.01	Margin used when extrusion is enabled.
--apx-mode {ch,box}	ch	Approximation mode: convex hull or box approximation.
--real-metric	false	Enable CoACD real_metric option if supported.
--seed <int>	0	Random seed for reproducible runs where CoACD supports it.

Practical parameter presets

Block-like assets such as doors, panels, and shells:

python scripts/run_convex_decomposition.py <raw_asset_dir> \
  --mesh visuals/door.obj \
  --variant coacd_v1 \
  --threshold 0.05 \
  --max-convex-hull 16

Coarser and faster collision:

python scripts/run_convex_decomposition.py <raw_asset_dir> \
  --mesh <mesh.obj> \
  --variant coacd_coarse_v1 \
  --threshold 0.08 \
  --max-convex-hull 8

Finer collision:

python scripts/run_convex_decomposition.py <raw_asset_dir> \
  --mesh <mesh.obj> \
  --variant coacd_fine_v1 \
  --threshold 0.02 \
  --max-convex-hull 32 \
  --no-merge \
  --mcts-max-depth 4

Rack/wire-like assets such as dishwasher racks:

python scripts/run_convex_decomposition.py <raw_asset_dir> \
  --mesh visuals/rack1.obj \
  --variant rack1_coacd_trial \
  --threshold 0.02 \
  --max-convex-hull 128 \
  --no-merge \
  --resolution 2000 \
  --mcts-max-depth 4

For rack/wire-like assets, inspect logs/parts_summary.csv and visualize the result before registering it. A valid result should not have convex hulls that bridge functional slots or gaps. If the decomposition needs hundreds of parts, primitive collision using capsules/boxes is usually a better engineering choice.

metadata.yaml

Minimum fields:

asset_id: robocasa.convex_decompositions.dishwashers.Dishwasher031_vhacd_v1
source: robocasa
asset_type: convex_decompositions
category: dishwashers
source_asset_id: Dishwasher031
format: obj_mjcf_include
entry_file: mjcf/convex_collision_include.xml
license: CC-BY-4.0
origin_url: https://robocasa.ai
path: assets/robocasa/derived/convex_decompositions/dishwashers/Dishwasher031_vhacd_v1
storage_mode: derived
derived_from:
  - assets/robocasa/raw/fixtures/dishwashers/Dishwasher031
derivation_method: convex_decomposition
decomposition_tool: vhacd
decomposition_version: TBD
decomposition_params_file: logs/decomposition.json
validation_status: pending
tags:
  - robocasa
  - convex_decomposition
  - collision
  - mujoco

source_refs.yaml

Record exact source files:

raw_asset: assets/robocasa/raw/fixtures/dishwashers/Dishwasher031
raw_entry_file: assets/robocasa/raw/fixtures/dishwashers/Dishwasher031/model.xml
raw_meshes:
  - assets/robocasa/raw/fixtures/dishwashers/Dishwasher031/visuals/door.obj
  - assets/robocasa/raw/fixtures/dishwashers/Dishwasher031/visuals/rack0.obj

logs/decomposition.json

Record tool parameters and hashes:

{
  "tool": "vhacd",
  "tool_version": "TBD",
  "created_at": "YYYY-MM-DD",
  "params": {},
  "inputs": [
    {
      "path": "assets/robocasa/raw/fixtures/dishwashers/Dishwasher031/visuals/door.obj",
      "sha256": "TBD"
    }
  ],
  "outputs": [
    {
      "path": "meshes/door/part_000.obj",
      "sha256": "TBD"
    }
  ]
}

MJCF integration rule

Do not edit raw model.xml in place. If MuJoCo collision integration is needed, write a generated include file under:

mjcf/

Generated MJCF should be split by include context:

mjcf/convex_assets_include.xml      # include at MJCF root/top level
mjcf/convex_geoms_include.xml       # include inside the target body
mjcf/convex_collision_include.xml   # standalone wrapper for preview/convenience

Recommended integration pattern:

<!-- top level -->
<include file="path/to/mjcf/convex_assets_include.xml"/>

<!-- inside the object body that should receive collision geoms -->
<include file="path/to/mjcf/convex_geoms_include.xml"/>

This split avoids mixing root-level <asset> declarations with body-level <geom> declarations. convex_collision_include.xml is provided as a convenience wrapper body and should not be used when replacing collision geometry inside an existing body hierarchy.

Manifest rule

Convex decomposition results are real derived assets and must be indexed in:

manifest/assets.jsonl

Link-only plans or empty placeholders should not be listed in the manifest.