Skip to content

🏠 Room Generation — Large-Scale Indoor Scenes

Generate multi-room, navigable, instance-editable indoor scenes as sim-ready backgrounds, built on top of Infinigen. Rooms are generated at a controllable complexity tier and can be exported to URDF and USD for direct use in simulators.

Large-scale multi-room scenes

Installation

Run bash install.sh room to install the additional dependencies (including the bundled Blender runtime) before using room generation.

⚡ Command-Line Usage

# Natural-language mode: a GPT router infers room type & complexity
room-cli -m embodied_gen.scripts.room_gen.gen_room \
  --output-root outputs/rooms \
  --prompt "Wipe the table in a simple dining room"

# Explicit mode
room-cli -m embodied_gen.scripts.room_gen.gen_room \
  --output-root outputs/rooms \
  --room-type Kitchen \
  --seed 42 \
  --complexity simple

Natural-language mode

--prompt feeds the description to a GPT router (route_room.py) that selects a plausible room type (or House for cross-room tasks) and a complexity tier, overriding --room-type / --complexity. It requires a configured GPT agent (embodied_gen/utils/gpt_config.yaml).

room-cli vs python -m

room-cli forwards its arguments to the bundled Blender Python (which has bpy and all room dependencies installed by bash install.sh room). Running python -m embodied_gen.scripts.room_gen.gen_room from the embodiedgen env is equivalent for generation, since heavy stages are dispatched to Blender Python internally either way. Run from the repository root; set BLENDER_PYTHON_BIN if Blender is installed elsewhere.

Parameters

  • --output-root (required): base output directory.
  • --prompt: natural-language task/scene description; infers room type and complexity via GPT, overriding the two flags below.
  • --room-type: Bedroom | LivingRoom | Kitchen | Bathroom | DiningRoom | Office | House.
  • --seed: random seed. Pass explicitly for reproducible runs; if omitted, a random seed is generated (check logs for the value).
  • --complexity: minimalist | simple | medium | detail.
  • --custom-params: gin file copied to Infinigen custom_solve.gin.
  • --large-scene: only for House; enables more rooms.
  • --gen/--no-gen, --urdf/--no-urdf, --usd/--no-usd: pipeline stage switches.

Complexity Tiers

Tier Description
minimalist Fastest, sparse furniture
simple Default, balanced quality/time
medium Richer layout, slower
detail Highest detail, longest runtime

Stage Control

# Generation only (no export)
python -m embodied_gen.scripts.room_gen.gen_room \
  --output-root outputs/rooms --room-type LivingRoom \
  --seed 100 --complexity medium --no-urdf --no-usd

# Export only, from an existing Blender output
python -m embodied_gen.scripts.room_gen.gen_room \
  --output-root outputs/rooms --room-type LivingRoom \
  --seed 100 --no-gen --urdf --usd

The generated results are organized as follows:

<output_root>/<RoomType>_seed<seed>
├── blender                     # Raw Infinigen/Blender scene output
├── urdf/export_scene
│   ├── scene.urdf              # Whole-room URDF referencing all instances
│   └── <instance>/             # Per-instance meshes with convex-decomposed collisions
└── usd/export_scene
    ├── export_scene.usdc       # USD scene for Isaac Sim
    └── textures                # Baked PBR textures (diffuse / normal / roughness)
  • urdf/export_scene/scene.urdf → Load directly in SAPIEN / PyBullet / IsaacGym; each furniture instance stays individually editable
  • usd/export_scene/export_scene.usdc → USD scene for Isaac Sim; requires a physics post-processing step before use (see below)
  • blender/ → Reusable for re-export (--no-gen --urdf --usd)

Add physics to the exported USD before loading it in Isaac Sim

The exported export_scene.usdc carries no physics collision properties yet. Post-process it with PhysicsUSDAdder (requires an Isaac Sim / USD environment), which applies convex-decomposition collision and rigid-body APIs to every mesh prim:

from embodied_gen.data.asset_converter import PhysicsUSDAdder

PhysicsUSDAdder().convert(
    "outputs/rooms/Kitchen_seed42/usd/export_scene/export_scene.usdc"
)
# In-place by default; pass output_file=... to write a separate copy.

Blender Python (room-cli)

room-cli is a wrapper that executes a script within the bundled Blender Python environment (which provides bpy). Steps that write USD require it, e.g. room-cli -m embodied_gen.skills.spatial-computing.cli.main ... when editing scenes with the spatial computing skill. If Blender is installed elsewhere, set BLENDER_PYTHON_BIN.

Next Steps

  • Place or remove objects in the generated rooms via natural language — see 3D Vibe Coding.
  • Use generated rooms as backgrounds for task-driven worlds — see Layout Generation.