KO | EN

GitLite

Minimalist Open Source Repository Search Engine

gitlite — search

#typescript #ai-agents #cli #claude-code #codex #developer-tools #ai #dsh-plugin #ai-agent #mcp #deepseek-harness #windows

scenesmith

★ 555 Open GitHub ↗

Code for "SceneSmith: Agentic Generation of Simulation-Ready Indoor Scenes", ICML 2026 Spotlight

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

<h1 align="center">
 <img src="media/symbol.svg" alt="" height="36" align="absmiddle">&nbsp;SceneSmith
</h1>

Code for "SceneSmith: Agentic Generation of Simulation-Ready Indoor Scenes".

<a href="https://nepfaff.github.io/">Nicholas Pfaff</a>1,
<a href="https://tommycohn.com/">Thomas Cohn</a>1,
<a href="https://zakharos.github.io/">Sergey Zakharov</a>2,
<a href="https://www.linkedin.com/in/rick-cory-31a68979/">Rick Cory</a>2,
<a href="https://locomotion.csail.mit.edu/russt.html">Russ Tedrake</a>1 
1Massachusetts Institute of Technology,
2Toyota Research Institute

Fully automated text-to-scene generation. This entire community center was generated by SceneSmith without any human intervention, from a single 151-word text prompt. Beyond explicitly specified elements, SceneSmith places additional objects from inferred contextual information, such as ping pong paddles and balls placed near a ping pong table. Objects are generated on-demand, are fully separable (non-composite), and include estimated physical properties, enabling direct interaction within a simulation. The resulting scenes are immediately usable in arbitrary physics simulators (robots added for demonstration).

## Citation

```bibtex
@misc{scenesmith2026,
  title={SceneSmith: Agentic Generation of Simulation-Ready Indoor Scenes},
  author={Nicholas Pfaff and Thomas Cohn and Sergey Zakharov and Rick Cory and Russ Tedrake},
  year={2026},
  eprint={2602.09153},
  archivePrefix={arXiv},
  primaryClass={cs.RO},
  url={https://arxiv.org/abs/2602.09153},
}
```

## Table of Contents

- [Installation](#-installation)
  - [Local Install](#local-install)
  - [Multi-GPU Rendering](#multi-gpu-rendering-support-optional)
  - [Asset Generation Backends](#asset-generation-backends)
  - [Environment Variables](#environment-variables)
  - [Docker](#-docker)
- [Data Setup](#-data-setup)
  - [HSSD](#hssd-dataset-setup)
  - [Objaverse](#objaverse-objectthor-setup)
  - [Articulated Objects](#articulated-object-datasets)
  - [AmbientCG Materials](#ambientcg-pbr-materials)
- [Scene Generation](#-scene-generation)
  - [Multi-GPU Support](#multi-gpu-support)
  - [Pipeline Stage Control](#pipeline-stage-control)
  - [Branching](#branching-from-a-previous-experiment)
- [Visualization](#-visualization)
- [Export](#-export)
- [Robot Evaluation](#-robot-evaluation)
- [Testing](#-testing)
- [License](#-license)

---

## 📦 Installation

> For Docker-based setup with GPU support, see [Docker](#-docker) below.

### Local Install

This repository uses [uv](https://docs.astral.sh/uv) for dependency management.

Install uv:
```sh
curl -LsSf https://astral.sh/uv/install.sh | sh
```

Install the dependencies (including dev tools like pytest) into `.venv`:
```sh
uv sync
```

To install without dev dependencies:
```sh
uv sync --no-dev
```

Activate the virtual env:
```sh
source .venv/bin/activate
```

Install the pre-commit hooks:
```sh
pre-commit install
```

> **Important:** Installation is not complete yet. The default configuration requires
> additional data and model checkpoints. After installing dependencies, continue with:
> 1. [SAM3D Backend](#sam3d-backend) - 3D asset generation (default backend)
> 2. [Articulated Objects (ArtVIP)](#artvip-setup) - Articulated furniture
> 3. [AmbientCG Materials](#ambientcg-pbr-materials) - PBR materials

### Multi-GPU Rendering Support (Optional)

For distributing Blender rendering across multiple GPUs (recommended for parallel
scene generation), install bubblewrap:

```sh
# Ubuntu/Debian
sudo apt-get install bubblewrap
```

This enables GPU isolation for EEVEE Next rendering, preventing OOM errors when
running many scenes in parallel. Each scene's BlenderServer is isolated to a single
GPU via OS-level namespacing. Without bubblewrap, all Blender instances share GPU 0.

### Asset Generation Backends

SceneSmith supports two backends for open-set 3D asset generation. If using
`strategy: "hssd"` (retrieval from HSSD library), no generation backend is needed.

We support distributing generations across GPUs and hence recommend a multi-GPU node
for the best scene generation throughput when using generated assets. We found AWS
`g6e.48xlarge` instances to work great with our system.

#### Hunyuan3D-2 Backend

Lower-quality asset generation but fits into 24GB GPU memory. The results will be significantly worse than when using SAM3D. We only recommend this as a proof-of-concept if your are highly GPU memory limited. Use SAM3D for production results.

1. **Install submodules**:
   ```sh
   git submodule update --init --recursive
   ```

2. **Install Hunyuan3D-2**:
   ```sh
   bash scripts/install_hunyuan3d.sh
   ```

3. **Enable Hunyuan3D-2 in configuration**:
   ```yaml
   # In your agent config files
   asset_manager:
     backend: "hunyuan3d"
   ```

#### SAM3D Backend

Higher-quality asset generation but requires 32GB GPU memory.
**This is the backend used in the paper.**

1. **Request access to the SAM models**:
   - [SAM-3D](https://huggingface.co/facebook/sam-3d-objects)
   - [SAM3](https://huggingface.co/facebook/sam3)

Automatic approval usually happens within 30min.

2. **Authenticate with HuggingFace** (required for checkpoint download):
   ```sh
   huggingface-cli login
   ```

3. **Install SAM3D**:
   ```sh
   bash scripts/install_sam3d.sh
   ```

This will:
   - Clone SAM3 and SAM 3D Objects repositories
   - Install dependencies (pytorch3d, gsplat, kaolin, nvdiffrast)
   - Download model checkpoints (~5GB)

4. **Enable SAM3D in configuration**:
   ```yaml
   # In your agent config files
   asset_manager:
     backend: "sam3d"
   ```

### Environment Variables

Set the following environment variables. These are required for both local and
Docker usage (Docker Compose forwards them from the host).

```sh
# Required: OpenAI API key for GPT-5 agents and default image generation
export OPENAI_API_KEY="your-openai-key"

# Optional: Google API key for Gemini image generation backend
# Only required if using image_generation.backend: "gemini" in config
export GOOGLE_API_KEY="your-google-key"

# Optional: Separate API key for OpenAI Agents tracing
# Allows traces to appear on a different account than is used for billing
export OPENAI_TRACING_KEY="your-tracing-key"
```

---

## 🐳 Docker

Alternatively, run SceneSmith in a Docker container with NVIDIA GPU support.
All servers (geometry generation, retrieval, blender, etc.) are auto-managed
by the pipeline inside the container.

### Prerequisites

- **Docker Engine** (apt-based, **not** snap): Install via
  [official instructions](https://docs.docker.com/engine/install/ubuntu/) or
  `curl -fsSL https://get.docker.com | sudo sh`. The snap version has
  sandboxing issues that prevent GPU access.
- **NVIDIA Container Toolkit**: Install from the
  [NVIDIA repo](https://docs.nvidia.com/datacenter/cloud-native/container-toolkit/install-guide.html),
  then configure and restart Docker:
  ```sh
  sudo nvidia-ctk runtime configure --runtime=docker
  sudo systemctl restart docker
  ```

### Build

```sh
docker build -t scenesmith .
```

### Data Preparation

Data directories are **not** baked into the image. Download them on the host
(see the dataset sections below) and mount into the container.

### SAM3D Checkpoints

SAM3D model weights require HuggingFace authentication and are mounted at
runtime:

```sh
huggingface-cli login

# Run the host install script (`scripts/install_sam3d.sh`, only the checkpoint download part)
# Or manually download:
mkdir -p external/checkpoints
huggingface-cli download facebook/sam3 sam3.pt --local-dir external/checkpoints
huggingface-cli download facebook/sam-3d-objects \
    --repo-type model \
    --local-dir external/checkpoints/sam-3d-objects-download \
    --include "checkpoints/*"
mv external/checkpoints/sam-3d-objects-download/checkpoints/* external/checkpoints/
rm -rf external/checkpoints/sam-3d-objects-download
```

### Run

```sh
# Interactive shell with all volumes and env vars
docker compose run --rm scenesmith bash

# Then inside the container:
python main.py +name=my_experiment
```

```sh
# Or run a one-off command directly
docker compose run --rm scenesmith python main.py +name=my_experiment

# Smoke test
docker compose run --rm scenesmith \
    python -c "import torch; print(torch.cuda.is_available()); import scenesmith"

# Unit tests
docker compose run --rm scenesmith pytest tests/unit/ -x
```

### Volume Mounts

| Host Path | Container Path | Content |
|-----------|---------------|---------|
| `./data/` | `/app/data/` | HSSD models, Objaverse assets, materials, preprocessed indices |
| `./external/checkpoints/` | `/app/external/checkpoints/` | SAM3D model weights |
| `./outputs/` | `/app/outputs/` | Generated scenes |

> **Important:** The Docker image does not include model checkpoints or datasets.
> Before running scene generation, you must download and mount the required data.
> The default configuration requires:
> 1. [SAM3D checkpoints](#sam3d-checkpoints) (mounted via `external/checkpoints/`)
> 2. [Articulated Objects (ArtVIP)](#artvip-setup) (mounted via `data/`)
> 3. [AmbientCG Materials](#ambientcg-pbr-materials) (mounted via `data/`)

---

## 💾 Data Setup

SceneSmith supports multiple asset and material sources. The **default configuration**
requires three data dependencies:

1. **SAM3D** - 3D asset generation backend (see [installation above](#sam3d-backend))
2. **ArtVIP** - Articulated furniture (cabinets, drawers, etc.)
3. **AmbientCG Materials** - PBR materials for walls, floors, and surfaces

HSSD, Objaverse, and PartNet-Mobility are **optional alternative** asset sources
that can be enabled via configuration overrides.

### HSSD Dataset Setup

For asset retrieval using the HSSD object library instead of generative 3D models:

1. **Accept the HSSD license** at https://huggingface.co/datasets/hssd/hssd-models

2. **Download HSSD models** (~72GB, requires Git LFS):
   ```sh
   cd data
   git lfs install
   git clone git@hf.co:datasets/hssd/hssd-models
   ```

3. **Download preprocessed data** (~60MB + ~2GB):

This data is from HSM (https://arxiv.org/abs/2503.16848) and includes:
   - CLIP indices and embeddings for semantic search (~60MB)
   - Pre-validated support surfaces (~2GB, provides ~10x speedup)

```sh
   bash scripts/download_hssd_data.sh
   ```

Or manually:
   ```sh
   # Download CLIP indices
   wget https://github.com/3dlg-hcvc/hsm/releases/latest/download/data.zip
   unzip data.zip -d data/preprocessed

# Download pre-validated support surfaces
   wget https://github.com/3dlg-hcvc/hsm/releases/latest/download/support-surfaces.zip
   unzip support-surfaces.zip -d data/hssd-models
   ```

4. **Enable HSSD in configuration**:
   ```yaml
   # In your experiment config file
   asset_manager:
     strategy: "hssd"  # Use HSSD retrieval instead of generation
   ```

The `strategy: "generated"` uses either Hunyuan3D or SAM3D for open-set asset generation (controlled by `backend` setting). HSSD assets automatically use pre-validated support surfaces when available, eliminating the need to recompute them.

### Objaverse (ObjectThor) Setup

For asset retrieval using the ObjectThor subset of Objaverse:

1. **Download ObjectThor data** (~50GB assets + ~200MB features):
   ```sh
   bash scripts/download_objaverse_data.sh
   ```

This downloads:
   - ObjectThor assets (GLB meshes)
   - Annotations with placement constraints and metadata
   - Pre-computed CLIP features (3 views, 768-dim)

2. **Preprocess for retrieval**:
   ```sh
   python scripts/prepare_objaverse.py
   ```

This creates `data/objathor-assets/preprocessed/` with:
   - Averaged CLIP embeddings (768-dim per object)
   - Metadata index with categories and bounding boxes
   - Object category mapping for filtering

3. **Enable Objaverse in configuration**:
   ```yaml
   # In your experiment config file
   asset_manager:
     general_asset_source: "objaverse"  # Use Objaverse retrieval
   ```

### Articulated Object Datasets

For articulated objects (cabinets with doors, drawers, etc.), we support two datasets:
PartNet-Mobility and ArtVIP. Both require preprocessing to convert to our SDF format.

Note that the quality (both mesh quality and joint quality) of the PartNet-Mobility dataset is very low.

#### PartNet-Mobility Setup

1. **Download PartNet-Mobility** from https://sapien.ucsd.edu/downloads

Extract to `data/partnet-mobility-v0/`.

2. **Convert to SDF format**:
   ```sh
   python scripts/convert_partnet_mobility.py \
       --input data/partnet-mobility-v0 \
       --output data/partnet_mobility_sdf
   ```

For parallel processing, use the wrapper script:
   ```sh
   bash scripts/convert_partnet_parallel.sh data/partnet-mobility-v0 data/partnet_mobility_sdf 8
   ```

This uses VLM analysis to determine:
   - Physics properties (mass, inertia per link)
   - Front-facing orientation (canonicalized to +Y forward, Z up)
   - Scale correction
   - Placement type (floor/wall/ceiling/on-object)

3. **Compute CLIP embeddings** for text-based retrieval:
   ```sh
   python scripts/compute_articulated_embeddings.py \
       --source partnet_mobility \
       --data-path data/partnet_mobility_sdf \
       --output-path data/partnet_mobility_sdf/embeddings
   ```

Optional: Keep rendered images for inspection:
   ```sh
   python scripts/compute_articulated_embeddings.py \
       --source partnet_mobility \
       --data-path data/partnet_mobility_sdf \
       --output-path data/partnet_mobility_sdf/embeddings \
       --keep-renders
   ```

#### ArtVIP Setup

We provide preprocessed ArtVIP assets (converted to Drake SDFormat with collision
geometries and CLIP embeddings) on HuggingFace:

```sh
# Download VHACD variant (recommended — tighter collision geometries)
huggingface-cli download nepfaff/scenesmith-preprocessed-data \
    artvip/artvip_vhacd.tar.gz --repo-type dataset --local-dir .
mkdir -p data/artvip_sdf
tar xzf artvip/artvip_vhacd.tar.gz -C data/artvip_sdf
rm -rf artvip
```

Alternatively, download the CoACD variant (can produce faster simulations):
```sh
huggingface-cli download nepfaff/scenesmith-preprocessed-data \
    artvip/artvip_coacd.tar.gz --repo-type dataset --local-dir .
mkdir -p data/artvip_sdf
tar xzf artvip/artvip_coacd.tar.gz -C data/artvip_sdf
rm -rf artvip
```

To preprocess ArtVIP assets yourself instead (e.g., with updated data or custom
settings), use [mesh-to-sim-asset](https://github.com/nepfaff/mesh-to-sim-asset)
to convert from USD to Drake SDFormat, then compute CLIP embeddings:
```sh
python scripts/compute_articulated_embeddings.py \
    --source artvip \
    --data-path data/artvip_sdf \
    --output-path data/artvip_sdf/embeddings
```

#### Test Retrieval

Test that retrieval works correctly:
```sh
python scripts/test_asset_retrieval.py \
    --source partnet_mobility \
    --query "wooden cabinet with drawers" \
    --top-k 5 \
    --output-path output/retrieval_test
```

This renders multi-view images of retrieved objects for visual inspection.

#### Enable Articulated Retrieval

Add articulated sources to your agent configuration:

```yaml
# In configurations/furniture_agent/base_furniture_agent.yaml
asset_manager:
  router:
    enabled: true
  articulated:
    use_top_k: 5  # Number of CLIP candidates before bbox ranking
    sources:
      partnet_mobility:
        enabled: false  # Disabled by default (low quality)
        data_path: data/partnet_mobility_sdf
        embeddings_path: data/partnet_mobility_sdf/embeddings
      artvip:
        enabled: true  # Enabled by default
        data_path: data/artvip_sdf
        embeddings_path: data/artvip_sdf/embeddings
```

### AmbientCG PBR Materials

Download free CC0 PBR materials from [AmbientCG](https://ambientcg.com/) for scene
rendering:

1. **Download materials**:
   ```sh
   python scripts/download_ambientcg.py --output data/materials
   ```

Options:
   ```sh
   # Download specific resolution/format
   python scripts/download_ambientcg.py -r 2K -f PNG --output data/materials

# Limit number of materials (for testing)
   python scripts/download_ambientcg.py --limit 100 --output data/materials

# Dry run to see what would be downloaded
   python scripts/download_ambientcg.py --dry-run
   ```

2. **Download pre-computed CLIP embeddings** (recommended):
   ```sh
   huggingface-cli download nepfaff/scenesmith-preprocessed-data \
       --repo-type dataset \
       --include "ambientcg/embeddings/**" \
       --local-dir data/scenesmith-preprocessed-data
   mv data/scenesmith-preprocessed-data/ambientcg/embeddings data/materials/embeddings
   rm -rf data/scenesmith-preprocessed-data
   ```

Or compute them yourself:
   ```sh
   python scripts/compute_ambientcg_embeddings.py --materials-dir data/materials
   ```

3. **Test retrieval**:
   ```sh
   python scripts/test_material_retrieval.py \
       --materials-dir data/materials \
       --query "red brick wall" \
       --top-k 5

# Save preview images for inspection
   python scripts/test_material_retrieval.py \
       --materials-dir data/materials \
       --query "wooden floor" \
       --top-k 5 \
       --output-path output/material_test
   ```

---

## 🎬 Scene Generation

```sh
python main.py +name=run_name
```

Set the scene prompts in `experiment.prompts` (`configurations/experiment/base_experiment.yaml`).
Set `floor_plan_agent.mode="house"` for house scenes and `floor_plan_agent.mode="room"` for single room scenes.

Note that you will need >=24GB of GPU memory for Hunyuan3D asset generation and >=32GB for SAM3D asset generation. The material and articulated retrieval servers require additional GPU memory when enabled. Hence, we recommend >=45GB of GPU memory to run the full pipeline as documented in the research paper. All code was tested with an L40S GPU.

### Multi-GPU Support

The geometry generation server automatically detects and uses all available GPUs,
spawning one worker process per GPU for parallel asset generation. This significantly
speeds up image-to-3D asset generation and thus scene generation.

To control which GPUs are used, set `CUDA_VISIBLE_DEVICES`:

```sh
# Use only GPUs 0, 1, 2, 3 (4 workers)
CUDA_VISIBLE_DEVICES=0,1,2,3 python main.py +name=my_experiment

# Force single GPU mode
CUDA_VISIBLE_DEVICES=0 python main.py +name=my_experiment

# Use all available GPUs (default - no env var needed)
python main.py +name=my_experiment
```

### Pipeline Stage Control

The scene generation pipeline has five stages that run in order:
1. **floor_plan** - Generate room geometry (walls, floor)
2. **furniture** - Place furniture in the room
3. **wall_mounted** - Place wall-mounted objects (mirrors, artwork, shelves, clocks)
4. **ceiling_mounted** - Place ceiling fixtures (chandeliers, pendant lights, ceiling fans)
5. **manipuland** - Place small objects on furniture surfaces

Use `pipeline.start_stage` and `pipeline.stop_stage` to control which stages run:

#### Generate full scenes (default)
```sh
python main.py +name=my_experiment
```

#### Stop after a specific stage
```sh
python main.py +name=my_experiment experiment.pipeline.stop_stage=furniture
```

#### Resume from a later stage
```sh
python main.py +name=my_experiment experiment.pipeline.start_stage=manipuland
```

**State persistence**: The pipeline automatically saves state after each stage:
- `house_layout.json` - After floor_plan stage
- `scene_states/scene_after_furniture/` - After furniture stage
- `scene_states/scene_after_wall_objects/` - After wall_mounted stage
- `scene_states/scene_after_ceiling_objects/` - After ceiling_mounted stage
- `scene_states/final_scene/` - After manipuland stage

When resuming from a later stage, the pipeline loads from the previous stage's
saved state. This enables iterative development: generate expensive stages once,
then quickly iterate on later stages.

### Branching from a Previous Experiment

Use `resume_from_path` to create multiple independent runs from a single
checkpoint. This is useful for A/B testing different configurations or debugging
a specific stage.

```sh
# First run: generate floor plans and furniture
python main.py +name=base_run experiment.pipeline.stop_stage=furniture

# Branch 1: add manipulands with default config
python main.py +name=branch_1 \
  experiment.pipeline.start_stage=manipuland \
  experiment.pipeline.resume_from_path=outputs/2025-12-21/10-30-45

# Branch 2: add manipulands with different config
python main.py +name=branch_2 \
  experiment.pipeline.start_stage=manipuland \
  experiment.pipeline.resume_from_path=outputs/2025-12-21/10-30-45 \
  manipuland_agent.some_param=different_value
```

When `resume_from_path` is set:
1. The source scene is copied to the new output directory
2. Absolute paths in checkpoint files are automatically fixed
3. The pipeline continues from `start_stage` with the new configuration

This creates fully independent output directories, preserving the original run.

---

## 🔍 Visualization

### Intermediate Scenes

Intermediate scene states are saved as `scene.dmd.yaml` files inside `scene_renders/`
directories during generation. View them interactively with Drake's model visualizer:

```sh
python -m pydrake.visualization.model_visualizer \
    outputs/YYYY-MM-DD/HH-MM-SS/scene_000/room_*/scene_renders/furniture/renders_001/scene.dmd.yaml
```

This opens an interactive 3D viewer in the browser where you can inspect object placement, collision geometry, articulated joints, and scene structure at any render checkpoint.

### Portable House Scenes

Final combined house scenes use `package://scene/` URIs for portability. These scenes
can be moved or shared without breaking file paths. To view them, set `ROS_PACKAGE_PATH`
to the scene directory:

```sh
# View a portable house scene
export ROS_PACKAGE_PATH=/path/to/outputs/YYYY-MM-DD/HH-MM-SS/scene_000:$ROS_PACKAGE_PATH
python -m pydrake.visualization.model_visualizer \
    outputs/YYYY-MM-DD/HH-MM-SS/scene_000/combined_house_after_furniture/house.dmd.yaml
```

Each scene directory contains a `package.xml` file that registers it as a ROS-style
package named "scene". Drake's model visualizer automatically discovers packages via
`ROS_PACKAGE_PATH`.

For programmatic loading (e.g., in Python scripts), register the package directly:

```python
from pydrake.multibody.parsing import Parser

parser = Parser(plant)
parser.package_map().Add("scene", "/path/to/scene_000")
```

Note that final scenes after all agent stages are saved to `outputs/YYYY-MM-DD/HH-MM-SS/scene_000/combined_house/`. This includes both a `house.blend` Blender file and a `house.dmd.yaml` Drake file. See below for converting this scene into alternative formats (MuJoCo or USD).

The generated output directories can be huge. To remove all files that aren't used in the final scene, use `scripts/clean_scene_output.py`.

---

## 📤 Export

SceneSmith's native output format is Drake Directives (`.dmd.yaml`).

### Export to MuJoCo and USD

MuJoCo and USD export require a separate virtual environment because:
- `mujoco` is not included in the main venv
- `mujoco-usd-converter` conflicts with `bpy` (Blender) due to incompatible OpenUSD (`pxr`) versions

Setup the MuJoCo export environment:
```sh
./scripts/setup_mujoco_export.sh
```

```sh
source .mujoco_venv/bin/activate

# Export a generated scene
python scripts/export_scene_to_mujoco.py outputs/YYYY-MM-DD/HH-MM-SS/scene_000 \
    -o mujoco_export

# Export a standalone SDF model (e.g., robot arm)
python scripts/export_scene_to_mujoco.py --sdf /path/to/model.sdf -o mujoco_export

# Export to USD format (adds usd/ subdirectory)
python scripts/export_scene_to_mujoco.py --sdf /path/to/model.sdf -o mujoco_export --usd

# View in MuJoCo
python -m mujoco.viewer --mjcf=mujoco_export/scene.xml
```

The USD output will be in `mujoco_export/usd/`.

> **Note:** Simulator export (MuJoCo, USD/Isaac Sim) is experimental.
> Scene quality may vary slightly across different simulators compared
> to the native Drake format. We welcome feedback and contributions
> for additional export targets.

---

## 🤖 Robot Evaluation

The robot evaluation module provides tools for task-based scene generation and validation. It converts human tasks (e.g., "Find a fruit and place it on the table") into generated scenes and validates task completion.

Robot manipulation evaluation pipeline. Given a manipulation task (e.g., "Pick a fruit from the fruit bowl and place it on a plate"), an LLM generates diverse scene prompts specifying scene constraints implied by the task. SceneSmith generates scenes from each prompt. A robot policy attempts the task in simulation, and an evaluation agent verifies success using simulator state queries and visual observations. This enables scalable policy evaluation without manual environment or success predicate design.

### Pipeline Stages

The evaluation pipeline has 4 stages:

1. **Generate Prompts** - LLM converts human task → diverse scene prompts
2. **Generate Scenes** - Use `main.py` with generated prompts (SceneSmith pipeline)
3. **Policy Interface** - Convert scene → robot-executable poses (optional, not needed for language-conditioned policies)
4. **Validate** - VLM agent checks if task is completed

### Stage 1: Generate Scene Prompts

Convert a human task description into diverse scene prompts:

```sh
python scripts/robot_eval/generate_prompts.py \
    --task "Find a fruit and place it on the kitchen table" \
    --output-dir outputs/eval_run \
    --num-prompts 5
```

This generates:
- `outputs/eval_run/prompts.csv` - Scene prompts for main.py
- `outputs/eval_run/task_metadata.yaml` - Task metadata for validation

### Stage 2: Generate Scenes

Run standard scene generation with the generated prompts:

```sh
python main.py +name=eval experiment.csv_path=outputs/eval_run/prompts.csv
```

### Stage 3: Policy Interface (Predicates → Robot Poses)

Extract robot-executable poses from a generated scene. This is useful for model-based policies that require explicit pose targets instead of natural language task descriptions. We used this as part of the proof-of-concept model-based policy in our paper. However, this component isn't very general and language-conditioned policies that could skip this step have potential to perform better.

```sh
python scripts/robot_eval/policy_interface.py \
    --scene-state outputs/.../scene_002/combined_house/house_state.json \
    --dmd outputs/.../scene_002/combined_house/house.dmd.yaml \
    --scene-dir outputs/.../scene_002 \
    --task "Find a speaker and place it on the bed" \
    --output-json robot_commands.json
```

Arguments:
- `--scene-state` - Path to scene_state.json (per-room) or house_state.json (combined house)
- `--dmd` - Path to scene.dmd.yaml or house.dmd.yaml (Drake scene with poses)
- `--scene-dir` - Scene root directory for package:// URI resolution (default: parent of DMD)

The policy interface uses a unified LLM agent with access to state tools (support, containment, spatial relations) and optionally vision tools (scene renders). The agent autonomously reasons about the task to:
- Parse goals and preconditions (e.g., "from the floor")
- Find objects matching categories
- Verify preconditions using tools
- Return ALL valid (target, reference) bindings ranked by confidence

The output includes both a sampled target pose and contained placement bounds (AABB shrunk by object half-extents) that robot policies can sample from.

### Stage 4: Validate Task Completion

After the robot executes the task, it writes the final object poses back to a `.dmd.yaml` file.
The validator then compares the original scene metadata with the robot's output poses to
determine task success.

**Robot policy contract:**
1. Robot receives initial `scene.dmd.yaml` (from scene generation)
2. Robot performs the task
3. Robot outputs modified `scene.dmd.yaml` with updated object poses

**Validation:**

```sh
python scripts/robot_eval/validate.py \
    --scene-state outputs/.../scene_002/combined_house/house_state.json \
    --dmd outputs/.../scene_002/combined_house/house.dmd.yaml \
    --scene-dir outputs/.../scene_002 \
    --task "Find a speaker and place it on the bed"
```

Arguments:
- `--scene-state` - Path to scene_state.json (per-room) or house_state.json (combined house)
- `--dmd` - Path to scene.dmd.yaml or house.dmd.yaml (poses from robot output)
- `--scene-dir` - Scene root directory for package:// URI resolution (default: parent of DMD)

The validator loads object metadata from scene_state.json and poses from the DMD file.
It uses geometric state tools and vision tools (renders) to assess task completion.

---

## 🧪 Testing

Run unit tests:
```sh
pytest tests/unit/ --testmon
```

Run integration tests:
```sh
pytest tests/integration/ --testmon
```

Run all tests:
```sh
pytest tests/ --testmon
```

Omit `--testmon` to run the full test suite without caching. Use `-x` to stop on first failure.

Note that some heavy integration tests are skipped in the GitHub CI and must be run locally.

---

## 📄 License

This project is licensed under the MIT License - see [LICENSE](LICENSE) for details.

scenesmith

Explore Similar Repositories

Was this content helpful?

Recent Feedback

Download README