⚠️ Alpha Release Notice
This is an alpha release for MOPS data generation. The public API might still change, and some bugs might be present.

This project builds on ManiSkill3, a simulation framework built on SAPIEN, which requires Python 3.10.
We recommend uv for environment and dependency management.

# Install uv (if not already installed)
curl -LsSf https://astral.sh/uv/install.sh | sh

# Create a Python 3.10 virtual environment and install all dependencies
uv venv --python 3.10
source .venv/bin/activate
uv pip install -e .

1. Download RoboCasa assets (required for kitchen dataset)

   python -m mani_skill.utils.download_asset RoboCasa

2. Download PartNet Mobility assets (required for all datasets)
   Get the assets from SAPIEN UCSD and place them under data/partnet_mobility/.

3. Expected directory structure

   mops-data/
   ├── data/
   │   ├── partnet_mobility/   # PartNet-Mobility assets
   │   └── mops_data/          # generated datasets (output)
   ├── scripts/                # dataset generation entry points
   ├── demos/                  # visualisation & exploration scripts
   ├── src/
   │   └── mops_data/          # core library
   └── ...
   

Three dataset variants are provided. Each has a debug mode (small, fast) and a full mode (production quality).

Isolated PartNet-Mobility objects rendered from multiple viewpoints with varied lighting.

# Debug run (quick sanity check)
python scripts/generate_single_object.py --debug

# Full generation
python scripts/generate_single_object.py

# Custom output path
python scripts/generate_single_object.py --output data/mops_data/my_single_obj

Objects placed inside RoboCasa kitchen environments, rendered from table-level and overhead viewpoints.

python scripts/generate_kitchen.py --debug   # debug run
python scripts/generate_kitchen.py           # full generation

Cluttered tabletop scenes with multiple objects rendered from top-down viewpoints.

python scripts/generate_clutter.py --debug   # debug run
python scripts/generate_clutter.py           # full generation

All scripts default to WebDataset (sharded TAR archives), which is the recommended format for publishing on Hugging Face Hub. To generate in legacy HDF5 format instead:

python scripts/generate_kitchen.py --debug --format hdf5

To convert an existing HDF5 dataset to WebDataset:

python scripts/convert_hdf5_to_webdataset.py data/mops_data/dataset.h5 [-o output_dir]

Each script accepts --output <path> to override the output directory and --format <webdataset|hdf5> to select the output format.
For deeper customisation, edit the corresponding config class in src/mops_data/generation/:

Key parameters shared by all configs (BaseDatasetConfig):

For code development, install the dev dependencies and configure pre-commit hooks:

uv pip install -e ".[dev]"
pre-commit install
pre-commit run

This project builds on ManiSkill3, so much of its official documentation is applicable to MOPS-data.

• Custom environments accept an optional np_rng keyword argument (a NumPy Generator object) for reproducible scene sampling.
• AffordanceKitchenEnv-v1 supports an optional preroll argument to deterministically recreate a specific kitchen setup.

• 🎨 Photoreal Simulation: High-quality visual rendering for computer vision
• 🤖 Robotic Manipulation: Specialised environments for manipulation tasks
• 🏠 Kitchen Environments: Realistic household scenarios built on RoboCasa
• 📦 Object Diversity: Support for cluttered tabletop and single-object scenarios
• 🔧 Extensible: Modular pipeline and config design for custom datasets

If you use MOPS-data in your research, please cite:

@article{li2026mops,
  title={Multi-Objective Photoreal Simulation (MOPS) Dataset for Computer Vision in Robotic Manipulation},
  author={
    Maximilian Xiling Li and
    Paul Mattes and
    Nils Blank and
    Rudolf Lioutikov
  },
  year={2026}
}

We welcome contributions! Please see the development setup above and feel free to submit issues and pull requests.

This project is licensed under the MIT License - see the LICENSE file for details.