MESA: An Evaluation Framework for Compositional, Semantic, and Spatial Generalization in Robotics

🌐 Project Website | 📚 Documentation | 🤗 Weights and Data | 📄 Paper

MESA is a framework for building, scaling, and evaluating language-conditioned tabletop manipulation policies. The project includes:

• MESA-Gen for scalable task and demonstration generation.
• MESA-Bench for standardized evaluation across in-distribution, spatial, category, instance, and compositional generalization settings.

Installation

MESA uses uv for environment and dependency management.

1. Install uv using the official guide: https://docs.astral.sh/uv/getting-started/installation/
2. Sync dependencies from the repository root:

   uv sync

3. Optional: install LeRobot export dependencies:

   uv sync --extra lerobot

4. Run setup scripts:

   ./scripts/setup.sh

For detailed setup instructions, see docs/getting_started/installation.md.

Acknowledgements

MESA builds on top of and draws inspiration from the following projects:

• MimicLabs
• MimicGen
• OpenPI
• LeRobot
• RoboCasa
• LIBERO
• tyro

See docs/other/acknowledgement.md for the documentation version of this list.

Citation

If you find MESA useful in your research, please cite:

@article{mesa2026,
  author    = {Albert Wilcox and Frank Chang and Aishani Chakraborty and Nhi Nguyen and Jeremy A. Collins and Vaibhav Saxena and Benjamin Joffe and Siddhath Karamcheti and Animesh Garg},
  title     = {MESA: An Evaluation Framework for Compositional, Semantic, and Spatial Generalization in Robotics},
  journal   = {arXiv preprint},
  year      = {2026},
}

License

This repository is released under the MIT License. See LICENSE for full terms.