R³L: Reasoning 3D Layouts from Relative Spatial Relations

Zhifeng Gu^*   Yuqi Wang^*   Bing Wang

Spatial Intelligence Group, The Hong Kong Polytechnic University

---

Get Started

Tested on macOS and Ubuntu. The Blender renderer auto-routes to METAL (macOS) or CUDA (Ubuntu).

Prerequisites: Python 3.10 (enforced by uv). Install uv if you haven't:

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

1. Sync dependencies and fetch data & retriever models:

uv sync
uv run python scripts/download_data.py    # objathor asset data
uv run python scripts/download_models.py  # retriever models

2. Set your API keys:

cp scripts/env.example.sh scripts/env.sh
vim scripts/env.sh

3. Run:

source scripts/env.sh
uv run -q python main.py

If you run into any trouble, please file an issue.

Acknowledgements

This project builds upon Holodeck and LayoutVLM. We sincerely thank the authors for their open-source contributions.

Citation

@inproceedings{gu2026r3l,
  title     = {R$^3$L: Reasoning 3D Layouts from Relative Spatial Relations},
  author    = {Gu, Zhifeng and Wang, Yuqi and Wang, Bing},
  booktitle = {International Conference on Machine Learning (ICML)},
  year      = {2026},
  url       = {https://arxiv.org/abs/2605.06758}
}

License

This project is licensed under the MIT License. Vendored third-party components (under planners/holodeck_v2/, retrievers/holodeck_v2/, utils/third_party/) retain their own licenses.