LabBuilder: Protocol-Grounded 3D Layout Generation
for Interactable and Safe Laboratory

ICML 2026

---

Overview

Given a free-form experimental description, LabBuilder generates end-to-end 3D laboratory layouts that are geometrically compliant, chemically safe, and robot-navigable. It operates through three tightly coupled components:

• LabForge — Curates a meta-dataset of annotated assets and chemical knowledge, translating natural language specifications into structured protocols.
• LabGen — Synthesizes laboratory layouts via hierarchical initialization, geometric & chemical optimization, and navigation-aware refinement.
• LabTouchstone — Evaluates layouts across geometric compliance, feasibility, chemical safety, and semantic plausibility.

Installation

git clone https://github.com/che-0212/LabBuilder.git
cd LabBuilder
pip install -r requirements.txt

API Configuration

LabBuilder uses OpenAI-compatible LLM APIs. Configure your credentials:

cp .env.example .env
# Edit .env with your API key and base URL

Or set environment variables directly:

export OPENAI_API_KEY="your-api-key"
export OPENAI_BASE_URL="https://api.openai.com/v1"

3D Assets

Download the annotated 3D asset files (USD format):

# Download from [link to be added]
# Extract to assets/

The asset metadata (data/assets_annotated.json) is included in this repository and contains geometry, semantic, and chemical safety annotations for 176 laboratory assets.

Release Status

Module	Status	Description
LabForge	Released	Protocol synthesis from experimental descriptions
LabGen	Coming soon	Hierarchical layout generation & optimization
LabTouchstone	Coming soon	Four-dimensional evaluation benchmark

Code Structure

LabBuilder/
├── labforge/                    # LabForge: Protocol Synthesis
│   ├── protocol_planner.py      #   Main protocol generation pipeline
│   ├── protocol_prompt.py       #   Prompt templates for protocol synthesis
│   ├── protocol_schema.py       #   Data schemas (Protocol, Step, Asset, Constraint)
│   └── ragflow_client.py        #   RAG-based knowledge retrieval (optional)
│
├── labgen/                      # [Coming Soon] LabGen: Layout Generation & Optimization
│
├── labtouchstone/               # [Coming Soon] LabTouchstone: Evaluation Suite
│
├── data/                        # Data & Knowledge Base
│   ├── assets_annotated.json    #   Asset Knowledge Base (176 lab assets)
│   ├── experiments.json         #   30 experiment descriptions (benchmark)
│   ├── rag_knowledge_base/      #   RAG retrieval documents
│   └── examples/                #   Example protocols
│
├── scripts/                     # Entry Scripts
│   ├── run_planner.py           #   Single protocol generation
│   └── batch_generate_protocols_parallel.py
│
├── utils/                       # Shared Utilities
├── requirements.txt
├── .env.example
└── LICENSE

Usage

Protocol Synthesis (LabForge)

Generate a structured protocol from a natural language experiment description:

python scripts/run_planner.py "Boc Deprotection of Hydrazine using TFA in DCM"

Batch Processing

# Generate protocols for all 30 experiments
python scripts/batch_generate_protocols_parallel.py \
    --experiments data/experiments.json \
    --output-dir output/protocols/

Layout Generation & Evaluation (Coming Soon)

LabGen and LabTouchstone will be released soon. Stay tuned!

Evaluation Metrics

LabTouchstone evaluates layouts across four dimensions:

Dimension	Metrics	Description
Geometric Compliance	Obj, OB, CN	Asset count, boundary violations, collisions
Feasibility Success Rate	Asset, Nav	Asset availability, navigation feasibility
Chemical Safety	Flam, Store, Incomp, Glass	Flammable isolation, storage, incompatibility
Semantic Plausibility	Real, Lay, Comp	Realism, layout rationality, completeness

Citation

If you find this work useful, please cite:

@article{cao2026labbuilder,
  title={LabBuilder: Protocol-Grounded 3D Layout Generation for Interactable and Safe Laboratory},
  author={Cao, Jianbao and Zhao, Zhangrui and Feng, Bohan and Hu, Zixuan and Li, Rui and Wan, Haiyuan and Li, Chenxi and Li, Jingyuan and Cai, Wenzhe and Bai, Lei and others},
  journal={arXiv preprint arXiv:2605.02288},
  year={2026}
}

License

• Code — Released under the MIT License.
• Data Assets — Released under the CC BY-NC 4.0 License. Free to use for research and educational purposes only.

Acknowledgements

This work was supported by Shanghai AI Laboratory, Wuhan University, Beihang University, Peking University, Tsinghua University, Shanghai Jiao Tong University, and The Chinese University of Hong Kong.