BTGenBot-2

AIRLab-POLIMI/llama-3.2-1b-it-ft-lora-bt is the fine-tuned Llama 3.2 1B Instruct model released with BTGenBot-2: Efficient Behavior Tree Generation with Small Language Models.

BTGenBot-2 generates executable robot Behavior Trees from:

1. a natural-language task description, and
2. a list of available robot action primitives.

The model outputs XML Behavior Trees compatible with BehaviorTree.CPP, supporting ROS 2 robotics behavior-tree pipelines.

For the complete project, examples, code, dataset, benchmark, and paper, visit:

👉 https://airlab-polimi.github.io/BTGenBot-2/

Model Details

• Developed by: AIRLab, Politecnico di Milano
• Authors: Riccardo Andrea Izzo, Gianluca Bardaro, Matteo Matteucci
• Base model: meta-llama/Llama-3.2-1B-Instruct
• Model type: Small language model for Behavior Tree generation
• Fine-tuning method: QLoRA / LoRA parameter-efficient fine-tuning
• Input: Natural-language robot task + available robot action primitives
• Output: XML Behavior Tree compatible with BehaviorTree.CPP
• Language: English
• Project page: https://airlab-polimi.github.io/BTGenBot-2/
• Paper: https://arxiv.org/abs/2602.01870
• Code: https://github.com/AIRLab-POLIMI/BTGenBot-2

Intended Use

This model is intended to generate robot Behavior Trees for research and development in robotics task planning.

Example applications include:

• ROS 2 / Nav2-compatible task planning;
• navigation Behavior Tree generation;
• manipulation Behavior Tree generation;
• simulation-based robot-task validation;
• benchmarking language-model-based Behavior Tree generation.

Input Format

The recommended input format is:

Task:
Describe the robot task in natural language.

Actions:
[ActionName(parameters: parameter_1, parameter_2), AnotherAction(parameters: parameter_1)]

Output Format

The model is expected to return XML only:

<root BTCPP_format="4">
  <BehaviorTree ID="MainTree">
    ...
  </BehaviorTree>
</root>

Training Data

BTGenBot-2 was trained on a synthetic instruction-following dataset of 5,204 natural-language instruction / Behavior Tree pairs.

Each sample contains:

• instruction: system-level instructions for Behavior Tree generation;
• input: task description and available robot actions;
• output: XML Behavior Tree.

The dataset was generated from real Behavior Trees and expanded through controlled synthetic generation.

See the full project page for details:

https://airlab-polimi.github.io/BTGenBot-2/

Training Procedure

The model was fine-tuned from meta-llama/Llama-3.2-1B-Instruct using QLoRA / LoRA.

Reported training details from the paper include:

• Train/test split: 95% / 5%
• Learning rate: 1e-4
• Warmup ratio: 0.1
• Batch size: 16
• Training duration: approximately 30 hours
• Hardware: 2 × NVIDIA RTX Quadro 6000 GPUs, 48 GB total VRAM

Citation

@article{izzo2026btgenbot,
  title={BTGenBot-2: Efficient Behavior Tree Generation with Small Language Models},
  author={Izzo, Riccardo Andrea and Bardaro, Gianluca and Matteucci, Matteo},
  journal={arXiv preprint arXiv:2602.01870},
  year={2026}
}

More Information

• Project page: https://airlab-polimi.github.io/BTGenBot-2/
• Code: https://github.com/AIRLab-POLIMI/BTGenBot-2
• Paper: https://arxiv.org/abs/2602.01870