LaMMA-P: Generalizable Multi-Agent Long-Horizon Task Allocation and Planning with LM-Driven PDDL Planner

This is the official repository for the LaMMA-P codebase. It includes instructions for configuring and running LaMMA-P on the MAT-THOR datasets in the AI2-THOR simulator. It is accepted as a conference paper by the IEEE International Conference on Robotics and Automation (ICRA), Atlanta, 2025.

Project Website | Paper | Video

Abstract: Language models (LMs) possess a strong capability to comprehend natural language, making them effective in translating human instructions into detailed plans for simple robot tasks. Nevertheless, it remains a significant challenge to handle long-horizon tasks, especially in subtask identification and allocation for cooperative heterogeneous robot teams. To address this issue, we propose a Language Model-Driven Multi-Agent PDDL Planner (LaMMA-P), a novel multi-agent task planning framework that achieves state-of-the-art performance on long-horizon tasks. LaMMA-P integrates the strengths of the LMs’ reasoning capability and the traditional heuristic search planner to achieve a high success rate and efficiency while demonstrating strong generalization across tasks. Additionally, we create MAT-THOR, a comprehensive benchmark that features household tasks with two different levels of complexity based on the AI2-THOR environment. The experimental results demonstrate that LaMMA-P achieves a 105% higher success rate and 36% higher efficiency than existing LM-based multi-agent planners.

Code Organization

Below are the details of various important directories

• resources/: Contains robot definitions and PDDL domain files
• scripts/: Main execution scripts adapted from SMART-LLM
• data/: Test datasets and example tasks extended from SMART-LLM
• downward/: Fast Downward planner from Fast Downward

Datasets

The repository includes various commands and robots with different skill sets for heterogeneous robot tasks:

• Test tasks: data/final_test/
• Robot definitions: resources/robots.py
• Floor plans: Refer to AI2Thor Demo for layouts

Environment Setup

1. Environment Setup

Create a conda environment (or virtualenv):

conda create -n lammap python==3.9
conda activate lammap

Install dependencies:

pip install -r requirements.txt

2. Fast Downward Planner Setup

The project requires the Fast Downward Planner. Follow these steps to set it up:

1. Clone the Fast Downward repository as a submodule:

git submodule update --init --recursive
cd downward

2. Build the planner:

./build.py

3. Verify the installation:

./fast-downward.py --help

3. OpenAI API Setup

The code relies on OpenAI's API for LLM functionality. To set this up:

1. Create an API Key at https://platform.openai.com/
2. Create a file named api_key.txt in the root folder
3. Paste your OpenAI API Key in the file

Quickstart

1. Generate PDDL Plans

To generate PDDL plans for tasks in AI2Thor floor plans, run:

python scripts/pddlrun_llmseparate.py --floor-plan <floor_plan_no>

Additional parameters:

• --gpt-version: Choose between 'gpt-3.5-turbo', 'gpt-4o', 'gpt-3.5-turbo-16k' (default: 'gpt-4o')
• --prompt-decompse-set: Set decomposition prompt set (default: 'pddl_train_task_decomposesep')
• --prompt-allocation-set: Set allocation prompt set (default: 'pddl_train_task_allocationsep')

The script will:

1. Decompose the high-level task into subtasks
2. Generate PDDL problem files for each subtask
3. Run the Fast Downward planner on each subtask
4. Combine the solutions into a complete plan

Output files are stored in the logs directory, organized by timestamp and task name.

2. Execute Plans in AI2Thor

To execute the generated plans in the AI2Thor environment:

python scripts/execute_plan.py --command <log_folder_name>

Replace <log_folder_name> with the specific folder name in the logs directory containing your generated plan.

Citation

If you find this work useful for your research, please consider citing:

@inproceedings{zhang2025lamma,
  title={LaMMA-P: Generalizable Multi-Agent Long-Horizon Task Allocation and Planning with LM-Driven PDDL Planner},
  author={Zhang, Xiaopan and Qin, Hao and Wang, Fuquan and Dong, Yue and Li, Jiachen},
  journal={2025 IEEE International Conference on Robotics and Automation (ICRA)},
  year={2025},
  organization={IEEE}
}

Acknowledgement

We sincerely thank the researchers and developers for SMART-LLM, AI2THOR, and Fast Downward for their amazing work.