LLM+P: Empowering Large Language Models with Optimal Planning Proficiency

This repo contains the source code for making plans based on problems decribed by natural language.

Dependency

1. Install OpenAI GPT API. Remember to put openai_keys under the keys folder.

2. Install fast-downward. For more details on fast-downward, please check the official github repo and the fast-downward website.

Running Code

To run a for a specific task in a specific domain using a specific method:

python main.py --domain DOMAIN --method METHOD --task TASK_ID

DOMAIN is selected from [barman, blocksworld, floortile, grippers, storage, termes, tyreworld]

METHOD is selected from [llm_ic_pddl_planner, llm_pddl_planner, llm_planner, llm_ic_planner]

Alternatively, you can just use:

bash run.sh DOMAIN METHOD TASK_ID

Citations

Please cite this pre-print if you find this repo useful.

@article{liu2023llmp,
  title={LLM+P: Empowering Large Language Models with Optimal Planning Proficiency},
  author={Liu, Bo and Jiang, Yuqian and Zhang, Xiaohan and Liu, Qiang and Zhang, Shiqi and Biswas, Joydeep and Stone, Peter},
  journal={arXiv preprint arXiv:2304.11477},
  year={2023}
}

The File Hierarchy:

llm-pddl
 └─main.py                         (the main python script)
 └─keys
    └─ openai_keys.txt             (you should place your openai keys here, one line each)
 └─domains                         (the generated domain files)
    └─ barman
        └─ description_geneator.py (generating natural language description)
        └─ p_example.nl            (example natural language)
        └─ p_example.pddl          (example problem pddl file)
        └─ domain.pddl             (the shared domain.pddl file for all problems)
        └─ xxx.nl                  (task natural language description)
        └─ xxx.pddl                (ground-truth problem pddl, might not be used)
    └─ blocksworld
    └─ floortile
    └─ grippers
    └─ storage
    └─ termes
    └─ tyreworld
 └─problems                        (the generated problem pddl files)
    └─ llm                         (empty, since llm -> plan does not generate pddl)
    └─ llm_ic                      (empty, since llm + context -> plan does not generate pddl)    
    └─ llm_pddl                    (baseline 2: llm -> p.pddl)
    └─ llm_ic_pddl                 (ours: llm + context -> p.pddl)
        └─ barman
        └─ ...
 └─plans                           (the tmp folder for storing raw solutions found by fast-downward)
    └─ llm                         (empty, since llm -> plan does not generate raw plans)
    └─ llm_ic                      (empty, since llm + context -> plan does not generate raw plans)
    └─ llm_pddl                    (baseline 2: llm -> p.pddl)
    └─ llm_ic_pddl                 (ours: llm + context -> p.pddl)
        └─ barman
        └─ ...
 └─results                         (the final plan in natural language)
    └─ llm                         (baseline 1: llm -> plan)
    └─ llm_ic                      (baseline 3: llm + context -> plan)
    └─ llm_pddl                    (baseline 2: llm -> p.pddl)
    └─ llm_ic_pddl                 (ours: llm + context -> p.pddl)
        └─ barman
        └─ ...