NL2PDDL : Large Language Models as Planning Domain Generators

Code and data for the paper Large Language Models as Planning Domain Generators accepted for publication at ICAPS 2024.

Installation and Setup

For automatic setup use the setup.sh script. This will init submodules, create a virtual environment, and install all python dependencies and build other dependencies (VAL).

Alternatively you can use the manual setup instructions.

Manual Setup Instructions

Clone the repo

Clone the repository with submodules

git clone --recurse-submodules <Link to this repo>

If you have already cloned without submodules, you can get the submodules with

git submodule update --init --recursive

Create a virtual python environment (Recommended)

We suggest creating a virtual python environment for the project. This will help ensure that the large list of pip dependencies required for this project do not conflict with alternate versions on your system.

Here is an example using python's built in venv module, but you could also use anaconda or pipenv.

# Create the virtual environment 
python -m venv .venv
# Activate the virtual environment
source .venv/bin/activate

Once the virtual environment is activated you are free to install pip dependencies.

Install Required Python Packages

Note that all code in this repository has only been tested on Python 3.11.4

pip install -r requirements.txt 

Build VAL into /VAL/build/bin

The VAL submodule must be built into /VAL/build/bin. To build VAL we need CMake and Make. You can run the following commands to build it.

cd VAL
mkdir build
cd build
cmake ..
make

Setup Credentials for Model Inference (Required For New Generation)

We use IBM Generative AI Python SDK. Create a .env file in the top level repo with the following credentials.

GENAI_KEY=<your GENAI key>
GENAI_API=<your API endpoint>

Note: Models supported by known ibm-genai endpoints change frequently, you may need to change models evaluated against or host a custom endpoint.

Running Experiments

Use driver.py to run the experiments on the raw LLM outputs located in data/outputs. It will generate a .json file with new output at each stage of the pipeline, and end by generating the figures and tables used by the paper in figures_and_tables.

Make sure to activate your virtual environment first if necessary!

source .venv/bin/activate
python driver.py