DELTA: Decomposed Efficient Long-Term Robot Task Planning using Large Language Models

This repository is the python implementation for the paper DELTA: Decomposed Efficient Long-Term Robot Task Planning using Large Language Models, which is accepted at IEEE International Conference on Robotics & Automation (ICRA) 2025.

DELTA uses pre-trained LLMs (GPT and Llama families) to automatically:

Generate a PDDL domain file given a natural language description of the domain knowledge
Prune the 3D Scene Graph (3DSG) by only keeping the task-relevant items
Generate a PDDL problem file given a natural language description of the task goal and a 3DSG representing the environment
Decompose the long-term task goal into a sequence of sub-goals, such that an automated task planner can solve the corresponding sub-problems more efficiently

The code is implemented in Python 3.8.17.

Project webpage: https://delta-llm.github.io/

Usage from source

Create a virtual environment

Assuming you have already installed Anaconda, run the following command to create a conda environment with Python 3.8:

conda create -n delta python=3.8

Access to the LLMs

The code inferences the GPT models via Azure OpenAI Service. For using the Azure OpenAI API, please first apply for the access, then obtain the end point and API key from Azure OpenAI Studio. Create a .env file in the root folder and place the your key and end point there:

AZURE_API_KEY = "YOUR_AZURE_API_KEY"
AZURE_ENDPOINT = "https://YOUR_AZURE_ENDPOINT.openai.azure.com/"

If your wish to use the regular OpenAI API, please adjust the corresponding part in the __init()__ function of the GPT class in file llm.py.

To use the Llama models, create a Hugging Face account and getting access to Hugging Face Llama 3.1 repository. Generate access tokens and use it to log in to your Hugging Face account:

pip install -U "huggingface_hub[cli]"
huggingface-cli login --token YOUR_HF_TOKEN

By default the models weights will be downloaded and stored in ~/.cache/huggingface/. If you wish to store the models weights in a different directory, please execute the following command:

export HF_HOME=/YOUR_DESIRED_DIR/

Installing dependencies

First execute pip install -r requirements.txt to install the required packages.

Clone the PDDLGym Planners repository and install from source.

pip install -e .

Install the PDDL validator VAL with the following commands (reference here):

git clone https://github.com/KCL-Planning/VAL.git
VAL/scripts/linux/build_linux64.sh all Release
sudo tar -xzvf VAL/build/linux64/Release/Val--Linux.tar.gz -C /usr/local --strip-components=1

Running experiments of DELTA

For reproducing the results of DELTA with all evaluation domains, please run the bash script:

bash scripts/run_delta.sh

You may also directly run the source file of DELTA python delta.py -h with the -h flag to see more options for specifying the experiments, LLMs, domains, problems, and other parameters.

Running experiments of other baselines

The scripts for evaluating other baselines can also be found in the /scripts folder. You can find the source files of the baselines in the /baselines folder.

Evaluation results

After running the experiments, the log files are saved in the /result folder. Execute the following command to print the evaluation results (replace YOUR_LOG_PATH with the desired path of the log files):

python eval.py -f result/YOUR_LOG_PATH/log.csv

Maintainers and Contributors

Current maintainer: Yuchen Liu

Contributor: Luigi Palmieri

3rd Party Licenses

This package depends on an open-source dataset and several open-source software. Following a list of the external dataset and software libraries:

3DSceneGraph LLM-GenPlan

Citation

@article{liu2024delta,
  title={Delta: Decomposed efficient long-term robot task planning using large language models},
  author={Liu, Yuchen and Palmieri, Luigi and Koch, Sebastian and Georgievski, Ilche and Aiello, Marco},
  journal={arXiv preprint arXiv:2404.03275},
  year={2024}
}

DELTA

Install / Use

README