Common sense-based Open-World Planning (COWP)

Please note that OpenAI regularly updates their APIs for accessing GPT responses. If the API mentioned in this repository is outdated, kindly update the corresponding APIs on your own.

Colab Version (new)

We highly recommend using the Colab version.

One of main difference is it generates PDDL code using a third-party package called pddl (https://github.com/AI-Planning/pddl).

By utilizing this package, it becomes simpler to inject knowledge from LLMs into the classical planner.

Environment

ubuntu (version 16.04, 18.04, and 20.04 have been tested)

Install GPT-Planner

step 1: download GPT-Planner

open terminal
cd
git clone https://github.com/yding25/GPT-Planner.git

Note that please place the GPT-planner folder in your home folder.

step 2: install dependencies

pip install numpy==1.19.5
pip install openai==0.8.0

step 3: install FastDownward (detailed instruction at https://www.fast-downward.org/ObtainingAndRunningFastDownward)

cd GPT-Planner
git clone https://github.com/aibasel/downward.git FastDownward
sudo apt install cmake g++ git make python3
cd FastDownward
python build.py
cd ..

step 4: prepare an GPT-3 API key

Here is a brief instruction to obtain a key.

log in https://beta.openai.com/
click the icon ``Personal'' on the top right
click ``view API keys''
click ``+ Create new secret key''
copy the key and paste it in GPT-Planner/dataset/openai_api_key.txt

Run GPT-Planner

change the value of task_id (line 19 in main.py file) to select an task

task_id = 4
task = {
    1: 'cleaning floor',
    4: 'drinking water',
    6: 'setting table',
    9: 'grasping plate',
    10: 'drinking soda',
    11: 'eating dinner'
    }
task_name = task[task_id]

python main.py

Here are one trial result, where the left and right figures show the situation and solution, respectively. The situation is "the cup_1 is missing", and the solution is that GPT-3 suggests using a drinking glasss to replace the cup in the task.

Experiment Results

All experiment results (log files) can be found in the folder "/GPT-Planner/results". For example, the below figure shows situation0 in task4 has been handled in some trials, and some are not.

Citation and Reference

@article{ding2023integrating,
  title={Integrating Action Knowledge and LLMs for Task Planning and Situation Handling in Open Worlds},
  author={Ding, Yan and Zhang, Xiaohan and Amiri, Saeid and Cao, Nieqing and Yang, Hao and Kaminski, Andy and Esselink, Chad and Zhang, Shiqi},
  journal={arXiv preprint arXiv:2305.17590},
  year={2023}
}
@article{ding2022robot,
  title={Robot task planning and situation handling in open worlds},
  author={Ding, Yan and Zhang, Xiaohan and Amiri, Saeid and Cao, Nieqing and Yang, Hao and Esselink, Chad and Zhang, Shiqi},
  journal={arXiv preprint arXiv:2210.01287},
  year={2022}
}

xuejianyong/GPT-Planner