Skip to content

ShaoZhang0115/COLE-Platform

 
 

Repository files navigation

COLE: Cooperative Open-ended Learning Framework for Zero-shot Human-AI Coordination

Welcome to the COLE repository! The official GitHub repository is for our series work "Cooperative Open-ended Learning Framework for Zero-shot Coordination" (accepted by ICML2023) and "Tackling Cooperative Incompatibility for Zero-Shot Human-AI Coordination". You can access our hands-on demo page. Below, you'll find a list of distinct features that our repository offers.

Update: New Features are Available:

    1. Support Human-Human Experiments (human branch)
    1. Play with LLM Agent like GPT-4 (human branch)
    1. Training your own COLE_SV agent (cole_training branch)
    1. ZSC baseline agents including SP, FCP, PBT, MEP (baseline_training branch).

Features

This repository presents a human-AI evaluation platform centered on the popular game Overcooked 2, created specifically to facilitate experiments involving human-AI interaction. Overcooked is an engaging, fully cooperative game that requires two players to work in concert. The architecture of the system is outlined below.

Here, you're granted the ability to:

  • Upload your weights (main branch)
  • Customize the human questionnaire (main branch)
  • Configure game settings (main branch)
  • Human-Human Play (human branch)
  • Play with LLM agent like gpt-4 (human branch)
  • Training your own COLE_SV agent (cole_training branch) and ZSC baseline agents including SP, FCP, PBT, MEP (baseline_training branch)
  • And many more!

The usage of each ability, please refer to the corresponding branch.

This repository introduces a human-AI evaluation platform built around the Overcooked game, designed to support Human-AI experiments. Overcooked, a two-player fully cooperative game. The system is shown as follows.


Here, you're granted the ability to:
  • Upload your weights (main branch)
  • Customize the human questionnaire (main branch)
  • Configure game settings (main branch)
  • Play with LLM agent like gpt-4 (human branch)
  • Training your own COLE_SV agent (cole branch) and ZSC baseline agents including SP, FCP, PBT, MEP (baseline branch)
  • And many more!

Getting Started

1. How to setup

Install PantheonRL in this repo

   conda create -n overcooked-vis python=3.7
   conda activate overcooked-vis
   pip install -r requirements.txt
   pip install -e .

Install mpi4py

conda install mpi4py

Install PyTorch (based on your CUDA version): https://pytorch.org/ (You don't actually need the GPU version to run the game)

Install human_aware_rl and its dependencies: overcooked_ai, baselines & stable_baselines

   cd overcookedgym/human_aware_rl
   pip install -e .
   cd overcooked_ai
   pip install -e .
   cd ..
   cd stable-baselines
   pip install -e .
   cd ..
   cd baselines
   pip install -e .

Here are instructions for building using npm.

You need to firstly install npm (if you can't do this, you can checkout our history version to get all built files)

If you want to utilize human-human play or llm agents, please refer to human branch and remember that you need to re-run npm run build in overcookedgym/overcooked-flask

In overcookedgym/human_aware_rl/overcooked_ai/overcooked_ai_js, run

npm install
sudo npm install browserify

Then

npm run build
npm run build-window

to build overcooked_ai game core.

Then, in overcookedgym/overcooked-flask

sed -i 's/overcook\"/overcooked\"/g' ../human_aware_rl/overcooked_ai/overcooked_ai_js/package.json
wget https://raw.githubusercontent.com/HumanCompatibleAI/overcooked_ai/37d14dd48ae93ad0363610a0a370221c47a79eb2/overcooked_ai_js/js/mdp.es6 -O ../human_aware_rl/overcooked_ai/overcooked_ai_js/js/mdp.es6
wget https://raw.githubusercontent.com/HumanCompatibleAI/overcooked_ai/37d14dd48ae93ad0363610a0a370221c47a79eb2/overcooked_ai_js/js/task.es6 -O ../human_aware_rl/overcooked_ai/overcooked_ai_js/js/task.es6
    
npm install
npm link ../human_aware_rl/overcooked_ai/overcooked_ai_js/

npm run build

2. How to load models

You need to put your model file in ./models. You can get our trained models here, including BC, self-play, population-based training, FCP, MEP, COLE.

Also, you need to put this BC data in ./data.

Note: The layout names in code and google drive are not aligned with the layout names in those papers. Here is the mapping:

PYTHON_LAYOUT_NAME_TO_ENV_NAME = {
    "unident_s": "Asymmetric Advantages",
    "simple": "Cramped Room",
    "random1": "Coordination Ring",
    "random0": "Forced Coordination",
    "random3": "Counter Circuit"
}

In addition, you can load your own models if they are trained using the Human-Aware-RL framework. Agents are loaded using the get_agent_from_saved_model() method, which loads tensorflow predictor models (.pb files), so you should save your agents in this style if you wish to load them into our framework. You can reference to the save method in human_aware_rl/pbt/pbt.py for saving agents that can be loaded.

To load your own models, you need to put them in the ./models folder in a named folder (the folder names need to be the same for all layouts), and the models would be loaded upon starting the server. For example. If your algo is named ABC, then the folder structure should look like this:

-- models
  | --  simple
       | -- SP        <---- Baseline 1 
       | -- PBT       <---- Baseline 2
          ...
       | -- ABC       <---- Your Algorithm
  | --  unident_s
       | -- SP        <---- Baseline 1 
       | -- PBT       <---- Baseline 2
          ...
       | -- ABC       <---- Your Algorithm
  | --  random1
       | -- SP        <---- Baseline 1 
       | -- PBT       <---- Baseline 2
          ...
       | -- ABC       <---- Your Algorithm
  ...

3. How to run

python overcookedgym/overcooked-flask/app.py --trajs_savepath ./trajs --ckpts ./models
  • --ckpts: Folder containing all the AI models to be loaded. Default is ./models.
  • --port: The port where you run the server process.
  • --trajs_savepath: Optional trajectory save path, default is ./trajs.
  • --questionnaire_savepath: Optional questionnaire save path, default is ./questionnaire.
  • --ip: Default is LOCALHOST, we recommend you replace it with your public network IP, because of a known bug of Flask that may cause extreme lag when playing the game. The same applies when debugging, you should visit your machine's IP in your browser instead of LOCALHOST.

4. How to customize

Customize experiment statements

You can replace configs/statement.md by your experiment statement markdown file, then restarting your web process.

Customize before game questionnaire.

You can modify configs/before_game.yaml to customize your settings of before game questionnaire.

5. How to collect data

Questionnaire data are saved in ./questionnaire, its corresponging co-play trajectorys is saved in ./trajs.

We also privide a simple data processing scripts named questionnaire_analyze.ipynb.

License

MIT License

Citation

Please cite

@inproceedings{10.5555/3618408.3619252,
author = {Li, Yang and Zhang, Shao and Sun, Jichen and Du, Yali and Wen, Ying and Wang, Xinbing and Pan, Wei},
title = {Cooperative Open-Ended Learning Framework for Zero-Shot Coordination},
year = {2023},
publisher = {JMLR.org},
booktitle = {Proceedings of the 40th International Conference on Machine Learning},
articleno = {844},
numpages = {15},
location = {Honolulu, Hawaii, USA},
series = {ICML'23}
}
@misc{li2023tackling,
      title={Tackling Cooperative Incompatibility for Zero-Shot Human-AI Coordination}, 
      author={Yang Li and Shao Zhang and Jichen Sun and Wenhao Zhang and Yali Du and Ying Wen and Xinbing Wang and Wei Pan},
      year={2023},
      eprint={2306.03034},
      archivePrefix={arXiv},
      primaryClass={cs.AI}
}
@inproceedings{sarkar2022pantheonrl,
 title={PantheonRL: A MARL Library for Dynamic Training Interactions},
 author={Sarkar, Bidipta and Talati, Aditi and Shih, Andy and Sadigh, Dorsa},
 booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
 volume={36},
 number={11},
 pages={13221--13223},
 year={2022}
}
@article{carroll2019utility,
 title={On the utility of learning about humans for human-ai coordination},
 author={Carroll, Micah and Shah, Rohin and Ho, Mark K and Griffiths, Tom and Seshia, Sanjit and Abbeel, Pieter and Dragan, Anca},
 journal={Advances in neural information processing systems},
 volume={32},
 year={2019}
}

About

Overcooked human-AI experiment platform

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages

  • Python 64.6%
  • HTML 17.4%
  • TypeScript 11.4%
  • JavaScript 5.8%
  • Shell 0.5%
  • Starlark 0.2%
  • Other 0.1%