🌎GPTWorld: An Experimental AI-engined Sandbox World

NOTE: We are now integrated into AgentVerse!

🌎GPTWorld is an experimental multi-agent sandbox world. Unlike typical sandbox games, interactions in GPTWorld are not fully defined by pre-written scripts 📜 and rules📏, but rather inferred by a world engine based on large models 🤖. In this world, the whole environment, including agents and objects, is supported by large models, providing great autonomy and diversity.

The creation of a sandbox world is often considered to require advanced game development techniques that are beyond the reach of most individuals. This is where 🌎GPTWorld comes to the rescue! 🛠️ Our goal is to make the process of creating a diverse range of sandboxes more accessible by enabling players to construct custom worlds using simple configuration files or even natural language 🤩.

Through this easy-to-use world creation feature, we hope to inspire everyone to construct their own unique worlds. As a result, a vibrant sandbox environment community will emerge, allowing intelligent agents to explore various creations through community sharing 🤝.

👨‍💻👩‍💻 Join us on this exciting journey of creating a groundbreaking sandbox world with limitless possibilities 🚀!

Project Status

1. Currently, 🌍GPTWorld Supports:

👨‍💻 Create with JSON: Easily creating your environment and entities using JSON and mounting them onto your world

🏃 Start World Engine: Starting the environment's autonomous operation with just one command.

👀 Demo: Observing the behavior of the agent in your web browser.

2. Experimental Feature

🗣️ Create with NL: Creating environments with natural languages.

3. TODOS!

❗️ Polishing: Feeling that the current behavior of agents is not very reasonable? Don't worry, we are improving it!

❗️ Refactoring: We are also using AgentVerse to refactor the behavior of agents!

🤝 Distributed: Allowing players to build and share environments in a distributed way.

Usage

Installation

We strongly recommend to use python=3.8 on all platforms.

1. Download project by

git clone https://github.com/ShengdingHu/GPT-World.git

2. (Optional) Create a python virtual enviornment by

conda create -n gptworld python=3.8

• Please note that Python 3.8 provides the best support for flask-socketio, whereas other versions of Python may currently encounter issues with web display. *

3. Go to project directory by

cd GPT-World

4. Install dependencies by

pip install -r requirements.txt

5. Then install this library by

pip install .

6. Build the front end

cd io/frontend/rpg-game/
npm install
npm run build

Run Examples

We currently provide some example sandboxes in world_instances/, choose one that you want to launch. Take alice_home as an example.

1. start the world engine

python gptworld/run.py -W alice_home

2. start the web server

python io/app.py -W alice_home

now open the 5001 port of your localhost(http://0.0.0.0:5001/), and you will be able to see a wonderful environment full of GPT-drived agents and objects!

If http://0.0.0.0:5001/ is not working, you may try http://127.0.0.1:5001/.

(🧪Experimental) Create Your Own World

If you want to create your world instance, run gptworld/create_world.py -W <world_instance> -I <info>, where <world_instance> stands for the name of the world directory in the ./world_instances/ directory, and <info> stands for the initial information about the world to be created. Remember to add specific information into the <info> argument, the more detailed, the better.

e.g. -W neighborhood -I "A neighborhood with some Chinese restaurants (with a cook, tables, chairs, delicious foods), and coffee bar (coffee machine, coffee staff), 711 shop (stocks, drinks, snacks, etc), trees, grass area (with flower, old people, etc), roads (with some cars, bikes), Alice's house and Bob's house (includes some furniture , like sofa, bed, bookshelf, cooker, and their families, a yard with two apple trees, etc), all above should be areas, each task could be a subtask, assign the jobs with lots of details, no overlapping areas. Add some people and animals and plants, evenly distributed. Expect more details. Areas should not overlap with each other."

Global Configs

1. Log Level:

export GPTWORLD_LOG_LEVEL=<log_level>

<log_level> can be DEBUG, INFO, WARNING ...) to set the logging level, DEBUG will have a lot of print in command line :

2. OpenAI Key

export OPENAI_API_KEY=<your_openai_api_key>

<your_openai_api_key> refers to your own api key to access to the ChatGPT.

In Windows OS, please add environment variables instead of export methods.

🙏Acknowledgements

This project was greatly inspired by Generative Agents: Interactive Simulacra of Human Behavior during development, where the mechanism of agents' behavior takes the reflection-summary-plan framework.

🚨 Disclaimer:

This project is for academic and experimental purposes only. We currently suspect that it is far from a usable game product.

Contributor

Made with contrib.rocks.