The MultiGrid library provides contains a collection of fast multi-agent discrete gridworld environments for reinforcement learning in Gymnasium. This is a multi-agent extension of the minigrid library, and the interface is designed to be as similar as possible.
The environments are designed to be fast and easily customizable. Compared to minigrid, the underlying gridworld logic is significantly optimized, with environment simulation 10x to 20x faster by our benchmarks.
Documentation for this library can be found at ini.github.io/docs/multigrid.
pip install multigrid
Or alternatively, for an editable install:
git clone https://github.com/ini/multigrid
cd multigrid
pip install -e .
This package requires Python 3.9 or later.
The multigrid.envs package provides implementations of several multi-agent environments. You can find the full list here.
MultiGrid follows the same pattern as RLlib's MultiAgentEnv API and PettingZoo's ParallelEnv API.
import gymnasium as gym
import multigrid.envs
env = gym.make('MultiGrid-Empty-8x8-v0', agents=2, render_mode='human')
observations, infos = env.reset()
while not env.is_done():
# this is where you would insert your policy / policies
actions = {agent.index: agent.action_space.sample() for agent in env.agents}
observations, rewards, terminations, truncations, infos = env.step(actions)
env.close()More information about using MultiGrid directly with other APIs:
See the scripts folder for an example training with RLlib.
Documentation for this package can be found at ini.github.io/docs/multigrid.
To cite this project please use:
@article{oguntola2023theory,
title={Theory of mind as intrinsic motivation for multi-agent reinforcement learning},
author={Oguntola, Ini and Campbell, Joseph and Stepputtis, Simon and Sycara, Katia},
journal={arXiv preprint arXiv:2307.01158},
year={2023}
}