Welcome to the MiniGrid Environment repository! This is a simple yet powerful environment designed for reinforcement learning agents. The goal of the game is to navigate a grid and reach a target location. The environment is highly customizable and can be used with both Fully Connected Networks (FCN) and Convolutional Neural Networks (CNN).
- Customizable Grid Size: You can easily adjust the size of the grid to challenge your agent with different levels of complexity.
- FCN and CNN Compatible: The environment is designed to work seamlessly with both FCN and CNN policies. This makes it a great playground to experiment with different types of neural networks.
- Human Render Mode: Watch your agent learn in real-time with the human render mode. It's not just about the numbers, it's also about the journey!
- OpenAI Gym / Gymnasium Compatible: MiniGrid follows the OpenAI Gym / Gymnasium interface, making it compatible with a wide range of reinforcement learning libraries and algorithms.
To use the MiniGrid environment, you can install it directly into your project using pip:
pip install gymnasium-minigridImport the MiniGridEnv class:
from gymnasium_minigrid import MiniGridEnvHere is a simple example of how to use the environment with a FCN policy:
from gymnasium_minigrid import MiniGridEnv
from stable_baselines3 import PPO
env = MiniGridEnv(size=5)
model = PPO("MlpPolicy", env, verbose=1)
model.learn(total_timesteps=10000)For detailed usage instructions and examples, please refer to the examples directory or check out our Colab Notebook.
The action space is discrete with 4 possible actions, corresponding to "right", "up", "left", "down".
The observation space includes PositionX, PositionY, TargetX, TargetY. If output_is_picture is set to true, the observation space is a 2D grid with 0 : empty, 1 self position, 2 target position
The agent receives a reward of 500 for reaching the target and a penalty proportional to the distance to the target for each step taken.
The episode terminates when the agent reaches the target.
We believe in the importance of testing. That's why we have included a suite of tests in the test directory. To run the tests, simply use the command pytest.
We have conducted a simple benchmark to compare the performance of different reinforcement learning algorithms on the MiniGrid environment. The following algorithms were tested:
- A2C
- DQN
- PPO
The benchmark was run with the default hyperparameters for each algorithm and was executed only once. The performance of each algorithm was evaluated every second for a total of 30 seconds. The graph below shows the evaluation score (y-axis) as a function of the total training time (x-axis).
Please note that this benchmark is not exhaustive and should be taken as an indicative comparison of the algorithms' performance. For a more rigorous evaluation, multiple runs with different random seeds and hyperparameter tuning should be performed. However, this graph provides a good representation of our experience with these algorithms on the MiniGrid environment.
We warmly welcome contributions from the community. If you have an idea for an improvement or found a bug, don't hesitate to open an issue or submit a pull request. Your input is greatly appreciated, and our project is made better by your participation!
If you find this repository useful, please give it a star!
This project is licensed under the MIT License. See the LICENSE file for details.