– The Benchmark Library

CARL (context adaptive RL) provides highly configurable contextual extensions to several well-known RL environments. It's designed to test your agent's generalization capabilities in all scenarios where intra-task generalization is important.

Feel free to check out our paper and our short blog post!

Benchmarks

Benchmarks include:

OpenAI gym classic control suite extended with several physics context features like gravity or friction
OpenAI gym Box2D BipedalWalker, LunarLander and CarRacing, each with their own modification possibilities like new vehicles to race
All Brax locomotion environments with exposed internal features like joint strength or torso mass
Super Mario (TOAD-GAN), a procedurally generated jump'n'run game with control over level similarity
dm_control, environments based on the MuJoCo physics engine. The environments are extended with different context features.

For more information, check out our documentation!

Installation

We recommend you use a virtual environment (e.g. Anaconda) to install CARL and its dependencies. We recommend and test with python 3.9 under Linux.

First, clone our repository and install the basic requirements:

git clone https://github.com/automl/CARL.git --recursive
cd CARL
pip install .

This will only install the basic classic control environments, which should run on most operating systems. For the full set of environments, use the install options:

pip install -e .[box2d, brax, mario, dm_control]

These may not be compatible with Windows systems. Box2D environment may need to be installed via conda on MacOS systems:

conda install -c conda-forge gym-box2d

In general, we test on Linux systems, but aim to keep the benchmark compatible with MacOS as much as possible. Mario at this point, however, will not run on any operation system besides Linux

To install the additional requirements for ToadGAN:

javac carl/envs/mario/Mario-AI-Framework/**/*.java

CARL's Contextual Extension

CARL contextually extends the environment by making the context visible and configurable. During training we therefore can encounter different contexts and train for generalization. We exemplarily show how Brax' Fetch is extended and embedded by CARL. Different instiations can be achieved by setting the context features to different values.

Cite Us

If you use CARL in your research, please cite our paper on the benchmark:

@inproceedings{BenEim2021a,
    title     = {CARL: A Benchmark for Contextual and Adaptive Reinforcement Learning},
    author    = {Carolin Benjamins and Theresa Eimer and Frederik Schubert and André Biedenkapp and Bodo Rosenhahn and Frank Hutter and Marius Lindauer},
    booktitle = {NeurIPS 2021 Workshop on Ecological Theory of Reinforcement Learning},
    year      = {2021},
    month     = dec
}

You can find the code and experiments for this paper in the neurips_ecorl_workshop_2021 branch.

References

OpenAI gym, Brockman et al., 2016. arXiv preprint arXiv:1606.01540

Brax -- A Differentiable Physics Engine for Large Scale Rigid Body Simulation, Freeman et al., NeurIPS 2021 (Dataset & Benchmarking Track)

TOAD-GAN: Coherent Style Level Generation from a Single Example, Awiszus et al., AIIDE 2020

dm_control: Software and Tasks for Continuous Control

License

CARL falls under the Apache License 2.0 (see file 'LICENSE') as is permitted by all work that we use. This includes CARLMario, which is not based on the Nintendo Game, but on TOAD-GAN and TOAD-GUI running under an MIT license. They in turn make use of the Mario AI framework (https://github.com/amidos2006/Mario-AI-Framework). This is not the original game but a replica, explicitly built for research purposes and includes a copyright notice (https://github.com/amidos2006/Mario-AI-Framework#copyrights ).

macmacal / CARL