CARL provides highly configurable contextual extensions to several well-known RL environments.

Last update: Dec 28, 2022

Related tags

Overview

The CARL 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.

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
RNADesign, an environment for RNA design given structure constraints with structures from different datasets to choose from

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, rna, mario]

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 src/envs/mario/Mario-AI-Framework/**/*.java

If you want to use the RNA design environment:

cd src/envs/rna/learna
make requirements
make data

In case you want to run our experiments or use our training files, also install the experiment dependencies:

pip install -e .[experiments]

Train an Agent

To get started with CARL, you can use our 'train.py' script. It will train a PPO agent on the environment of your choice with custom context variations that are sampled from a standard deviation.

To use MetaCartPole with variations in gravity and friction by 20% compared to the default, run:

python train.py 
--env CARLCartPoleEnv 
--context_args gravity friction
--default_sample_std_percentage 0.2
--outdir <result_location>

You can use the plotting scripts in src/eval to view the results.

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

@misc{CARL,
  author    = {C. Benjamins and 
               T. Eimer and 
               F. Schubert and 
               A. Biedenkapp and 
               B. Rosenhahn and 
               F. Hutter and 
               M. Lindauer},
  title     = {CARL: A Benchmark for Contextual and Adaptive Reinforcement Learning},
  howpublished = {https://github.com/automl/CARL},
  year      = {2021},
  month     = aug,
}

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

Learning to Design RNA, Runge et al., ICRL 2019

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 ).

Comments

Rna fixup

RNA is now better documented and more easily runnable. There's also an option to subsample the datasets instead of always using all instances per context.

The thing that's missing right now are more context options like filtering by solvers or GC-content, but those aren't easily extractable from our data right now, so that's a separate work package all together.

opened by TheEimer 6
Gym 0.22.0
update required minimum gym version number

added pygame as a requirement because it is not picked up by the gym requirements

getting rid of CustomBipedalWalkerEnv because the functionality of changing the gravity is covered by CARLEnv (same for CustomLunarLanderEnv)

add high game over penalty for LunarLander by a wrapper
opened by benjamc 6
Instance selection

Instance selection now is a class. Default is still roundrobin selection. An instance is only selected when env.reset() (or to be more specific, _progress_instance() is called.

opened by benjamc 4
Added Encoders

Context encoders have been added as a folder and an experiment for running the encoder added in the experiments folders. Since the working directory is the experiment one, I had to add an absolute path for the saved weights. This might need to be changed in the config file

opened by amsks 4
Update References with correct conference

Thanks for the pointer to the survey, but it hasn't been published anywhere, so that detail is incorrect (I wouldn't want to claim that it's published somewhere when it isn't).

opened by RobertKirk 3
Performance Deviations in Brax

Comparing HalfCheetah in Brax (via gym.make and then wrapped as here: https://github.com/google/brax/blob/main/notebooks/training_torch.ipynb) vs in CARL makes a big difference in return even when the context is kept static. Do we do any unexpected reward normalization? Does the way we reset the env make a difference compared to theirs (as we actually update the simluation)?
bug

opened by TheEimer 2
Integrate DM Control
[ ] (convert test file to jupyter notebook. I would like to keep that)

[ ] check tests / write more to increase coverage

[x] update README.md

[x] update documentation

[x] add dm_control to requirements

[x] support dict observation space

documentation tests
opened by benjamc 2
Fix gym version

Gym released a new version where the signature of the step function has changed. This affects our code and requires a separate PR. For now, fix the gym version.

opened by benjamc 1
Initial statedistrs #48
#48 Make initial state distribution configurable. So far, only uniform distributions are used and the bounds can be adjusted.

Classic control:

[x] Acrobot

[x] Pendulum

[x] MountainCar (normal distribution instead of uniform)

[x] MountainCarContinuous (uniform distribution)

[x] CartPole

Box2d

[x] LunarLander

[ ] (maybe/later) Make distributions fully configurable by passing the distribution class and its parameters.

[x] Update documentation: Contexts are automatically filled with the default context if underspecified.
opened by benjamc 1
Integrate dmcontrol

Add support for dm control environments. Integrated walker, quadruped and fish.

In dmc environments there is an additional setting for the context, namely the context mask, which can reduce the amount of context features.

opened by sebidoe 1
use appropriate library for building states

So far, when we do not hide the context, we concatenate the context to the state. For jax based environments (brax) this means that the state is converted from a jax to a numpy array. Now, the state builder checks which library to use and keeps jax states as jax arrays and numpy states as numpy arrays.

Noticed in #42.

opened by benjamc 1
AttributeError: 'System' object has no attribute 'body_idx' in brax

when running test/test_all_envs.py, there is AttributeError: 'System' object has no attribute 'body_idx' in carl_fetch and carl_humanoid environments.

opened by andy-james0310 3

Releases(v0.2.0)

v0.2.0(Jul 12, 2022)
Integrate dm control environments (#55)

Add context masks to only append those to the state (#54)

Extend classic control environments to parametrize initial state distributions (#52)

Remove RNA environment for maintenance (#61)

Fixed pre-commit (mypy, black, flake8, isort) (#62)

Source code(tar.gz)
Source code(zip)

Owner

AutoML-Freiburg-Hannover

GitHub Repository

(EI 2022) Controllable Confidence-Based Image Denoising

Image Denoising with Control over Deep Network Hallucination Paper and arXiv preprint -- Our frequency-domain insights derive from SFM and the concept

5 Dec 18, 2022

It is a simple library to speed up CLIP inference up to 3x (K80 GPU)

CLIP-ONNX It is a simple library to speed up CLIP inference up to 3x (K80 GPU) Usage Install clip-onnx module and requirements first. Use this trick !

93 Dec 20, 2022

From a body shape, infer the anatomic skeleton.

OSSO: Obtaining Skeletal Shape from Outside (CVPR 2022) This repository contains the official implementation of the skeleton inference from: OSSO: Obt

166 Dec 28, 2022

Here is the implementation of our paper S2VC: A Framework for Any-to-Any Voice Conversion with Self-Supervised Pretrained Representations.

S2VC Here is the implementation of our paper S2VC: A Framework for Any-to-Any Voice Conversion with Self-Supervised Pretrained Representations. In thi

81 Dec 15, 2022

Deal or No Deal? End-to-End Learning for Negotiation Dialogues

Introduction This is a PyTorch implementation of the following research papers: (1) Hierarchical Text Generation and Planning for Strategic Dialogue (

1.4k Dec 29, 2022

Faster RCNN with PyTorch

Faster RCNN with PyTorch Note: I re-implemented faster rcnn in this project when I started learning PyTorch. Then I use PyTorch in all of my projects.

1.6k Dec 23, 2022

Reducing Information Bottleneck for Weakly Supervised Semantic Segmentation (NeurIPS 2021)

Reducing Information Bottleneck for Weakly Supervised Semantic Segmentation (NeurIPS 2021) The implementation of Reducing Infromation Bottleneck for W

81 Dec 16, 2022

PyTorch implementation of CDistNet: Perceiving Multi-Domain Character Distance for Robust Text Recognition

PyTorch implementation of CDistNet: Perceiving Multi-Domain Character Distance for Robust Text Recognition The unofficial code of CDistNet. Now, we ha

25 Jul 20, 2022

My solution for the 7th place / 245 in the Umoja Hack 2022 challenge

Umoja Hack 2022 : Insurance Claim Challenge My solution for the 7th place / 245 in the Umoja Hack 2022 challenge Umoja Hack Africa is a yearly hackath

17 Jun 03, 2022

ParaGen is a PyTorch deep learning framework for parallel sequence generation

ParaGen is a PyTorch deep learning framework for parallel sequence generation. Apart from sequence generation, ParaGen also enhances various NLP tasks, including sequence-level classification, extrac

169 Dec 22, 2022

The official implementation of ELSA: Enhanced Local Self-Attention for Vision Transformer

ELSA: Enhanced Local Self-Attention for Vision Transformer By Jingkai Zhou, Pich

87 Dec 19, 2022

Experimental solutions to selected exercises from the book [Advances in Financial Machine Learning by Marcos Lopez De Prado]

Advances in Financial Machine Learning Exercises Experimental solutions to selected exercises from the book Advances in Financial Machine Learning by

1.4k Jan 04, 2023