Multi Task RL Algorithms

1. Introduction

2. Setup

3. Usage

4. Documentation

5. Contributing to MTRL

6. Community

7. Acknowledgements

MTRL is a library of multi-task reinforcement learning algorithms. It has two main components:

• Building blocks and agents that implement the multi-task RL algorithms.

• Experiment setups that enable training/evaluation on different setups.

Together, these two components enable use of MTRL across different environments and setups.

• Learning Robust State Abstractions for Hidden-Parameter Block MDPs
• Multi-Task Reinforcement Learning with Context-based Representations
  • We use the af8417bfc82a3e249b4b02156518d775f29eb289 commit for the MetaWorld environments for our experiments.

• MTRL uses MIT License.

• Terms of Use

• Privacy Policy

If you use MTRL in your research, please use the following BibTeX entry:

@Misc{Sodhani2021MTRL,
  author =       {Shagun Sodhani and Amy Zhang},
  title =        {MTRL - Multi Task RL Algorithms},
  howpublished = {Github},
  year =         {2021},
  url =          {https://github.com/facebookresearch/mtrl}
}

• Clone the repository: git clone git@github.com:facebookresearch/mtrl.git.

• Install dependencies: pip install -r requirements/dev.txt

• MTRL supports 8 different multi-task RL algorithms as described here.

• MTRL supports multi-task environments using MTEnv. These environments include MetaWorld and multi-task variants of DMControl Suite

• Refer the tutorial to get started with MTRL.

https://mtrl.readthedocs.io

There are several ways to contribute to MTRL.

1. Use MTRL in your research.

2. Contribute a new algorithm. We currently support 8 multi-task RL algorithms and are looking forward to adding more environments.

3. Check out the good-first-issues on GitHub and contribute to fixing those issues.

4. Check out additional details here.

Ask questions in the chat or github issues:

• Chat
• Issues

• Our implementation of SAC is inspired by Denis Yarats' implementation of SAC.
• Project file pre-commit, mypy config, towncrier config, circleci etc are based on same files from Hydra.