Surreal 0.2.1

Stanford University Repository for Reinforcement Algorithms

**`SURREAL <https://surreal.stanford.edu>`__**
==============================================

| `About <#open-source-distributed-reinforcement-learning-framework>`__
| `Installation <#installation>`__
| `Benchmarking <#benchmarking>`__
| `Citation <#citation>`__

Open-Source Distributed Reinforcement Learning Framework
--------------------------------------------------------

*Stanford Vision and Learning Lab*

`SURREAL <https://surreal.stanford.edu>`__ is a fully integrated
framework that runs state-of-the-art distributed reinforcement learning
(RL) algorithms.

.. raw:: html

<div align="center">

.. raw:: html

</div>

- **Scalability**. RL algorithms are data hungry by nature. Even the
simplest Atari games, like Breakout, typically requires up to a
billion frames to learn a good solution. To accelerate training
significantly, SURREAL parallelizes the environment simulation and
learning. The system can easily scale to thousands of CPUs and
hundreds of GPUs.

- **Flexibility**. SURREAL unifies distributed on-policy and off-policy
learning into a single algorithmic formulation. The key is to
separate experience generation from learning. Parallel actors
generate massive amount of experience data, while a *single,
centralized* learner performs model updates. Each actor interacts
with the environment independently, which allows them to diversify
the exploration for hard long-horizon robotic tasks. They send the
experiences to a centralized buffer, which can be instantiated as a
FIFO queue for on-policy mode and replay memory for off-policy mode.

.. raw:: html

<!--<img src=".README_images/distributed.png" alt="drawing" width="500" />-->

- **Reproducibility**. RL algorithms are notoriously hard to reproduce
[Henderson et al., 2017], due to multiple sources of variations like
algorithm implementation details, library dependencies, and hardware
types. We address this by providing an *end-to-end integrated
pipeline* that replicates our full cluster hardware and software
runtime setup.

.. raw:: html

<!--<img src=".README_images/pipeline.png" alt="drawing" height="250" />-->

Installation
------------

| Surreal algorithms can be deployed at various scales. It can run on a
single laptop and solve easier locomotion tasks, or run on hundreds of
machines to solve complex manipulation tasks.
| \* `Surreal on your Laptop <docs/surreal_subproc.md>`__ \* `Surreal on
Google Cloud Kubenetes Engine <docs/surreal_kube_gke.md>`__
| \* `Customizing Surreal <docs/contributing.md>`__
| \* `Documentation Index <docs/index.md>`__

Benchmarking
------------

- Scalability of Surreal-PPO with up to 1024 actors on Surreal Robotics
Suite.

.. figure:: .README_images/scalability-robotics.png
:alt:

- Training curves of 16 actors on OpenAI Gym tasks for 3 hours,
compared to other baselines.

Citation
--------

Please cite our CORL paper if you use this repository in your
publications:

::

@inproceedings{corl2018surreal,
title={SURREAL: Open-Source Reinforcement Learning Framework and Robot Manipulation Benchmark},
author={Fan, Linxi and Zhu, Yuke and Zhu, Jiren and Liu, Zihua and Zeng, Orien and Gupta, Anchit and Creus-Costa, Joan and Savarese, Silvio and Fei-Fei, Li},
booktitle={Conference on Robot Learning},
year={2018}
}