Status: Maintenance (expect bug fixes and minor updates)
OpenAI Gym is a toolkit for developing and comparing reinforcement learning algorithms. This is the gym open-source library, which gives you access to a standardized set of environments.
gym makes no assumptions about the structure of your agent, and is compatible with any numerical computation library, such as TensorFlow or Theano. You can use it from Python code, and soon from other languages.
If you're not sure where to start, we recommend beginning with the docs on our site. See also the FAQ.
A whitepaper for OpenAI Gym is available at http://arxiv.org/abs/1606.01540, and here's a BibTeX entry that you can use to cite it in a publication:
@misc{1606.01540,
Author = {Greg Brockman and Vicki Cheung and Ludwig Pettersson and Jonas Schneider and John Schulman and Jie Tang and Wojciech Zaremba},
Title = {OpenAI Gym},
Year = {2016},
Eprint = {arXiv:1606.01540},
}
Contents of this document
There are two basic concepts in reinforcement learning: the environment (namely, the outside world) and the agent (namely, the algorithm you are writing). The agent sends actions to the environment, and the environment replies with observations and rewards (that is, a score).
The core gym interface is Env, which is
the unified environment interface. There is no interface for agents;
that part is left to you. The following are the Env methods you
should know:
- reset(self): Reset the environment's state. Returns observation.
- step(self, action): Step the environment by one timestep. Returns observation, reward, done, info.
- render(self, mode='human', close=False): Render one frame of the environment. The default mode will do something human friendly, such as pop up a window. Passing the close flag signals the renderer to close any such windows.
You can perform a minimal install of gym with:
git clone https://github.com/openai/gym.git
cd gym
pip install -e .If you prefer, you can do a minimal install of the packaged version directly from PyPI:
pip install gymYou'll be able to run a few environments right away:
- algorithmic
- toy_text
- classic_control (you'll need
pygletto render though)
We recommend playing with those environments at first, and then later installing the dependencies for the remaining environments.
To install the full set of environments, you'll need to have some system packages installed. We'll build out the list here over time; please let us know what you end up installing on your platform. Also, take a look at the docker files (test.dockerfile.xx.xx) to see the composition of our CI-tested images.
On OSX:
brew install cmake boost boost-python sdl2 swig wgetOn Ubuntu 14.04 (non-mujoco only):
apt-get install libjpeg-dev cmake swig python-pyglet python3-opengl libboost-all-dev \
libsdl2-2.0.0 libsdl2-dev libglu1-mesa libglu1-mesa-dev libgles2-mesa-dev \
freeglut3 xvfb libav-toolsOn Ubuntu 16.04:
apt-get install -y python-pyglet python3-opengl zlib1g-dev libjpeg-dev patchelf \
cmake swig libboost-all-dev libsdl2-dev libosmesa6-dev xvfb ffmpegOn Ubuntu 18.04:
apt install -y python3-dev zlib1g-dev libjpeg-dev cmake swig python-pyglet python3-opengl libboost-all-dev libsdl2-dev \
libosmesa6-dev patchelf ffmpeg xvfbMuJoCo has a proprietary dependency we can't set up for you. Follow
the
instructions
in the mujoco-py package for help.
Once you're ready to install everything, run pip install -e '.[all]' (or pip install 'gym[all]').
We currently support Linux and OS X running Python 2.7 or 3.5. Some users on OSX + Python3 may need to run
brew install boost-python --with-python3If you want to access Gym from languages other than python, we have limited support for non-python frameworks, such as lua/Torch, using the OpenAI Gym HTTP API.
To run pip install -e '.[all]', you'll need a semi-recent pip.
Please make sure your pip is at least at version 1.5.0. You can
upgrade using the following: pip install --ignore-installed
pip. Alternatively, you can open setup.py and
install the dependencies by hand.
If you're trying to render video on a server, you'll need to connect a
fake display. The easiest way to do this is by running under
xvfb-run (on Ubuntu, install the xvfb package):
xvfb-run -s "-screen 0 1400x900x24" bashIf you'd like to install the dependencies for only specific environments, see setup.py. We maintain the lists of dependencies on a per-environment group basis.
The code for each environment group is housed in its own subdirectory gym/envs. The specification of each task is in gym/envs/__init__.py. It's worth browsing through both.
These are a variety of algorithmic tasks, such as learning to copy a sequence.
import gym
env = gym.make('Copy-v0')
env.reset()
env.render()The Atari environments are a variety of Atari video games. If you didn't do the full install, you can install dependencies via pip install -e '.[atari]' (you'll need cmake installed) and then get started as follow:
import gym
env = gym.make('SpaceInvaders-v0')
env.reset()
env.render()This will install atari-py, which automatically compiles the Arcade Learning Environment. This can take quite a while (a few minutes on a decent laptop), so just be prepared.
Box2d is a 2D physics engine. You can install it via pip install -e '.[box2d]' and then get started as follow:
import gym
env = gym.make('LunarLander-v2')
env.reset()
env.render()These are a variety of classic control tasks, which would appear in a typical reinforcement learning textbook. If you didn't do the full install, you will need to run pip install -e '.[classic_control]' to enable rendering. You can get started with them via:
import gym
env = gym.make('CartPole-v0')
env.reset()
env.render()MuJoCo is a physics engine which can do
very detailed efficient simulations with contacts. It's not
open-source, so you'll have to follow the instructions in mujoco-py
to set it up. You'll have to also run pip install -e '.[mujoco]' if you didn't do the full install.
import gym
env = gym.make('Humanoid-v2')
env.reset()
env.render()MuJoCo is a physics engine which can do
very detailed efficient simulations with contacts and we use it for all robotics environments. It's not
open-source, so you'll have to follow the instructions in mujoco-py
to set it up. You'll have to also run pip install -e '.[robotics]' if you didn't do the full install.
import gym
env = gym.make('HandManipulateBlock-v0')
env.reset()
env.render()You can also find additional details in the accompanying technical report and blog post. If you use these environments, you can cite them as follows:
@misc{1802.09464,
Author = {Matthias Plappert and Marcin Andrychowicz and Alex Ray and Bob McGrew and Bowen Baker and Glenn Powell and Jonas Schneider and Josh Tobin and Maciek Chociej and Peter Welinder and Vikash Kumar and Wojciech Zaremba},
Title = {Multi-Goal Reinforcement Learning: Challenging Robotics Environments and Request for Research},
Year = {2018},
Eprint = {arXiv:1802.09464},
}
Toy environments which are text-based. There's no extra dependency to install, so to get started, you can just do:
import gym
env = gym.make('FrozenLake-v0')
env.reset()
env.render()See the examples directory.
- Run examples/agents/random_agent.py to run an simple random agent.
- Run examples/agents/cem.py to run an actual learning agent (using the cross-entropy method).
- Run examples/scripts/list_envs to generate a list of all environments.
We are using pytest for tests. You can run them via:
pytest- 2019-02-06 (v0.11.0)
- remove gym.spaces.np_random common PRNG; use per-instance PRNG instead.
- support for kwargs in gym.make
- lots of bugfixes
2018-02-28: Release of a set of new robotics environments.
2018-01-25: Made some aesthetic improvements and removed unmaintained parts of gym. This may seem like a downgrade in functionality, but it is actually a long-needed cleanup in preparation for some great new things that will be released in the next month.
- Now your Env and Wrapper subclasses should define step, reset, render, close, seed rather than underscored method names.
- Removed the board_game, debugging, safety, parameter_tuning environments since they're not being maintained by us at OpenAI. We encourage authors and users to create new repositories for these environments.
- Changed MultiDiscrete action space to range from [0, ..., n-1] rather than [a, ..., b-1].
- No more render(close=True), use env-specific methods to close the rendering.
- Removed scoreboard directory, since site doesn't exist anymore.
- Moved gym/monitoring to gym/wrappers/monitoring
- Add dtype to Space.
- Not using python's built-in module anymore, using gym.logger
2018-01-24: All continuous control environments now use mujoco_py >= 1.50. Versions have been updated accordingly to -v2, e.g. HalfCheetah-v2. Performance should be similar (see openai#834) but there are likely some differences due to changes in MuJoCo.
2017-06-16: Make env.spec into a property to fix a bug that occurs when you try to print out an unregistered Env.
2017-05-13: BACKWARDS INCOMPATIBILITY: The Atari environments are now at v4. To keep using the old v3 environments, keep gym <= 0.8.2 and atari-py <= 0.0.21. Note that the v4 environments will not give identical results to existing v3 results, although differences are minor. The v4 environments incorporate the latest Arcade Learning Environment (ALE), including several ROM fixes, and now handle loading and saving of the emulator state. While seeds still ensure determinism, the effect of any given seed is not preserved across this upgrade because the random number generator in ALE has changed. The *NoFrameSkip-v4 environments should be considered the canonical Atari environments from now on.
2017-03-05: BACKWARDS INCOMPATIBILITY: The configure method has been removed from Env. configure was not used by gym, but was used by some dependent libraries including universe. These libraries will migrate away from the configure method by using wrappers instead. This change is on master and will be released with 0.8.0.
2016-12-27: BACKWARDS INCOMPATIBILITY: The gym monitor is now a wrapper. Rather than starting monitoring as env.monitor.start(directory), envs are now wrapped as follows: env = wrappers.Monitor(env, directory). This change is on master and will be released with 0.7.0.
2016-11-1: Several experimental changes to how a running monitor interacts with environments. The monitor will now raise an error if reset() is called when the env has not returned done=True. The monitor will only record complete episodes where done=True. Finally, the monitor no longer calls seed() on the underlying env, nor does it record or upload seed information.
2016-10-31: We're experimentally expanding the environment ID format to include an optional username.
2016-09-21: Switch the Gym automated logger setup to configure the root logger rather than just the 'gym' logger.
2016-08-17: Calling close on an env will also close the monitor and any rendering windows.
2016-08-17: The monitor will no longer write manifest files in real-time, unless write_upon_reset=True is passed.
2016-05-28: For controlled reproducibility, envs now support seeding (cf #91 and #135). The monitor records which seeds are used. We will soon add seed information to the display on the scoreboard.