Skip to content
Branch: master
Find file History

Unity ML-Agents Gym Wrapper

A common way in which machine learning researchers interact with simulation environments is via a wrapper provided by OpenAI called gym. For more information on the gym interface, see here.

We provide a gym wrapper and instructions for using it with existing machine learning algorithms which utilize gyms. Both wrappers provide interfaces on top of our UnityEnvironment class, which is the default way of interfacing with a Unity environment via Python.


The gym wrapper can be installed using:

pip install gym_unity

or by running the following from the /gym-unity directory of the repository:

pip install .

Using the Gym Wrapper

The gym interface is available from gym_unity.envs. To launch an environment from the root of the project repository use:

from gym_unity.envs import UnityEnv

env = UnityEnv(environment_filename, worker_id, use_visual, uint8_visual, multiagent)
  • environment_filename refers to the path to the Unity environment.

  • worker_id refers to the port to use for communication with the environment. Defaults to 0.

  • use_visual refers to whether to use visual observations (True) or vector observations (False) as the default observation provided by the reset and step functions. Defaults to False.

  • uint8_visual refers to whether to output visual observations as uint8 values (0-255). Many common Gym environments (e.g. Atari) do this. By default they will be floats (0.0-1.0). Defaults to False.

  • multiagent refers to whether you intent to launch an environment which contains more than one agent. Defaults to False.

  • flatten_branched will flatten a branched discrete action space into a Gym Discrete. Otherwise, it will be converted into a MultiDiscrete. Defaults to False.

  • allow_multiple_visual_obs will return a list of observation instead of only one if disabled. Defaults to False.

The returned environment env will function as a gym.

For more on using the gym interface, see our Jupyter Notebook tutorial.


  • It is only possible to use an environment with a single Brain.
  • By default, the first visual observation is provided as the observation, if present. Otherwise, vector observations are provided. You can receive all visual observations by using the allow_multiple_visual_obs=True option in the gym parameters. If set to True, you will receive a list of observation instead of only the first one.
  • All BrainInfo output from the environment can still be accessed from the info provided by env.step(action).
  • Stacked vector observations are not supported.
  • Environment registration for use with gym.make() is currently not supported.

Running OpenAI Baselines Algorithms

OpenAI provides a set of open-source maintained and tested Reinforcement Learning algorithms called the Baselines.

Using the provided Gym wrapper, it is possible to train ML-Agents environments using these algorithms. This requires the creation of custom training scripts to launch each algorithm. In most cases these scripts can be created by making slight modifications to the ones provided for Atari and Mujoco environments.

Example - DQN Baseline

In order to train an agent to play the GridWorld environment using the Baselines DQN algorithm, you first need to install the baselines package using pip:

pip install git+git://

Next, create a file called Then create an /envs/ directory and build the GridWorld environment to that directory. For more information on building Unity environments, see here. Add the following code to the file:

import gym

from baselines import deepq
from baselines import logger

from gym_unity.envs.unity_env import UnityEnv

def main():
    env = UnityEnv("./envs/GridWorld", 0, use_visual=True, uint8_visual=True)
    logger.configure('./logs') # Çhange to log in a different directory
    act = deepq.learn(
        "cnn", # conv_only is also a good choice for GridWorld
        checkpoint_path='./logs', # Change to save model in a different directory
    print("Saving model to unity_model.pkl")"unity_model.pkl")

if __name__ == '__main__':

To start the training process, run the following from the directory containing

python -m train_unity

Other Algorithms

Other algorithms in the Baselines repository can be run using scripts similar to the examples from the baselines package. In most cases, the primary changes needed to use a Unity environment are to import UnityEnv, and to replace the environment creation code, typically gym.make(), with a call to UnityEnv(env_path) passing the environment binary path.

A typical rule of thumb is that for vision-based environments, modification should be done to Atari training scripts, and for vector observation environments, modification should be done to Mujoco scripts.

Some algorithms will make use of make_env() or make_mujoco_env() functions. You can define a similar function for Unity environments. An example of such a method using the PPO2 baseline:

from gym_unity.envs import UnityEnv
from baselines.common.vec_env.subproc_vec_env import SubprocVecEnv
from baselines.bench import Monitor
from baselines import logger
import baselines.ppo2.ppo2 as ppo2

import os

    from mpi4py import MPI
except ImportError:
    MPI = None

def make_unity_env(env_directory, num_env, visual, start_index=0):
    Create a wrapped, monitored Unity environment.
    def make_env(rank, use_visual=True): # pylint: disable=C0111
        def _thunk():
            env = UnityEnv(env_directory, rank, use_visual=use_visual, uint8_visual=True)
            env = Monitor(env, logger.get_dir() and os.path.join(logger.get_dir(), str(rank)))
            return env
        return _thunk
    if visual:
        return SubprocVecEnv([make_env(i + start_index) for i in range(num_env)])
        rank = MPI.COMM_WORLD.Get_rank() if MPI else 0
        return make_env(rank, use_visual=False)

def main():
    env = make_unity_env('./envs/GridWorld', 4, True)

if __name__ == '__main__':

Run Google Dopamine Algorithms

Google provides a framework Dopamine, and implementations of algorithms, e.g. DQN, Rainbow, and the C51 variant of Rainbow.
Using the Gym wrapper, we can run Unity environments using Dopamine.

First, after installing the Gym wrapper, clone the Dopamine repository.

git clone

Then, follow the appropriate install instructions as specified on Dopamine's homepage. Note that the Dopamine guide specifies using a virtualenv. If you choose to do so, make sure your unity_env package is also installed within the same virtualenv as Dopamine.

Adapting Dopamine's Scripts

First, open dopamine/atari/ Alternatively, copy the entire atari folder, and name it something else (e.g. unity). If you choose the copy approach, be sure to change the package names in the import statements in to your new directory.

Within, we will need to make changes to which environment is instantiated, just as in the Baselines example. At the top of the file, insert

from gym_unity.envs import UnityEnv

to import the Gym Wrapper. Navigate to the create_atari_environment method in the same file, and switch to instantiating a Unity environment by replacing the method with the following code.

    game_version = 'v0' if sticky_actions else 'v4'
    full_game_name = '{}NoFrameskip-{}'.format(game_name, game_version)
    env = UnityEnv('./envs/GridWorld', 0, use_visual=True, uint8_visual=True)
    return env

./envs/GridWorld is the path to your built Unity executable. For more information on building Unity environments, see here, and note the Limitations section below.

Note that we are not using the preprocessor from Dopamine, as it uses many Atari-specific calls. Furthermore, frame-skipping can be done from within Unity, rather than on the Python side.


Since Dopamine is designed around variants of DQN, it is only compatible with discrete action spaces, and specifically the Discrete Gym space. For environments that use branched discrete action spaces (e.g. VisualBanana), you can enable the flatten_branched parameter in UnityEnv, which treats each combination of branched actions as separate actions.

Furthermore, when building your environments, ensure that your Learning Brain is using visual observations with greyscale enabled, and that the dimensions of the visual observations is 84 by 84 (matches the parameter found in and Dopamine's agents currently do not automatically adapt to the observation dimensions or number of channels.


The hyperparameters provided by Dopamine are tailored to the Atari games, and you will likely need to adjust them for ML-Agents environments. Here is a sample dopamine/agents/rainbow/configs/rainbow.gin file that is known to work with GridWorld.

import dopamine.agents.rainbow.rainbow_agent
import dopamine.unity.run_experiment
import dopamine.replay_memory.prioritized_replay_buffer

RainbowAgent.num_atoms = 51
RainbowAgent.stack_size = 1
RainbowAgent.vmax = 10.
RainbowAgent.gamma = 0.99
RainbowAgent.update_horizon = 3
RainbowAgent.min_replay_history = 20000  # agent steps
RainbowAgent.update_period = 5
RainbowAgent.target_update_period = 50  # agent steps
RainbowAgent.epsilon_train = 0.1
RainbowAgent.epsilon_eval = 0.01
RainbowAgent.epsilon_decay_period = 50000  # agent steps
RainbowAgent.replay_scheme = 'prioritized'
RainbowAgent.tf_device = '/cpu:0'  # use '/cpu:*' for non-GPU version
RainbowAgent.optimizer = @tf.train.AdamOptimizer()

tf.train.AdamOptimizer.learning_rate = 0.00025
tf.train.AdamOptimizer.epsilon = 0.0003125

Runner.game_name = "Unity" # any name can be used here
Runner.sticky_actions = False
Runner.num_iterations = 200
Runner.training_steps = 10000  # agent steps
Runner.evaluation_steps = 500  # agent steps
Runner.max_steps_per_episode = 27000  # agent steps

WrappedPrioritizedReplayBuffer.replay_capacity = 1000000
WrappedPrioritizedReplayBuffer.batch_size = 32

This example assumed you copied atari to a separate folder named unity. Replace unity in import dopamine.unity.run_experiment with the folder you copied your and files to. If you directly modified the existing files, then use atari here.

Starting a Run

You can now run Dopamine as you would normally:

python -um dopamine.unity.train \
  --agent_name=rainbow \
  --base_dir=/tmp/dopamine \

Again, we assume that you've copied atari into a separate folder. Remember to replace unity with the directory you copied your files into. If you edited the Atari files directly, this should be atari.

Example: GridWorld

As a baseline, here are rewards over time for the three algorithms provided with Dopamine as run on the GridWorld example environment. All Dopamine (DQN, Rainbow, C51) runs were done with the same epsilon, epsilon decay, replay history, training steps, and buffer settings as specified above. Note that the first 20000 steps are used to pre-fill the training buffer, and no learning happens.

We provide results from our PPO implementation and the DQN from Baselines as reference. Note that all runs used the same greyscale GridWorld as Dopamine. For PPO, num_layers was set to 2, and all other hyperparameters are the default for GridWorld in trainer_config.yaml. For Baselines DQN, the provided hyperparameters in the previous section are used. Note that Baselines implements certain features (e.g. dueling-Q) that are not enabled in Dopamine DQN.

Dopamine on GridWorld

Example: VisualBanana

As an example of using the flatten_branched option, we also used the Rainbow algorithm to train on the VisualBanana environment, and provide the results below. The same hyperparameters were used as in the GridWorld case, except that replay_history and epsilon_decay were increased to 100000.

Dopamine on VisualBanana

You can’t perform that action at this time.