This is a packaged version of the distracting_control suite from Stone et al. We provide OpenAI gym bindings, to make the original code base easier to use.

pip install distracting_control

Then in your python script:

import gym

env = gyn.make('gdc:Hopper-hop-easy-v1', from_pixel=True)
obs = env.reset()
doc.figure(obs, "figures/hopper_readme.png")

Take a look at the test file in the ./specs folder, and the source code. DeepMind control has a lot of low level binding burried in the source code.

def test_max_episode_steps():
    env = gym.make('distracting_control:Walker-walk-easy-v1')
    assert env._max_episode_steps == 250


def test_flat_obs():
    env = gym.make('distracting_control:Walker-walk-easy-v1', frame_skip=4)
    env.env.env.env.observation_spec()
    assert env.reset().shape == (24,)


def test_frame_skip():
    env = gym.make('distracting_control:Walker-walk-easy-v1', from_pixels=True, frame_skip=8)
    assert env._max_episode_steps == 125


def test_channel_first():
    env = gym.make('distracting_control:Walker-walk-easy-v1', from_pixels=True, channels_first=True)
    assert env.reset().shape == (3, 84, 84)


def test_channel_last():
    env = gym.make('distracting_control:Walker-walk-easy-v1', from_pixels=True, frame_skip=8, channels_first=False)
    assert env._max_episode_steps == 125
    assert env.reset().shape == (84, 84, 3)

1. [planned] remove tensorflow dependency
2. [planned] increase ground floor transparency in Hopper

distracting_control extends dm_control with static or dynamic visual distractions in the form of changing colors, backgrounds, and camera poses. Details and experimental results can be found in our paper.

• Clone this repository
• sh run.sh
• Follow the instructions and install dm_control. Make sure you setup your MuJoCo keys correctly.
• Download the DAVIS 2017 dataset. Make sure to select the 2017 TrainVal - Images and Annotations (480p). The training images will be used as distracting backgrounds.

• You can run the distracting_control_demo to generate sample images of the different tasks at different difficulties:

  python distracting_control_demo --davis_path=$HOME/DAVIS/JPEGImages/480p/
  --output_dir=/tmp/distrtacting_control_demo
  

• As seen from the demo to generate an instance of the environment you simply need to import the suite and use suite.load while specifying the dm_control domain and task, then choosing a difficulty and providing the dataset_path.

• Note the environment follows the dm_control environment APIs.

If you use this code, please cite the accompanying paper as:

@article{stone2021distracting,
      title={The Distracting Control Suite -- A Challenging Benchmark for Reinforcement Learning from Pixels}, 
      author={Austin Stone and Oscar Ramirez and Kurt Konolige and Rico Jonschkowski},
      year={2021},
      journal={arXiv preprint arXiv:2101.02722},
}

This is not an official Google product.