gym-dmc, OpenAI Gym Plugin for DeepMind Control Suite

Update Log

• 2023-06-14: Test gym (V0.21.0) compatibility, Embed gym_dmc into rl-baselines3-zoo (v1.8.0)
• ========== ZZM Matain ↑ ==========
• 2022-01-13: Add space_dtype for overriding the dtype for the state and action spaces. Default to None, need to set to float/np.float32 for pytorch_SAC implementation.
• 2022-01-11: Added a env._get_obs() method to allow one to obtain the observation after resetting the environment. Version: v0.2.1

How To Use

Install:

download https://github.com/TneitapSimHand/gym-dmc.git as zip

pip install gym-dmc.zip

Usage pattern:

import gym
import gym_dmc # perform registeration with 'env make'
env = gym.make("dmc:Pendulum-swingup-v1")

For the full list of environments, you can print:

from dm_control.suite import ALL_TASKS

print(*ALL_TASKS, sep="\n")

# Out[2]: ('acrobot', 'swingup')
#         ('acrobot', 'swingup_sparse')
...

We register all of these environments using the following pattern:

acrobot task "swingup_sparse" becomes dmc:Acrobot-swingup_sparse-v1

You can see the usage pattern in ./specs/test_gym_dmc.py:

import gym, gym_dmc
env = gym.make('dmc:Walker-walk-v1', frame_skip=4, space_dtype=np.float32)
assert env.action_space.dtype is np.float32
assert env.observation_space.dtype is np.float32

env = gym.make('dmc:Walker-walk-v1', frame_skip=4)
assert env._max_episode_steps == 250
assert env.reset().shape == (24,)

env = gym.make('dmc:Walker-walk-v1', from_pixels=True, frame_skip=4)
assert env._max_episode_steps == 250

env = gym.make('dmc:Cartpole-balance-v1', from_pixels=True, frame_skip=8)
assert env._max_episode_steps == 125
assert env.reset().shape == (3, 84, 84)

env = gym.make('dmc:Cartpole-balance-v1', from_pixels=True, frame_skip=8, channels_first=False)
assert env._max_episode_steps == 125
assert env.reset().shape == (84, 84, 3)

env = gym.make('dmc:Cartpole-balance-v1', from_pixels=True, frame_skip=8, channels_first=False, gray_scale=True)
assert env._max_episode_steps == 125
assert env.reset().shape == (84, 84, 1)

Note, the max_episode_steps is calculated based on the frame_skip. All DeepMind control domains terminate after 1000 simulation steps. So for frame_skip=4, the max_episode_steps should be 250.

Built with ❤️ by Ge Yang