Deepmind Control Suite

February 9, 2023 ยท View on GitHub

Tasks

We follow the settings in PAD and experiment with 8 tasks from DMControl. We measure generalization to (1) randomized colors of the background and robot itself, and (2) natural videos as dynamic background. SECANT significantly outperforms prior SOTA in all but one task, often by substantial margins up to 88.3%.

Installation

Please refer to Installation.

Usage

Use secant.envs.dm_control.make_dmc() to create a standardized DMControl Gym environment with image observation modality.

from secant.envs.dm_control import make_dmc

env = make_dmc(
    task="walker_walk",
    background="original",
)
env.reset()

done = False
while not done:
    action = env.action_space.sample()
    obs, reward, done, info = env.step(action)

Important Note:

  1. task can be set to one of ["cheetah_run", "ball_in_cup_catch", "cartpole_swingup", "cartpole_balance", "walker_walk", "walker_stand", "finger_spin", "reacher_easy"]

  2. background can be set to one of ["original", "color_easy", "color_hard", "video[0-9]"].

  3. The returned observation is a uint8 numpy array of the image observation. Don't forget to scale it to floats between [0., 1.)

  4. The created environment instance has properties observation_space and action_space. Please refer to OpenAI Gym's API. Don't forget to clamp your actions to fit the allowable action_space.

  5. All supported tasks are recorded in secant.envs.dm_control.ALL_TASKS. They can be specified either as a tuple ("cartpole", "balance") or as a single string "cartpole_balance".