def make_environment(suite: str, task: str) -> dm_env.Environment:
"""Makes the requested continuous control environment.
Args:
suite: One of 'gym' or 'control'.
task: Task to load. If `suite` is 'control', the task must be formatted as
f'{domain_name}:{task_name}'
Returns:
An environment satisfying the dm_env interface expected by Acme agents.
"""
if suite not in _VALID_TASK_SUITES:
raise ValueError(
f'Unsupported suite: {suite}. Expected one of {_VALID_TASK_SUITES}'
)
if suite == 'gym':
env = gym.make(task)
# Make sure the environment obeys the dm_env.Environment interface.
env = wrappers.GymWrapper(env)
elif suite == 'control':
# Load dm_suite lazily not require Mujoco license when not using it.
from dm_control import suite as dm_suite # pylint: disable=g-import-not-at-top
domain_name, task_name = task.split(':')
env = dm_suite.load(domain_name, task_name)
env = wrappers.ConcatObservationWrapper(env)
# Wrap the environment so the expected continuous action spec is [-1, 1].
# Note: this is a no-op on 'control' tasks.
env = wrappers.CanonicalSpecWrapper(env, clip=True)
env = wrappers.SinglePrecisionWrapper(env)
return env
def make_environment(
task: str = 'MountainCarContinuous-v0') -> dm_env.Environment:
"""Creates an OpenAI Gym environment."""
# Load the gym environment.
environment = gym.make(task)
# Make sure the environment obeys the dm_env.Environment interface.
environment = wrappers.GymWrapper(environment)
# Clip the action returned by the agent to the environment spec.
environment = wrappers.CanonicalSpecWrapper(environment, clip=True)
environment = wrappers.SinglePrecisionWrapper(environment)
return environment
def make_environment(evaluation: bool = False,
task: str = 'HalfCheetah-v3') -> dm_env.Environment:
"""Creates an OpenAI Gym environment."""
del evaluation
# Load the gym environment.
environment = gym.make(task)
# Make sure the environment obeys the dm_env.Environment interface.
environment = wrappers.GymWrapper(environment)
# Clip the action returned by the agent to the environment spec.
environment = wrappers.CanonicalSpecWrapper(environment, clip=True)
environment = wrappers.SinglePrecisionWrapper(environment)
return environment
def make_environment(task, evaluation = False):
"""Creates an OpenAI Gym environment."""
# Load the gym environment.
environment = gym.make(task)
environment = env_wrappers.AdroitSparseRewardWrapper(environment)
# Make sure the environment obeys the dm_env.Environment interface.
environment = wrappers.GymWrapper(environment)
# Clip the action returned by the agent to the environment spec.
environment = wrappers.CanonicalSpecWrapper(environment, clip=True)
environment = wrappers.SinglePrecisionWrapper(environment)
if evaluation:
environment = env_wrappers.SuccessRewardWrapper(environment,
success_threshold=1.)
return environment
def make_environment(
evaluation: bool = False,
domain_name: str = 'cartpole',
task_name: str = 'balance',
from_pixels: bool = False,
frames_to_stack: int = 3,
flatten_stack: bool = False,
num_action_repeats: Optional[int] = None,
) -> dm_env.Environment:
"""Implements a control suite environment factory."""
# Load dm_suite lazily not require Mujoco license when not using it.
from dm_control import suite # pylint: disable=g-import-not-at-top
from acme.wrappers import mujoco as mujoco_wrappers # pylint: disable=g-import-not-at-top
# Load raw control suite environment.
environment = suite.load(domain_name, task_name)
# Maybe wrap to get pixel observations from environment state.
if from_pixels:
environment = mujoco_wrappers.MujocoPixelWrapper(environment)
environment = wrappers.FrameStackingWrapper(environment,
num_frames=frames_to_stack,
flatten=flatten_stack)
environment = wrappers.CanonicalSpecWrapper(environment, clip=True)
if num_action_repeats:
environment = wrappers.ActionRepeatWrapper(
environment, num_repeats=num_action_repeats)
environment = wrappers.SinglePrecisionWrapper(environment)
if evaluation:
# The evaluator in the distributed agent will set this to True so you can
# use this clause to, e.g., set up video recording by the evaluator.
pass
return environment