"""An async vector environment."""
import multiprocessing as mp
import sys
import time
from copy import deepcopy
from enum import Enum
from typing import Any, Callable, List, Optional, Sequence, Tuple, Union
import numpy as np
from numpy.typing import NDArray
import gymnasium as gym
from gymnasium import logger
from gymnasium.core import Env, ObsType
from gymnasium.error import (
AlreadyPendingCallError,
ClosedEnvironmentError,
CustomSpaceError,
NoAsyncCallError,
)
from gymnasium.vector.utils import (
CloudpickleWrapper,
clear_mpi_env_vars,
concatenate,
create_empty_array,
create_shared_memory,
iterate,
read_from_shared_memory,
write_to_shared_memory,
)
from gymnasium.vector.vector_env import VectorEnv
__all__ = ["AsyncVectorEnv"]
class AsyncState(Enum):
DEFAULT = "default"
WAITING_RESET = "reset"
WAITING_STEP = "step"
WAITING_CALL = "call"
[docs]class AsyncVectorEnv(VectorEnv):
"""Vectorized environment that runs multiple environments in parallel.
It uses ``multiprocessing`` processes, and pipes for communication.
Example:
>>> import gymnasium as gym
>>> env = gym.vector.AsyncVectorEnv([
... lambda: gym.make("Pendulum-v1", g=9.81),
... lambda: gym.make("Pendulum-v1", g=1.62)
... ])
>>> env.reset(seed=42)
(array([[-0.14995256, 0.9886932 , -0.12224312],
[ 0.5760367 , 0.8174238 , -0.91244936]], dtype=float32), {})
"""
def __init__(
self,
env_fns: Sequence[Callable[[], Env]],
observation_space: Optional[gym.Space] = None,
action_space: Optional[gym.Space] = None,
shared_memory: bool = True,
copy: bool = True,
context: Optional[str] = None,
daemon: bool = True,
worker: Optional[Callable] = None,
):
"""Vectorized environment that runs multiple environments in parallel.
Args:
env_fns: Functions that create the environments.
observation_space: Observation space of a single environment. If ``None``,
then the observation space of the first environment is taken.
action_space: Action space of a single environment. If ``None``,
then the action space of the first environment is taken.
shared_memory: If ``True``, then the observations from the worker processes are communicated back through
shared variables. This can improve the efficiency if the observations are large (e.g. images).
copy: If ``True``, then the :meth:`~AsyncVectorEnv.reset` and :meth:`~AsyncVectorEnv.step` methods
return a copy of the observations.
context: Context for `multiprocessing`_. If ``None``, then the default context is used.
daemon: If ``True``, then subprocesses have ``daemon`` flag turned on; that is, they will quit if
the head process quits. However, ``daemon=True`` prevents subprocesses to spawn children,
so for some environments you may want to have it set to ``False``.
worker: If set, then use that worker in a subprocess instead of a default one.
Can be useful to override some inner vector env logic, for instance, how resets on termination or truncation are handled.
Warnings:
worker is an advanced mode option. It provides a high degree of flexibility and a high chance
to shoot yourself in the foot; thus, if you are writing your own worker, it is recommended to start
from the code for ``_worker`` (or ``_worker_shared_memory``) method, and add changes.
Raises:
RuntimeError: If the observation space of some sub-environment does not match observation_space
(or, by default, the observation space of the first sub-environment).
ValueError: If observation_space is a custom space (i.e. not a default space in Gym,
such as gymnasium.spaces.Box, gymnasium.spaces.Discrete, or gymnasium.spaces.Dict) and shared_memory is True.
"""
ctx = mp.get_context(context)
self.env_fns = env_fns
self.shared_memory = shared_memory
self.copy = copy
dummy_env = env_fns[0]()
self.metadata = dummy_env.metadata
if (observation_space is None) or (action_space is None):
observation_space = observation_space or dummy_env.observation_space
action_space = action_space or dummy_env.action_space
dummy_env.close()
del dummy_env
super().__init__(
num_envs=len(env_fns),
observation_space=observation_space,
action_space=action_space,
)
if self.shared_memory:
try:
_obs_buffer = create_shared_memory(
self.single_observation_space, n=self.num_envs, ctx=ctx
)
self.observations = read_from_shared_memory(
self.single_observation_space, _obs_buffer, n=self.num_envs
)
except CustomSpaceError as e:
raise ValueError(
"Using `shared_memory=True` in `AsyncVectorEnv` "
"is incompatible with non-standard Gymnasium observation spaces "
"(i.e. custom spaces inheriting from `gymnasium.Space`), and is "
"only compatible with default Gymnasium spaces (e.g. `Box`, "
"`Tuple`, `Dict`) for batching. Set `shared_memory=False` "
"if you use custom observation spaces."
) from e
else:
_obs_buffer = None
self.observations = create_empty_array(
self.single_observation_space, n=self.num_envs, fn=np.zeros
)
self.parent_pipes, self.processes = [], []
self.error_queue = ctx.Queue()
target = _worker_shared_memory if self.shared_memory else _worker
target = worker or target
with clear_mpi_env_vars():
for idx, env_fn in enumerate(self.env_fns):
parent_pipe, child_pipe = ctx.Pipe()
process = ctx.Process(
target=target,
name=f"Worker<{type(self).__name__}>-{idx}",
args=(
idx,
CloudpickleWrapper(env_fn),
child_pipe,
parent_pipe,
_obs_buffer,
self.error_queue,
),
)
self.parent_pipes.append(parent_pipe)
self.processes.append(process)
process.daemon = daemon
process.start()
child_pipe.close()
self._state = AsyncState.DEFAULT
self._check_spaces()
def reset_async(
self,
seed: Optional[Union[int, List[int]]] = None,
options: Optional[dict] = None,
):
"""Send calls to the :obj:`reset` methods of the sub-environments.
To get the results of these calls, you may invoke :meth:`reset_wait`.
Args:
seed: List of seeds for each environment
options: The reset option
Raises:
ClosedEnvironmentError: If the environment was closed (if :meth:`close` was previously called).
AlreadyPendingCallError: If the environment is already waiting for a pending call to another
method (e.g. :meth:`step_async`). This can be caused by two consecutive
calls to :meth:`reset_async`, with no call to :meth:`reset_wait` in between.
"""
self._assert_is_running()
if seed is None:
seed = [None for _ in range(self.num_envs)]
if isinstance(seed, int):
seed = [seed + i for i in range(self.num_envs)]
assert len(seed) == self.num_envs
if self._state != AsyncState.DEFAULT:
raise AlreadyPendingCallError(
f"Calling `reset_async` while waiting for a pending call to `{self._state.value}` to complete",
self._state.value,
)
for pipe, single_seed in zip(self.parent_pipes, seed):
single_kwargs = {}
if single_seed is not None:
single_kwargs["seed"] = single_seed
if options is not None:
single_kwargs["options"] = options
pipe.send(("reset", single_kwargs))
self._state = AsyncState.WAITING_RESET
def reset_wait(
self,
timeout: Optional[Union[int, float]] = None,
seed: Optional[int] = None,
options: Optional[dict] = None,
) -> Union[ObsType, Tuple[ObsType, dict]]:
"""Waits for the calls triggered by :meth:`reset_async` to finish and returns the results.
Args:
timeout: Number of seconds before the call to `reset_wait` times out. If `None`, the call to `reset_wait` never times out.
seed: ignored
options: ignored
Returns:
A tuple of batched observations and list of dictionaries
Raises:
ClosedEnvironmentError: If the environment was closed (if :meth:`close` was previously called).
NoAsyncCallError: If :meth:`reset_wait` was called without any prior call to :meth:`reset_async`.
TimeoutError: If :meth:`reset_wait` timed out.
"""
self._assert_is_running()
if self._state != AsyncState.WAITING_RESET:
raise NoAsyncCallError(
"Calling `reset_wait` without any prior " "call to `reset_async`.",
AsyncState.WAITING_RESET.value,
)
if not self._poll(timeout):
self._state = AsyncState.DEFAULT
raise mp.TimeoutError(
f"The call to `reset_wait` has timed out after {timeout} second(s)."
)
results, successes = zip(*[pipe.recv() for pipe in self.parent_pipes])
self._raise_if_errors(successes)
self._state = AsyncState.DEFAULT
infos = {}
results, info_data = zip(*results)
for i, info in enumerate(info_data):
infos = self._add_info(infos, info, i)
if not self.shared_memory:
self.observations = concatenate(
self.single_observation_space, results, self.observations
)
return (deepcopy(self.observations) if self.copy else self.observations), infos
def step_async(self, actions: np.ndarray):
"""Send the calls to :obj:`step` to each sub-environment.
Args:
actions: Batch of actions. element of :attr:`~VectorEnv.action_space`
Raises:
ClosedEnvironmentError: If the environment was closed (if :meth:`close` was previously called).
AlreadyPendingCallError: If the environment is already waiting for a pending call to another
method (e.g. :meth:`reset_async`). This can be caused by two consecutive
calls to :meth:`step_async`, with no call to :meth:`step_wait` in
between.
"""
self._assert_is_running()
if self._state != AsyncState.DEFAULT:
raise AlreadyPendingCallError(
f"Calling `step_async` while waiting for a pending call to `{self._state.value}` to complete.",
self._state.value,
)
actions = iterate(self.action_space, actions)
for pipe, action in zip(self.parent_pipes, actions):
pipe.send(("step", action))
self._state = AsyncState.WAITING_STEP
def step_wait(
self, timeout: Optional[Union[int, float]] = None
) -> Tuple[Any, NDArray[Any], NDArray[Any], NDArray[Any], dict]:
"""Wait for the calls to :obj:`step` in each sub-environment to finish.
Args:
timeout: Number of seconds before the call to :meth:`step_wait` times out. If ``None``, the call to :meth:`step_wait` never times out.
Returns:
The batched environment step information, (obs, reward, terminated, truncated, info)
Raises:
ClosedEnvironmentError: If the environment was closed (if :meth:`close` was previously called).
NoAsyncCallError: If :meth:`step_wait` was called without any prior call to :meth:`step_async`.
TimeoutError: If :meth:`step_wait` timed out.
"""
self._assert_is_running()
if self._state != AsyncState.WAITING_STEP:
raise NoAsyncCallError(
"Calling `step_wait` without any prior call " "to `step_async`.",
AsyncState.WAITING_STEP.value,
)
if not self._poll(timeout):
self._state = AsyncState.DEFAULT
raise mp.TimeoutError(
f"The call to `step_wait` has timed out after {timeout} second(s)."
)
observations_list, rewards, terminateds, truncateds, infos = [], [], [], [], {}
successes = []
for i, pipe in enumerate(self.parent_pipes):
result, success = pipe.recv()
successes.append(success)
if success:
obs, rew, terminated, truncated, info = result
observations_list.append(obs)
rewards.append(rew)
terminateds.append(terminated)
truncateds.append(truncated)
infos = self._add_info(infos, info, i)
self._raise_if_errors(successes)
self._state = AsyncState.DEFAULT
if not self.shared_memory:
self.observations = concatenate(
self.single_observation_space,
observations_list,
self.observations,
)
return (
deepcopy(self.observations) if self.copy else self.observations,
np.array(rewards),
np.array(terminateds, dtype=np.bool_),
np.array(truncateds, dtype=np.bool_),
infos,
)
def call_async(self, name: str, *args, **kwargs):
"""Calls the method with name asynchronously and apply args and kwargs to the method.
Args:
name: Name of the method or property to call.
*args: Arguments to apply to the method call.
**kwargs: Keyword arguments to apply to the method call.
Raises:
ClosedEnvironmentError: If the environment was closed (if :meth:`close` was previously called).
AlreadyPendingCallError: Calling `call_async` while waiting for a pending call to complete
"""
self._assert_is_running()
if self._state != AsyncState.DEFAULT:
raise AlreadyPendingCallError(
"Calling `call_async` while waiting "
f"for a pending call to `{self._state.value}` to complete.",
self._state.value,
)
for pipe in self.parent_pipes:
pipe.send(("_call", (name, args, kwargs)))
self._state = AsyncState.WAITING_CALL
def call_wait(self, timeout: Optional[Union[int, float]] = None) -> list:
"""Calls all parent pipes and waits for the results.
Args:
timeout: Number of seconds before the call to `step_wait` times out.
If `None` (default), the call to `step_wait` never times out.
Returns:
List of the results of the individual calls to the method or property for each environment.
Raises:
NoAsyncCallError: Calling `call_wait` without any prior call to `call_async`.
TimeoutError: The call to `call_wait` has timed out after timeout second(s).
"""
self._assert_is_running()
if self._state != AsyncState.WAITING_CALL:
raise NoAsyncCallError(
"Calling `call_wait` without any prior call to `call_async`.",
AsyncState.WAITING_CALL.value,
)
if not self._poll(timeout):
self._state = AsyncState.DEFAULT
raise mp.TimeoutError(
f"The call to `call_wait` has timed out after {timeout} second(s)."
)
results, successes = zip(*[pipe.recv() for pipe in self.parent_pipes])
self._raise_if_errors(successes)
self._state = AsyncState.DEFAULT
return results
def set_attr(self, name: str, values: Union[list, tuple, object]):
"""Sets an attribute of the sub-environments.
Args:
name: Name of the property to be set in each individual environment.
values: Values of the property to be set to. If ``values`` is a list or
tuple, then it corresponds to the values for each individual
environment, otherwise a single value is set for all environments.
Raises:
ValueError: Values must be a list or tuple with length equal to the number of environments.
AlreadyPendingCallError: Calling `set_attr` while waiting for a pending call to complete.
"""
self._assert_is_running()
if not isinstance(values, (list, tuple)):
values = [values for _ in range(self.num_envs)]
if len(values) != self.num_envs:
raise ValueError(
"Values must be a list or tuple with length equal to the "
f"number of environments. Got `{len(values)}` values for "
f"{self.num_envs} environments."
)
if self._state != AsyncState.DEFAULT:
raise AlreadyPendingCallError(
"Calling `set_attr` while waiting "
f"for a pending call to `{self._state.value}` to complete.",
self._state.value,
)
for pipe, value in zip(self.parent_pipes, values):
pipe.send(("_setattr", (name, value)))
_, successes = zip(*[pipe.recv() for pipe in self.parent_pipes])
self._raise_if_errors(successes)
def close_extras(
self, timeout: Optional[Union[int, float]] = None, terminate: bool = False
):
"""Close the environments & clean up the extra resources (processes and pipes).
Args:
timeout: Number of seconds before the call to :meth:`close` times out. If ``None``,
the call to :meth:`close` never times out. If the call to :meth:`close`
times out, then all processes are terminated.
terminate: If ``True``, then the :meth:`close` operation is forced and all processes are terminated.
Raises:
TimeoutError: If :meth:`close` timed out.
"""
timeout = 0 if terminate else timeout
try:
if self._state != AsyncState.DEFAULT:
logger.warn(
f"Calling `close` while waiting for a pending call to `{self._state.value}` to complete."
)
function = getattr(self, f"{self._state.value}_wait")
function(timeout)
except mp.TimeoutError:
terminate = True
if terminate:
for process in self.processes:
if process.is_alive():
process.terminate()
else:
for pipe in self.parent_pipes:
if (pipe is not None) and (not pipe.closed):
pipe.send(("close", None))
for pipe in self.parent_pipes:
if (pipe is not None) and (not pipe.closed):
pipe.recv()
for pipe in self.parent_pipes:
if pipe is not None:
pipe.close()
for process in self.processes:
process.join()
def _poll(self, timeout=None):
self._assert_is_running()
if timeout is None:
return True
end_time = time.perf_counter() + timeout
delta = None
for pipe in self.parent_pipes:
delta = max(end_time - time.perf_counter(), 0)
if pipe is None:
return False
if pipe.closed or (not pipe.poll(delta)):
return False
return True
def _check_spaces(self):
self._assert_is_running()
spaces = (self.single_observation_space, self.single_action_space)
for pipe in self.parent_pipes:
pipe.send(("_check_spaces", spaces))
results, successes = zip(*[pipe.recv() for pipe in self.parent_pipes])
self._raise_if_errors(successes)
same_observation_spaces, same_action_spaces = zip(*results)
if not all(same_observation_spaces):
raise RuntimeError(
"Some environments have an observation space different from "
f"`{self.single_observation_space}`. In order to batch observations, "
"the observation spaces from all environments must be equal."
)
if not all(same_action_spaces):
raise RuntimeError(
"Some environments have an action space different from "
f"`{self.single_action_space}`. In order to batch actions, the "
"action spaces from all environments must be equal."
)
def _assert_is_running(self):
if self.closed:
raise ClosedEnvironmentError(
f"Trying to operate on `{type(self).__name__}`, after a call to `close()`."
)
def _raise_if_errors(self, successes):
if all(successes):
return
num_errors = self.num_envs - sum(successes)
assert num_errors > 0
for i in range(num_errors):
index, exctype, value = self.error_queue.get()
logger.error(
f"Received the following error from Worker-{index}: {exctype.__name__}: {value}"
)
logger.error(f"Shutting down Worker-{index}.")
self.parent_pipes[index].close()
self.parent_pipes[index] = None
if i == num_errors - 1:
logger.error("Raising the last exception back to the main process.")
raise exctype(value)
def __del__(self):
"""On deleting the object, checks that the vector environment is closed."""
if not getattr(self, "closed", True) and hasattr(self, "_state"):
self.close(terminate=True)
def _worker(index, env_fn, pipe, parent_pipe, shared_memory, error_queue):
assert shared_memory is None
env = env_fn()
parent_pipe.close()
try:
while True:
command, data = pipe.recv()
if command == "reset":
observation, info = env.reset(**data)
pipe.send(((observation, info), True))
elif command == "step":
(
observation,
reward,
terminated,
truncated,
info,
) = env.step(data)
if terminated or truncated:
old_observation, old_info = observation, info
observation, info = env.reset()
info["final_observation"] = old_observation
info["final_info"] = old_info
pipe.send(((observation, reward, terminated, truncated, info), True))
elif command == "seed":
env.seed(data)
pipe.send((None, True))
elif command == "close":
pipe.send((None, True))
break
elif command == "_call":
name, args, kwargs = data
if name in ["reset", "step", "seed", "close"]:
raise ValueError(
f"Trying to call function `{name}` with "
f"`_call`. Use `{name}` directly instead."
)
function = getattr(env, name)
if callable(function):
pipe.send((function(*args, **kwargs), True))
else:
pipe.send((function, True))
elif command == "_setattr":
name, value = data
setattr(env, name, value)
pipe.send((None, True))
elif command == "_check_spaces":
pipe.send(
(
(data[0] == env.observation_space, data[1] == env.action_space),
True,
)
)
else:
raise RuntimeError(
f"Received unknown command `{command}`. Must "
"be one of {`reset`, `step`, `seed`, `close`, `_call`, "
"`_setattr`, `_check_spaces`}."
)
except (KeyboardInterrupt, Exception):
error_queue.put((index,) + sys.exc_info()[:2])
pipe.send((None, False))
finally:
env.close()
def _worker_shared_memory(index, env_fn, pipe, parent_pipe, shared_memory, error_queue):
assert shared_memory is not None
env = env_fn()
observation_space = env.observation_space
parent_pipe.close()
try:
while True:
command, data = pipe.recv()
if command == "reset":
observation, info = env.reset(**data)
write_to_shared_memory(
observation_space, index, observation, shared_memory
)
pipe.send(((None, info), True))
elif command == "step":
(
observation,
reward,
terminated,
truncated,
info,
) = env.step(data)
if terminated or truncated:
old_observation, old_info = observation, info
observation, info = env.reset()
info["final_observation"] = old_observation
info["final_info"] = old_info
write_to_shared_memory(
observation_space, index, observation, shared_memory
)
pipe.send(((None, reward, terminated, truncated, info), True))
elif command == "seed":
env.seed(data)
pipe.send((None, True))
elif command == "close":
pipe.send((None, True))
break
elif command == "_call":
name, args, kwargs = data
if name in ["reset", "step", "seed", "close"]:
raise ValueError(
f"Trying to call function `{name}` with "
f"`_call`. Use `{name}` directly instead."
)
function = getattr(env, name)
if callable(function):
pipe.send((function(*args, **kwargs), True))
else:
pipe.send((function, True))
elif command == "_setattr":
name, value = data
setattr(env, name, value)
pipe.send((None, True))
elif command == "_check_spaces":
pipe.send(
((data[0] == observation_space, data[1] == env.action_space), True)
)
else:
raise RuntimeError(
f"Received unknown command `{command}`. Must "
"be one of {`reset`, `step`, `seed`, `close`, `_call`, "
"`_setattr`, `_check_spaces`}."
)
except (KeyboardInterrupt, Exception):
error_queue.put((index,) + sys.exc_info()[:2])
pipe.send((None, False))
finally:
env.close()