test_libs.py

# Copyright (c) Meta Platforms, Inc. and affiliates.
#
# This source code is licensed under the MIT license found in the
# LICENSE file in the root directory of this source tree.
from __future__ import annotations

import collections
import functools
import gc
import importlib.util
import urllib.error

_has_isaac = importlib.util.find_spec("isaacgym") is not None

if _has_isaac:
    # isaac gym asks to be imported before torch...
    import isaacgym  # noqa
    import isaacgymenvs  # noqa
    from torchrl.envs.libs.isaacgym import IsaacGymEnv
import argparse
import importlib
import os

import time
import urllib
from contextlib import nullcontext
from pathlib import Path
from sys import platform
from unittest import mock

import numpy as np
import pytest
import torch

if os.getenv("PYTORCH_TEST_FBCODE"):
    from pytorch.rl.test._utils_internal import (
        _make_multithreaded_env,
        CARTPOLE_VERSIONED,
        get_available_devices,
        get_default_devices,
        HALFCHEETAH_VERSIONED,
        PENDULUM_VERSIONED,
        PONG_VERSIONED,
        rand_reset,
        retry,
        rollout_consistency_assertion,
    )
else:
    from _utils_internal import (
        _make_multithreaded_env,
        CARTPOLE_VERSIONED,
        get_available_devices,
        get_default_devices,
        HALFCHEETAH_VERSIONED,
        PENDULUM_VERSIONED,
        PONG_VERSIONED,
        rand_reset,
        retry,
        rollout_consistency_assertion,
    )
from packaging import version
from tensordict import (
    assert_allclose_td,
    is_tensor_collection,
    LazyStackedTensorDict,
    TensorDict,
)
from tensordict.nn import (
    ProbabilisticTensorDictModule,
    TensorDictModule,
    TensorDictSequential,
)
from torch import nn

from torchrl._utils import implement_for, logger as torchrl_logger
from torchrl.collectors.collectors import SyncDataCollector
from torchrl.data import (
    Binary,
    Bounded,
    Categorical,
    Composite,
    MultiCategorical,
    MultiOneHot,
    NonTensor,
    OneHot,
    ReplayBuffer,
    ReplayBufferEnsemble,
    Unbounded,
    UnboundedDiscreteTensorSpec,
)
from torchrl.data.datasets.atari_dqn import AtariDQNExperienceReplay
from torchrl.data.datasets.d4rl import D4RLExperienceReplay

from torchrl.data.datasets.gen_dgrl import GenDGRLExperienceReplay
from torchrl.data.datasets.minari_data import MinariExperienceReplay
from torchrl.data.datasets.openml import OpenMLExperienceReplay
from torchrl.data.datasets.openx import OpenXExperienceReplay
from torchrl.data.datasets.roboset import RobosetExperienceReplay
from torchrl.data.datasets.vd4rl import VD4RLExperienceReplay
from torchrl.data.replay_buffers import SamplerWithoutReplacement
from torchrl.data.utils import CloudpickleWrapper
from torchrl.envs import (
    CatTensors,
    Compose,
    DoubleToFloat,
    EnvBase,
    EnvCreator,
    RemoveEmptySpecs,
    RenameTransform,
)
from torchrl.envs.batched_envs import SerialEnv
from torchrl.envs.libs.brax import _has_brax, BraxEnv, BraxWrapper
from torchrl.envs.libs.dm_control import _has_dmc, DMControlEnv, DMControlWrapper
from torchrl.envs.libs.envpool import _has_envpool, MultiThreadedEnvWrapper
from torchrl.envs.libs.gym import (
    _gym_to_torchrl_spec_transform,
    _has_gym,
    _is_from_pixels,
    _torchrl_to_gym_spec_transform,
    gym_backend,
    GymEnv,
    GymWrapper,
    MOGymEnv,
    MOGymWrapper,
    register_gym_spec_conversion,
    set_gym_backend,
)
from torchrl.envs.libs.habitat import _has_habitat, HabitatEnv
from torchrl.envs.libs.jumanji import _has_jumanji, JumanjiEnv
from torchrl.envs.libs.meltingpot import MeltingpotEnv, MeltingpotWrapper
from torchrl.envs.libs.openml import OpenMLEnv
from torchrl.envs.libs.openspiel import _has_pyspiel, OpenSpielEnv, OpenSpielWrapper
from torchrl.envs.libs.pettingzoo import _has_pettingzoo, PettingZooEnv
from torchrl.envs.libs.robohive import _has_robohive, RoboHiveEnv
from torchrl.envs.libs.smacv2 import _has_smacv2, SMACv2Env
from torchrl.envs.libs.unity_mlagents import (
    _has_unity_mlagents,
    UnityMLAgentsEnv,
    UnityMLAgentsWrapper,
)
from torchrl.envs.libs.vmas import _has_vmas, VmasEnv, VmasWrapper

from torchrl.envs.transforms import ActionMask, TransformedEnv
from torchrl.envs.utils import (
    check_env_specs,
    ExplorationType,
    MarlGroupMapType,
    RandomPolicy,
)
from torchrl.modules import (
    ActorCriticOperator,
    MaskedCategorical,
    MLP,
    SafeModule,
    ValueOperator,
)

_has_d4rl = importlib.util.find_spec("d4rl") is not None

_has_mo = importlib.util.find_spec("mo_gymnasium") is not None

_has_sklearn = importlib.util.find_spec("sklearn") is not None

_has_gym_robotics = importlib.util.find_spec("gymnasium_robotics") is not None

_has_minari = importlib.util.find_spec("minari") is not None

_has_gymnasium = importlib.util.find_spec("gymnasium") is not None
_has_gym_regular = importlib.util.find_spec("gym") is not None
if _has_gymnasium:
    set_gym_backend("gymnasium").set()
    import gymnasium

    assert gym_backend() is gymnasium
elif _has_gym:
    set_gym_backend("gym").set()
    import gym

    assert gym_backend() is gym

_has_meltingpot = importlib.util.find_spec("meltingpot") is not None

_has_minigrid = importlib.util.find_spec("minigrid") is not None


@pytest.fixture(scope="session", autouse=True)
def maybe_init_minigrid():
    if _has_minigrid and _has_gymnasium:
        import minigrid

        minigrid.register_minigrid_envs()


@implement_for("gym")
def get_gym_pixel_wrapper():
    try:
        # works whenever gym_version > version.parse("0.19")
        PixelObservationWrapper = gym_backend(
            "wrappers.pixel_observation"
        ).PixelObservationWrapper
    except Exception:
        from torchrl.envs.libs.utils import (
            GymPixelObservationWrapper as PixelObservationWrapper,
        )
    return PixelObservationWrapper


@implement_for("gymnasium", None, "1.1.0")
def get_gym_pixel_wrapper():  # noqa: F811
    try:
        # works whenever gym_version > version.parse("0.19")
        PixelObservationWrapper = gym_backend(
            "wrappers.pixel_observation"
        ).PixelObservationWrapper
    except Exception:
        from torchrl.envs.libs.utils import (
            GymPixelObservationWrapper as PixelObservationWrapper,
        )
    return PixelObservationWrapper


@implement_for("gymnasium", "1.1.0")
def get_gym_pixel_wrapper():  # noqa: F811
    # works whenever gym_version > version.parse("0.19")
    PixelObservationWrapper = lambda *args, pixels_only=False, **kwargs: gym_backend(
        "wrappers"
    ).AddRenderObservation(*args, render_only=pixels_only, **kwargs)
    return PixelObservationWrapper


if _has_gym:
    try:
        from gymnasium import __version__ as gym_version

        gym_version = version.parse(gym_version)
    except ModuleNotFoundError:
        from gym import __version__ as gym_version

        gym_version = version.parse(gym_version)


if _has_envpool:
    import envpool

_has_pytree = True
try:
    from torch.utils._pytree import tree_flatten, tree_map
except ImportError:
    _has_pytree = False
IS_OSX = platform == "darwin"
RTOL = 1e-1
ATOL = 1e-1


@pytest.mark.skipif(not _has_gym, reason="no gym library found")
class TestGym:
    class DummyEnv(EnvBase):
        def __init__(self, arg1, *, arg2, **kwargs):
            super().__init__(**kwargs)

            assert arg1 == 1
            assert arg2 == 2

            self.observation_spec = Composite(
                observation=Unbounded((*self.batch_size, 3)),
                other=Composite(
                    another_other=Unbounded((*self.batch_size, 3)),
                    shape=self.batch_size,
                ),
                shape=self.batch_size,
            )
            self.action_spec = Unbounded((*self.batch_size, 3))
            self.done_spec = Categorical(2, (*self.batch_size, 1), dtype=torch.bool)
            self.full_done_spec["truncated"] = self.full_done_spec["terminated"].clone()

        def _reset(self, tensordict):
            return self.observation_spec.rand()

        def _step(self, tensordict):
            action = tensordict.get("action")
            return TensorDict(
                {
                    "observation": action.clone(),
                    "other": {"another_other": torch.zeros_like(action)},
                    "reward": action.sum(-1, True),
                    "done": ~action.any(-1, True),
                    "terminated": ~action.any(-1, True),
                    "truncated": torch.zeros((*self.batch_size, 1), dtype=torch.bool),
                },
                batch_size=[],
            )

        def _set_seed(self, seed: int | None) -> None:
            ...

    @implement_for("gym", None, "0.18")
    def _make_spec(self, batch_size, cat, cat_shape, multicat, multicat_shape):
        return Composite(
            a=Unbounded(shape=(*batch_size, 1)),
            b=Composite(c=cat(5, shape=cat_shape, dtype=torch.int64), shape=batch_size),
            d=cat(5, shape=cat_shape, dtype=torch.int64),
            e=multicat([2, 3], shape=(*batch_size, multicat_shape), dtype=torch.int64),
            f=Bounded(-3, 4, shape=(*batch_size, 1)),
            # g=UnboundedDiscreteTensorSpec(shape=(*batch_size, 1), dtype=torch.long),
            h=Binary(n=5, shape=(*batch_size, 5)),
            shape=batch_size,
        )

    @implement_for("gym", "0.18", None)
    def _make_spec(  # noqa: F811
        self, batch_size, cat, cat_shape, multicat, multicat_shape
    ):
        return Composite(
            a=Unbounded(shape=(*batch_size, 1)),
            b=Composite(c=cat(5, shape=cat_shape, dtype=torch.int64), shape=batch_size),
            d=cat(5, shape=cat_shape, dtype=torch.int64),
            e=multicat([2, 3], shape=(*batch_size, multicat_shape), dtype=torch.int64),
            f=Bounded(-3, 4, shape=(*batch_size, 1)),
            g=UnboundedDiscreteTensorSpec(shape=(*batch_size, 1), dtype=torch.long),
            h=Binary(n=5, shape=(*batch_size, 5)),
            shape=batch_size,
        )

    @implement_for("gymnasium", None, "1.0.0")
    def _make_spec(  # noqa: F811
        self, batch_size, cat, cat_shape, multicat, multicat_shape
    ):
        return Composite(
            a=Unbounded(shape=(*batch_size, 1)),
            b=Composite(c=cat(5, shape=cat_shape, dtype=torch.int64), shape=batch_size),
            d=cat(5, shape=cat_shape, dtype=torch.int64),
            e=multicat([2, 3], shape=(*batch_size, multicat_shape), dtype=torch.int64),
            f=Bounded(-3, 4, shape=(*batch_size, 1)),
            g=UnboundedDiscreteTensorSpec(shape=(*batch_size, 1), dtype=torch.long),
            h=Binary(n=5, shape=(*batch_size, 5)),
            shape=batch_size,
        )

    @implement_for("gymnasium", "1.1.0")
    def _make_spec(  # noqa: F811
        self, batch_size, cat, cat_shape, multicat, multicat_shape
    ):
        return Composite(
            a=Unbounded(shape=(*batch_size, 1)),
            b=Composite(c=cat(5, shape=cat_shape, dtype=torch.int64), shape=batch_size),
            d=cat(5, shape=cat_shape, dtype=torch.int64),
            e=multicat([2, 3], shape=(*batch_size, multicat_shape), dtype=torch.int64),
            f=Bounded(-3, 4, shape=(*batch_size, 1)),
            g=UnboundedDiscreteTensorSpec(shape=(*batch_size, 1), dtype=torch.long),
            h=Binary(n=5, shape=(*batch_size, 5)),
            shape=batch_size,
        )

    @pytest.mark.parametrize("categorical", [True, False])
    def test_gym_spec_cast(self, categorical):
        batch_size = [3, 4]
        cat = Categorical if categorical else OneHot
        cat_shape = batch_size if categorical else (*batch_size, 5)
        multicat = MultiCategorical if categorical else MultiOneHot
        multicat_shape = 2 if categorical else 5
        spec = self._make_spec(batch_size, cat, cat_shape, multicat, multicat_shape)
        recon = _gym_to_torchrl_spec_transform(
            _torchrl_to_gym_spec_transform(
                spec, categorical_action_encoding=categorical
            ),
            categorical_action_encoding=categorical,
            batch_size=batch_size,
        )
        for (key0, spec0), (key1, spec1) in zip(
            spec.items(True, True), recon.items(True, True)
        ):
            assert spec0 == spec1, (key0, key1, spec0, spec1)
        assert spec == recon
        assert recon.shape == spec.shape

    def test_gym_new_spec_reg(self):
        Space = gym_backend("spaces").Space

        class MySpaceParent(Space):
            ...

        s_parent = MySpaceParent()

        class MySpaceChild(MySpaceParent):
            ...

        # We intentionally register first the child then the parent
        @register_gym_spec_conversion(MySpaceChild)
        def convert_myspace_child(spec, **kwargs):
            return NonTensor((), example_data="child")

        @register_gym_spec_conversion(MySpaceParent)
        def convert_myspace_parent(spec, **kwargs):
            return NonTensor((), example_data="parent")

        s_child = MySpaceChild()
        assert _gym_to_torchrl_spec_transform(s_parent).example_data == "parent"
        assert _gym_to_torchrl_spec_transform(s_child).example_data == "child"

        class NoConversionSpace(Space):
            ...

        s_no_conv = NoConversionSpace()
        with pytest.raises(
            KeyError, match="No conversion tool could be found with the gym space"
        ):
            _gym_to_torchrl_spec_transform(s_no_conv)

    @pytest.mark.parametrize("order", ["tuple_seq"])
    @implement_for("gym")
    def test_gym_spec_cast_tuple_sequential(self, order):
        torchrl_logger.info("Sequence not available in gym")
        return

    @pytest.mark.parametrize("order", ["tuple_seq"])
    @implement_for("gymnasium", "1.1.0")
    def test_gym_spec_cast_tuple_sequential(self, order):  # noqa: F811
        self._test_gym_spec_cast_tuple_sequential(order)

    @pytest.mark.parametrize("order", ["tuple_seq"])
    @implement_for("gymnasium", None, "1.0.0")
    def test_gym_spec_cast_tuple_sequential(self, order):  # noqa: F811
        self._test_gym_spec_cast_tuple_sequential(order)

    def _test_gym_spec_cast_tuple_sequential(self, order):  # noqa: F811
        with set_gym_backend("gymnasium"):
            if order == "seq_tuple":
                # Requires nested tensors to be created along dim=1, disabling
                space = gym_backend("spaces").Dict(
                    feature=gym_backend("spaces").Sequence(
                        gym_backend("spaces").Tuple(
                            (
                                gym_backend("spaces").Box(-1, 1, shape=(2, 2)),
                                gym_backend("spaces").Box(-1, 1, shape=(1, 2)),
                            )
                        ),
                        stack=True,
                    )
                )
            elif order == "tuple_seq":
                space = gym_backend("spaces").Dict(
                    feature=gym_backend("spaces").Tuple(
                        (
                            gym_backend("spaces").Sequence(
                                gym_backend("spaces").Box(-1, 1, shape=(2, 2)),
                                stack=True,
                            ),
                            gym_backend("spaces").Sequence(
                                gym_backend("spaces").Box(-1, 1, shape=(1, 2)),
                                stack=True,
                            ),
                        ),
                    )
                )
            else:
                raise NotImplementedError
            sample = space.sample()
            partial_tree_map = functools.partial(
                tree_map, is_leaf=lambda x: isinstance(x, (tuple, torch.Tensor))
            )

            def stack_tuples(item):
                if isinstance(item, tuple):
                    try:
                        return torch.stack(
                            [partial_tree_map(stack_tuples, x) for x in item]
                        )
                    except RuntimeError:
                        item = [partial_tree_map(stack_tuples, x) for x in item]
                        try:
                            return torch.nested.nested_tensor(item)
                        except RuntimeError:
                            return tuple(item)
                return torch.as_tensor(item)

            sample_pt = partial_tree_map(stack_tuples, sample)
            # sample_pt = torch.utils._pytree.tree_map(lambda x: torch.stack(list(x)), sample_pt, is_leaf=lambda x: isinstance(x, tuple))
            spec = _gym_to_torchrl_spec_transform(space)
            rand = spec.rand()

            assert spec.contains(rand), (rand, spec)
            assert spec.contains(sample_pt), (rand, sample_pt)

            space_recon = _torchrl_to_gym_spec_transform(spec)
            assert space_recon == space, (space_recon, space)
            rand_numpy = rand.numpy()
            assert space.contains(rand_numpy)

    _BACKENDS = [None]
    if _has_gymnasium:
        _BACKENDS += ["gymnasium"]
    if _has_gym_regular:
        _BACKENDS += ["gym"]

    @pytest.mark.skipif(not _has_pytree, reason="pytree needed for torchrl_to_gym test")
    @pytest.mark.parametrize("backend", _BACKENDS)
    @pytest.mark.parametrize("numpy", [True, False])
    def test_torchrl_to_gym(self, backend, numpy):
        from torchrl.envs.libs.gym import gym_backend, set_gym_backend

        gb = gym_backend()
        try:
            EnvBase.register_gym(
                f"Dummy-{numpy}-{backend}-v0",
                entry_point=self.DummyEnv,
                to_numpy=numpy,
                backend=backend,
                arg1=1,
                arg2=2,
            )

            with set_gym_backend(backend) if backend is not None else nullcontext():
                envgym = gym_backend().make(f"Dummy-{numpy}-{backend}-v0")
                envgym.reset()
                obs, *_ = envgym.step(envgym.action_space.sample())
                assert "observation" in obs
                assert "other" in obs
                if numpy:
                    assert all(
                        isinstance(val, np.ndarray) for val in tree_flatten(obs)[0]
                    )
                else:
                    assert all(
                        isinstance(val, torch.Tensor) for val in tree_flatten(obs)[0]
                    )

                # with a transform
                transform = Compose(
                    CatTensors(["observation", ("other", "another_other")]),
                    RemoveEmptySpecs(),
                )
                envgym = gym_backend().make(
                    f"Dummy-{numpy}-{backend}-v0",
                    transform=transform,
                )
                envgym.reset()
                obs, *_ = envgym.step(envgym.action_space.sample())
                assert "observation_other" not in obs
                assert "observation" not in obs
                assert "other" not in obs
                if numpy:
                    assert all(
                        isinstance(val, np.ndarray) for val in tree_flatten(obs)[0]
                    )
                else:
                    assert all(
                        isinstance(val, torch.Tensor) for val in tree_flatten(obs)[0]
                    )

            # register with transform
            transform = Compose(
                CatTensors(["observation", ("other", "another_other")]),
                RemoveEmptySpecs(),
            )
            EnvBase.register_gym(
                f"Dummy-{numpy}-{backend}-transform-v0",
                entry_point=self.DummyEnv,
                backend=backend,
                to_numpy=numpy,
                arg1=1,
                arg2=2,
                transform=transform,
            )

            with set_gym_backend(backend) if backend is not None else nullcontext():
                envgym = gym_backend().make(f"Dummy-{numpy}-{backend}-transform-v0")
                envgym.reset()
                obs, *_ = envgym.step(envgym.action_space.sample())
                assert "observation_other" not in obs
                assert "observation" not in obs
                assert "other" not in obs
                if numpy:
                    assert all(
                        isinstance(val, np.ndarray) for val in tree_flatten(obs)[0]
                    )
                else:
                    assert all(
                        isinstance(val, torch.Tensor) for val in tree_flatten(obs)[0]
                    )

            # register with transform
            EnvBase.register_gym(
                f"Dummy-{numpy}-{backend}-noarg-v0",
                entry_point=self.DummyEnv,
                backend=backend,
                to_numpy=numpy,
            )
            with set_gym_backend(backend) if backend is not None else nullcontext():
                with pytest.raises(AssertionError):
                    envgym = gym_backend().make(
                        f"Dummy-{numpy}-{backend}-noarg-v0", arg1=None, arg2=None
                    )
                envgym = gym_backend().make(
                    f"Dummy-{numpy}-{backend}-noarg-v0", arg1=1, arg2=2
                )

            # Get info dict
            gym_info_at_reset = version.parse(
                gym_backend().__version__
            ) >= version.parse("0.26.0")
            with set_gym_backend(backend) if backend is not None else nullcontext():
                envgym = gym_backend().make(
                    f"Dummy-{numpy}-{backend}-noarg-v0",
                    arg1=1,
                    arg2=2,
                    info_keys=("other",),
                )
                if gym_info_at_reset:
                    out, info = envgym.reset()
                    if numpy:
                        assert all(
                            isinstance(val, np.ndarray)
                            for val in tree_flatten((obs, info))[0]
                        )
                    else:
                        assert all(
                            isinstance(val, torch.Tensor)
                            for val in tree_flatten((obs, info))[0]
                        )
                else:
                    out = envgym.reset()
                    info = {}
                    if numpy:
                        assert all(
                            isinstance(val, np.ndarray)
                            for val in tree_flatten((obs, info))[0]
                        )
                    else:
                        assert all(
                            isinstance(val, torch.Tensor)
                            for val in tree_flatten((obs, info))[0]
                        )
                assert "observation" in out
                assert "other" not in out

                if gym_info_at_reset:
                    assert "other" in info

                out, *_, info = envgym.step(envgym.action_space.sample())
                assert "observation" in out
                assert "other" not in out
                assert "other" in info
                if numpy:
                    assert all(
                        isinstance(val, np.ndarray)
                        for val in tree_flatten((obs, info))[0]
                    )
                else:
                    assert all(
                        isinstance(val, torch.Tensor)
                        for val in tree_flatten((obs, info))[0]
                    )

            EnvBase.register_gym(
                f"Dummy-{numpy}-{backend}-info-v0",
                entry_point=self.DummyEnv,
                backend=backend,
                to_numpy=numpy,
                info_keys=("other",),
            )
            with set_gym_backend(backend) if backend is not None else nullcontext():
                envgym = gym_backend().make(
                    f"Dummy-{numpy}-{backend}-info-v0", arg1=1, arg2=2
                )
                if gym_info_at_reset:
                    out, info = envgym.reset()
                    if numpy:
                        assert all(
                            isinstance(val, np.ndarray)
                            for val in tree_flatten((obs, info))[0]
                        )
                    else:
                        assert all(
                            isinstance(val, torch.Tensor)
                            for val in tree_flatten((obs, info))[0]
                        )
                else:
                    out = envgym.reset()
                    info = {}
                    if numpy:
                        assert all(
                            isinstance(val, np.ndarray)
                            for val in tree_flatten((obs, info))[0]
                        )
                    else:
                        assert all(
                            isinstance(val, torch.Tensor)
                            for val in tree_flatten((obs, info))[0]
                        )
                assert "observation" in out
                assert "other" not in out

                if gym_info_at_reset:
                    assert "other" in info

                out, *_, info = envgym.step(envgym.action_space.sample())
                assert "observation" in out
                assert "other" not in out
                assert "other" in info
                if numpy:
                    assert all(
                        isinstance(val, np.ndarray)
                        for val in tree_flatten((obs, info))[0]
                    )
                else:
                    assert all(
                        isinstance(val, torch.Tensor)
                        for val in tree_flatten((obs, info))[0]
                    )
        finally:
            set_gym_backend(gb).set()

    @implement_for("gym", None, "0.26")
    def test_gym_dict_action_space(self):
        torchrl_logger.info("tested for gym > 0.26 - no backward issue")
        return

    @implement_for("gym", "0.26", None)
    def test_gym_dict_action_space(self):  # noqa: F811
        import gym
        from gym import Env

        class CompositeActionEnv(Env):
            def __init__(self):
                self.action_space = gym.spaces.Dict(
                    a0=gym.spaces.Discrete(2), a1=gym.spaces.Box(-1, 1)
                )
                self.observation_space = gym.spaces.Box(-1, 1)

            def step(self, action):
                assert isinstance(action, dict)
                assert isinstance(action["a0"], np.ndarray)
                assert isinstance(action["a1"], np.ndarray)
                return (0.5, 0.0, False, False, {})

            def reset(
                self,
                *,
                seed: int | None = None,
                options: dict | None = None,
            ):
                return (0.0, {})

        env = CompositeActionEnv()
        torchrl_env = GymWrapper(env)
        assert isinstance(torchrl_env.action_spec, Composite)
        assert len(torchrl_env.action_keys) == 2
        r = torchrl_env.rollout(10)
        assert isinstance(r[0]["a0"], torch.Tensor)
        assert isinstance(r[0]["a1"], torch.Tensor)
        assert r[0]["observation"] == 0
        assert r[1]["observation"] == 0.5

    @implement_for("gymnasium")
    def test_gym_dict_action_space(self):  # noqa: F811
        import gymnasium as gym
        from gymnasium import Env

        class CompositeActionEnv(Env):
            def __init__(self):
                self.action_space = gym.spaces.Dict(
                    a0=gym.spaces.Discrete(2), a1=gym.spaces.Box(-1, 1)
                )
                self.observation_space = gym.spaces.Box(-1, 1)

            def step(self, action):
                assert isinstance(action, dict)
                assert isinstance(action["a0"], np.ndarray)
                assert isinstance(action["a1"], np.ndarray)
                return (0.5, 0.0, False, False, {})

            def reset(
                self,
                *,
                seed: int | None = None,
                options: dict | None = None,
            ):
                return (0.0, {})

        env = CompositeActionEnv()
        torchrl_env = GymWrapper(env)
        assert isinstance(torchrl_env.action_spec, Composite)
        assert len(torchrl_env.action_keys) == 2
        r = torchrl_env.rollout(10)
        assert isinstance(r[0]["a0"], torch.Tensor)
        assert isinstance(r[0]["a1"], torch.Tensor)
        assert r[0]["observation"] == 0
        assert r[1]["observation"] == 0.5

    @pytest.mark.parametrize(
        "env_name",
        [
            HALFCHEETAH_VERSIONED(),
            PONG_VERSIONED(),
            # PENDULUM_VERSIONED,
        ],
    )
    @pytest.mark.parametrize("frame_skip", [1, 3])
    @pytest.mark.parametrize(
        "from_pixels,pixels_only",
        [
            [True, True],
            [True, False],
            [False, False],
        ],
    )
    def test_gym(self, env_name, frame_skip, from_pixels, pixels_only):
        if env_name == PONG_VERSIONED() and not from_pixels:
            # raise pytest.skip("already pixel")
            # we don't skip because that would raise an exception
            return
        elif (
            env_name != PONG_VERSIONED()
            and from_pixels
            and torch.cuda.device_count() < 1
        ):
            raise pytest.skip("no cuda device")

        def non_null_obs(batched_td):
            if from_pixels:
                pix_norm = batched_td.get("pixels").flatten(-3, -1).float().norm(dim=-1)
                pix_norm_next = (
                    batched_td.get(("next", "pixels"))
                    .flatten(-3, -1)
                    .float()
                    .norm(dim=-1)
                )
                idx = (pix_norm > 1) & (pix_norm_next > 1)
                # eliminate batch size: all idx must be True (otherwise one could be filled with 0s)
                while idx.ndim > 1:
                    idx = idx.all(0)
                idx = idx.nonzero().squeeze(-1)
                assert idx.numel(), "Did not find pixels with norm > 1"
                return idx
            return slice(None)

        tdreset = []
        tdrollout = []
        final_seed = []
        for _ in range(2):
            env0 = GymEnv(
                env_name,
                frame_skip=frame_skip,
                from_pixels=from_pixels,
                pixels_only=pixels_only,
            )
            torch.manual_seed(0)
            np.random.seed(0)
            final_seed.append(env0.set_seed(0))
            tdreset.append(env0.reset())
            rollout = env0.rollout(max_steps=50)
            tdrollout.append(rollout)
            assert env0.from_pixels is from_pixels
            env0.close()
            env_type = type(env0._env)

        assert_allclose_td(*tdreset, rtol=RTOL, atol=ATOL)
        tdrollout = torch.stack(tdrollout, 0)

        # custom filtering of non-null obs: mujoco rendering sometimes fails
        # and renders black images. To counter this in the tests, we select
        # tensordicts with all non-null observations
        idx = non_null_obs(tdrollout)
        assert_allclose_td(
            tdrollout[0][..., idx], tdrollout[1][..., idx], rtol=RTOL, atol=ATOL
        )
        final_seed0, final_seed1 = final_seed
        assert final_seed0 == final_seed1

        if env_name == PONG_VERSIONED():
            base_env = gym_backend().make(env_name, frameskip=frame_skip)
            frame_skip = 1
        else:
            base_env = _make_gym_environment(env_name)

        if from_pixels and not _is_from_pixels(base_env):
            PixelObservationWrapper = get_gym_pixel_wrapper()
            base_env = PixelObservationWrapper(base_env, pixels_only=pixels_only)
        assert type(base_env) is env_type

        # Compare GymEnv output with GymWrapper output
        env1 = GymWrapper(base_env, frame_skip=frame_skip)
        assert env0.get_library_name(env0._env) == env1.get_library_name(env1._env)
        # check that we didn't do more wrapping
        assert type(env0._env) == type(env1._env)  # noqa: E721
        assert env0.output_spec == env1.output_spec
        assert env0.input_spec == env1.input_spec
        del env0
        torch.manual_seed(0)
        np.random.seed(0)
        final_seed2 = env1.set_seed(0)
        tdreset2 = env1.reset()
        rollout2 = env1.rollout(max_steps=50)
        assert env1.from_pixels is from_pixels
        env1.close()
        del env1, base_env

        assert_allclose_td(tdreset[0], tdreset2, rtol=RTOL, atol=ATOL)
        assert final_seed0 == final_seed2
        # same magic trick for mujoco as above
        tdrollout = torch.stack([tdrollout[0], rollout2], 0)
        idx = non_null_obs(tdrollout)
        assert_allclose_td(
            tdrollout[0][..., idx], tdrollout[1][..., idx], rtol=RTOL, atol=ATOL
        )

    @pytest.mark.parametrize(
        "env_name",
        [
            PONG_VERSIONED(),
            # PENDULUM_VERSIONED,
            HALFCHEETAH_VERSIONED(),
        ],
    )
    @pytest.mark.parametrize("frame_skip", [1, 3])
    @pytest.mark.parametrize(
        "from_pixels,pixels_only",
        [
            [False, False],
            [True, True],
            [True, False],
        ],
    )
    def test_gym_fake_td(self, env_name, frame_skip, from_pixels, pixels_only):
        if env_name == PONG_VERSIONED() and not from_pixels:
            # raise pytest.skip("already pixel")
            return
        elif (
            env_name != PONG_VERSIONED()
            and from_pixels
            and (not torch.has_cuda or not torch.cuda.device_count())
        ):
            raise pytest.skip("no cuda device")

        env = GymEnv(
            env_name,
            frame_skip=frame_skip,
            from_pixels=from_pixels,
            pixels_only=pixels_only,
        )
        check_env_specs(env)

    @pytest.mark.parametrize("frame_skip", [1, 3])
    @pytest.mark.parametrize(
        "from_pixels,pixels_only",
        [
            [False, False],
            [True, True],
            [True, False],
        ],
    )
    @pytest.mark.parametrize("wrapper", [True, False])
    def test_mo(self, frame_skip, from_pixels, pixels_only, wrapper):
        if importlib.util.find_spec("gymnasium") is not None and not _has_mo:
            raise pytest.skip("mo-gym not found")
        else:
            # avoid skipping, which we consider as errors in the gym CI
            return

        def make_env():
            import mo_gymnasium

            if wrapper:
                return MOGymWrapper(
                    mo_gymnasium.make("minecart-v0"),
                    frame_skip=frame_skip,
                    from_pixels=from_pixels,
                    pixels_only=pixels_only,
                )
            else:
                return MOGymEnv(
                    "minecart-v0",
                    frame_skip=frame_skip,
                    from_pixels=from_pixels,
                    pixels_only=pixels_only,
                )

        env = make_env()
        check_env_specs(env)
        env = SerialEnv(2, make_env)
        check_env_specs(env)

    def test_info_reader_mario(self):
        try:
            import gym_super_mario_bros as mario_gym
        except ImportError as err:
            try:
                gym = gym_backend()

                # with 0.26 we must have installed gym_super_mario_bros
                # Since we capture the skips as errors, we raise a skip in this case
                # Otherwise, we just return
                gym_version = version.parse(gym.__version__)
                if version.parse(
                    "0.26.0"
                ) <= gym_version and gym_version < version.parse("0.27"):
                    raise pytest.skip(f"no super mario bros: error=\n{err}")
            except ImportError:
                pass
            return

        gb = gym_backend()
        try:
            with set_gym_backend("gym"):
                env = mario_gym.make("SuperMarioBros-v0")
                env = GymWrapper(env)
                check_env_specs(env)

                def info_reader(info, tensordict):
                    assert isinstance(info, dict)  # failed before bugfix

                env.info_dict_reader = info_reader
                check_env_specs(env)
        finally:
            set_gym_backend(gb).set()

    @implement_for("gymnasium", "1.1.0")
    def test_one_hot_and_categorical(self):
        self._test_one_hot_and_categorical()

    @implement_for("gymnasium", None, "1.0.0")
    def test_one_hot_and_categorical(self):  # noqa
        self._test_one_hot_and_categorical()

    def _test_one_hot_and_categorical(self):
        # tests that one-hot and categorical work ok when an integer is expected as action
        cliff_walking = GymEnv("CliffWalking-v0", categorical_action_encoding=True)
        cliff_walking.rollout(10)
        check_env_specs(cliff_walking)

        cliff_walking = GymEnv("CliffWalking-v0", categorical_action_encoding=False)
        cliff_walking.rollout(10)
        check_env_specs(cliff_walking)

    @implement_for("gym")
    def test_one_hot_and_categorical(self):  # noqa: F811
        # we do not skip (bc we may want to make sure nothing is skipped)
        # but CliffWalking-v0 in earlier Gym versions uses np.bool, which
        # was deprecated after np 1.20, and we don't want to install multiple np
        # versions.
        return

    @implement_for("gymnasium", "1.1.0")
    @pytest.mark.parametrize(
        "envname",
        ["HalfCheetah-v4", "CartPole-v1", "ALE/Pong-v5"]
        + (["FetchReach-v2"] if _has_gym_robotics else []),
    )
    @pytest.mark.flaky(reruns=5, reruns_delay=1)
    def test_vecenvs_wrapper(self, envname):
        import gymnasium

        with set_gym_backend("gymnasium"):
            self._test_vecenvs_wrapper(
                envname,
                kwargs={"autoreset_mode": gymnasium.vector.AutoresetMode.SAME_STEP},
            )

    @implement_for("gymnasium", None, "1.0.0")
    @pytest.mark.parametrize(
        "envname",
        ["HalfCheetah-v4", "CartPole-v1", "ALE/Pong-v5"]
        + (["FetchReach-v2"] if _has_gym_robotics else []),
    )
    @pytest.mark.flaky(reruns=5, reruns_delay=1)
    def test_vecenvs_wrapper(self, envname):  # noqa
        with set_gym_backend("gymnasium"):
            self._test_vecenvs_wrapper(envname)

    def _test_vecenvs_wrapper(self, envname, kwargs=None):
        import gymnasium

        if kwargs is None:
            kwargs = {}
        # we can't use parametrize with implement_for
        env = GymWrapper(
            gymnasium.vector.SyncVectorEnv(
                2 * [lambda envname=envname: gymnasium.make(envname)], **kwargs
            )
        )
        assert env.batch_size == torch.Size([2])
        check_env_specs(env)
        env = GymWrapper(
            gymnasium.vector.AsyncVectorEnv(
                2 * [lambda envname=envname: gymnasium.make(envname)], **kwargs
            )
        )
        assert env.batch_size == torch.Size([2])
        check_env_specs(env)

    @implement_for("gymnasium", "1.1.0")
    # this env has Dict-based observation which is a nice thing to test
    @pytest.mark.parametrize(
        "envname",
        ["HalfCheetah-v4", "CartPole-v1", "ALE/Pong-v5"]
        + (["FetchReach-v2"] if _has_gym_robotics else []),
    )
    @pytest.mark.flaky(reruns=5, reruns_delay=1)
    def test_vecenvs_env(self, envname):
        self._test_vecenvs_env(envname)

    @implement_for("gymnasium", None, "1.0.0")
    # this env has Dict-based observation which is a nice thing to test
    @pytest.mark.parametrize(
        "envname",
        ["HalfCheetah-v4", "CartPole-v1", "ALE/Pong-v5"]
        + (["FetchReach-v2"] if _has_gym_robotics else []),
    )
    @pytest.mark.flaky(reruns=5, reruns_delay=1)
    def test_vecenvs_env(self, envname):  # noqa
        self._test_vecenvs_env(envname)

    def _test_vecenvs_env(self, envname):

        gb = gym_backend()
        try:
            with set_gym_backend("gymnasium"):
                env = GymEnv(envname, num_envs=2, from_pixels=False)
                env.set_seed(0)
                assert env.get_library_name(env._env) == "gymnasium"
            # rollouts can be executed without decorator
            check_env_specs(env)
            rollout = env.rollout(100, break_when_any_done=False)
            for obs_key in env.observation_spec.keys(True, True):
                rollout_consistency_assertion(
                    rollout,
                    done_key="done",
                    observation_key=obs_key,
                    done_strict="CartPole" in envname,
                )
            env.close()
            del env
        finally:
            set_gym_backend(gb).set()

    @implement_for("gym", "0.18")
    @pytest.mark.parametrize(
        "envname",
        ["CartPole-v1", "HalfCheetah-v4"],
    )
    @pytest.mark.flaky(reruns=5, reruns_delay=1)
    def test_vecenvs_wrapper(self, envname):  # noqa: F811
        with set_gym_backend("gym"):
            gym = gym_backend()
            # we can't use parametrize with implement_for
            for envname in ["CartPole-v1", "HalfCheetah-v4"]:
                env = GymWrapper(
                    gym.vector.SyncVectorEnv(
                        2 * [lambda envname=envname: gym.make(envname)]
                    )
                )
                assert env.batch_size == torch.Size([2])
                check_env_specs(env)
                env = GymWrapper(
                    gym.vector.AsyncVectorEnv(
                        2 * [lambda envname=envname: gym.make(envname)]
                    )
                )
                assert env.batch_size == torch.Size([2])
                check_env_specs(env)
                env.close()
                del env

    @implement_for("gym", "0.18")
    @pytest.mark.parametrize(
        "envname",
        ["cp", "hc"],
    )
    @pytest.mark.flaky(reruns=5, reruns_delay=1)
    def test_vecenvs_env(self, envname):  # noqa: F811
        gb = gym_backend()
        try:
            with set_gym_backend("gym"):
                if envname == "hc":
                    envname = HALFCHEETAH_VERSIONED()
                else:
                    envname = CARTPOLE_VERSIONED()
                env = GymEnv(envname, num_envs=2, from_pixels=False)
                env.set_seed(0)
                assert env.get_library_name(env._env) == "gym"
                # rollouts can be executed without decorator
                check_env_specs(env)
                rollout = env.rollout(100, break_when_any_done=False)
                for obs_key in env.observation_spec.keys(True, True):
                    rollout_consistency_assertion(
                        rollout,
                        done_key="done",
                        observation_key=obs_key,
                        done_strict="CartPole" in envname,
                    )
                env.close()
            del env
            if envname != "CartPole-v1":
                with set_gym_backend("gym"):
                    env = GymEnv(envname, num_envs=2, from_pixels=True)
                    env.set_seed(0)
                # rollouts can be executed without decorator
                check_env_specs(env)
                env.close()
                del env
        finally:
            set_gym_backend(gb).set()

    @implement_for("gym", None, "0.18")
    @pytest.mark.parametrize(
        "envname",
        ["CartPole-v1", "HalfCheetah-v4"],
    )
    def test_vecenvs_wrapper(self, envname):  # noqa: F811
        # skipping tests for older versions of gym
        ...

    @implement_for("gym", None, "0.18")
    @pytest.mark.parametrize(
        "envname",
        ["CartPole-v1", "HalfCheetah-v4"],
    )
    def test_vecenvs_env(self, envname):  # noqa: F811
        # skipping tests for older versions of gym
        ...

    @implement_for("gym", None, "0.26")
    @pytest.mark.parametrize("wrapper", [True, False])
    def test_gym_output_num(self, wrapper):
        # gym has 4 outputs, no truncation
        gym = gym_backend()
        try:
            if wrapper:
                env = GymWrapper(gym.make(PENDULUM_VERSIONED()))
            else:
                with set_gym_backend("gym"):
                    env = GymEnv(PENDULUM_VERSIONED())
            # truncated is read from the info
            assert "truncated" in env.done_keys
            assert "terminated" in env.done_keys
            assert "done" in env.done_keys
            check_env_specs(env)
        finally:
            set_gym_backend(gym).set()

    @implement_for("gym", "0.26")
    @pytest.mark.parametrize("wrapper", [True, False])
    def test_gym_output_num(self, wrapper):  # noqa: F811
        # gym has 5 outputs, with truncation
        gym = gym_backend()
        try:
            if wrapper:
                env = GymWrapper(gym.make(PENDULUM_VERSIONED()))
            else:
                with set_gym_backend("gym"):
                    env = GymEnv(PENDULUM_VERSIONED())
            assert "truncated" in env.done_keys
            assert "terminated" in env.done_keys
            assert "done" in env.done_keys
            check_env_specs(env)

            if wrapper:
                # let's further test with a wrapper that exposes the env with old API
                from gym.wrappers.compatibility import EnvCompatibility

                with pytest.raises(
                    ValueError,
                    match="GymWrapper does not support the gym.wrapper.compatibility.EnvCompatibility",
                ):
                    GymWrapper(EnvCompatibility(gym.make("CartPole-v1")))
        finally:
            set_gym_backend(gym).set()

    @implement_for("gymnasium", "1.1.0")
    @pytest.mark.parametrize("wrapper", [True, False])
    def test_gym_output_num(self, wrapper):  # noqa: F811
        self._test_gym_output_num(wrapper)

    @implement_for("gymnasium", None, "1.0.0")
    @pytest.mark.parametrize("wrapper", [True, False])
    def test_gym_output_num(self, wrapper):  # noqa: F811
        self._test_gym_output_num(wrapper)

    def _test_gym_output_num(self, wrapper):  # noqa: F811
        # gym has 5 outputs, with truncation
        gym = gym_backend()
        try:
            if wrapper:
                env = GymWrapper(gym.make(PENDULUM_VERSIONED()))
            else:
                with set_gym_backend("gymnasium"):
                    env = GymEnv(PENDULUM_VERSIONED())
            assert "truncated" in env.done_keys
            assert "terminated" in env.done_keys
            assert "done" in env.done_keys
            check_env_specs(env)
        finally:
            set_gym_backend(gym).set()

    def test_gym_gymnasium_parallel(self, maybe_fork_ParallelEnv):
        # tests that both gym and gymnasium work with wrappers without
        # decorating with set_gym_backend during execution
        gym = gym_backend()
        try:
            if importlib.util.find_spec("gym") is not None:
                with set_gym_backend("gym"):
                    gym = gym_backend()

                old_api = version.parse(gym.__version__) < version.parse("0.26")
                make_fun = EnvCreator(
                    lambda: GymWrapper(gym.make(PENDULUM_VERSIONED()))
                )
            elif importlib.util.find_spec("gymnasium") is not None:
                import gymnasium

                old_api = False
                make_fun = EnvCreator(
                    lambda: GymWrapper(gymnasium.make(PENDULUM_VERSIONED()))
                )
            else:
                raise ImportError  # unreachable under pytest.skipif
            penv = maybe_fork_ParallelEnv(2, make_fun)
            rollout = penv.rollout(2)
            if old_api:
                assert "terminated" in rollout.keys()
                # truncated is read from info
                assert "truncated" in rollout.keys()
            else:
                assert "terminated" in rollout.keys()
                assert "truncated" in rollout.keys()
            check_env_specs(penv)
        finally:
            set_gym_backend(gym).set()

    @implement_for("gym", None, "0.22.0")
    def test_vecenvs_nan(self):  # noqa: F811
        # old versions of gym must return nan for next values when there is a done state
        torch.manual_seed(0)
        env = GymEnv("CartPole-v0", num_envs=2)
        env.set_seed(0)
        rollout = env.rollout(200)
        assert torch.isfinite(rollout.get("observation")).all()
        assert not torch.isfinite(rollout.get(("next", "observation"))).all()
        env.close()
        del env

        # same with collector
        env = GymEnv("CartPole-v0", num_envs=2)
        env.set_seed(0)
        c = SyncDataCollector(
            env, RandomPolicy(env.action_spec), total_frames=2000, frames_per_batch=200
        )
        for rollout in c:
            assert torch.isfinite(rollout.get("observation")).all()
            assert not torch.isfinite(rollout.get(("next", "observation"))).all()
            break
        del c
        return

    @implement_for("gym", "0.22.0", None)
    def test_vecenvs_nan(self):  # noqa: F811
        # new versions of gym must never return nan for next values when there is a done state
        torch.manual_seed(0)
        env = GymEnv("CartPole-v0", num_envs=2)
        env.set_seed(0)
        rollout = env.rollout(200)
        assert torch.isfinite(rollout.get("observation")).all()
        assert torch.isfinite(rollout.get(("next", "observation"))).all()
        env.close()
        del env

        # same with collector
        env = GymEnv("CartPole-v0", num_envs=2)
        env.set_seed(0)
        c = SyncDataCollector(
            env, RandomPolicy(env.action_spec), total_frames=2000, frames_per_batch=200
        )
        for rollout in c:
            assert torch.isfinite(rollout.get("observation")).all()
            assert torch.isfinite(rollout.get(("next", "observation"))).all()
            break
        del c
        return

    @implement_for("gymnasium", "1.1.0")
    def test_vecenvs_nan(self):  # noqa: F811
        self._test_vecenvs_nan()

    @implement_for("gymnasium", None, "1.0.0")
    def test_vecenvs_nan(self):  # noqa: F811
        self._test_vecenvs_nan()

    def _test_vecenvs_nan(self):  # noqa: F811
        # new versions of gym must never return nan for next values when there is a done state
        torch.manual_seed(0)
        env = GymEnv("CartPole-v1", num_envs=2)
        env.set_seed(0)
        rollout = env.rollout(200)
        assert torch.isfinite(rollout.get("observation")).all()
        assert torch.isfinite(rollout.get(("next", "observation"))).all()
        env.close()
        del env

        # same with collector
        env = GymEnv("CartPole-v1", num_envs=2)
        env.set_seed(0)
        c = SyncDataCollector(
            env, RandomPolicy(env.action_spec), total_frames=2000, frames_per_batch=200
        )
        for rollout in c:
            assert torch.isfinite(rollout.get("observation")).all()
            assert torch.isfinite(rollout.get(("next", "observation"))).all()
            break
        del c
        return

    def _get_dummy_gym_env(self, backend, **kwargs):
        with set_gym_backend(backend):

            class CustomEnv(gym_backend().Env):
                def __init__(self, dim=3, use_termination=True, max_steps=4):
                    self.dim = dim
                    self.use_termination = use_termination
                    self.observation_space = gym_backend("spaces").Box(
                        low=-np.inf, high=np.inf, shape=(self.dim,), dtype=np.float32
                    )
                    self.action_space = gym_backend("spaces").Box(
                        low=-np.inf, high=np.inf, shape=(1,), dtype=np.float32
                    )
                    self.max_steps = max_steps

                def _get_info(self):
                    return {"field1": self.state**2}

                def _get_obs(self):
                    return self.state.copy()

                def reset(self, seed=0, options=None):
                    self.state = np.zeros(
                        self.observation_space.shape, dtype=np.float32
                    )
                    observation = self._get_obs()
                    info = self._get_info()
                    assert (observation < self.max_steps).all()
                    return observation, info

                def step(self, action):
                    # self.state += action.item()
                    self.state += 1
                    truncated, terminated = False, False
                    if self.use_termination:
                        terminated = self.state[0] == 4
                    reward = 1 if terminated else 0  # Binary sparse rewards
                    observation = self._get_obs()
                    info = self._get_info()
                    return observation, reward, terminated, truncated, info

            return CustomEnv(**kwargs)

    def counting_env(self):
        import gymnasium as gym
        from gymnasium import Env

        class CountingEnvRandomReset(Env):
            def __init__(self, i=0):
                self.counter = 1
                self.i = i
                self.observation_space = gym.spaces.Box(-np.inf, np.inf, shape=(1,))
                self.action_space = gym.spaces.Box(-np.inf, np.inf, shape=(1,))
                self.rng = np.random.RandomState(0)

            def step(self, action):
                self.counter += 1
                done = bool(self.rng.random() < 0.05)
                return (
                    np.asarray(
                        [
                            self.counter,
                        ]
                    ),
                    0,
                    done,
                    done,
                    {},
                )

            def reset(
                self,
                *,
                seed: int | None = None,
                options=None,
            ):
                self.counter = 1
                if seed is not None:
                    self.rng = np.random.RandomState(seed)
                return (
                    np.asarray(
                        [
                            self.counter,
                        ]
                    ),
                    {},
                )

        return CountingEnvRandomReset

    @implement_for("gym")
    def test_gymnasium_autoreset(self, venv):
        return

    @implement_for("gymnasium", None, "1.1.0")
    def test_gymnasium_autoreset(self, venv):  # noqa
        return

    @implement_for("gymnasium", "1.1.0")
    @pytest.mark.parametrize("venv", ["sync", "async"])
    def test_gymnasium_autoreset(self, venv):  # noqa
        import gymnasium as gym

        set_gym_backend("gymnasium").set()

        counting_env = self.counting_env()
        if venv == "sync":
            venv = gym.vector.SyncVectorEnv
        else:
            venv = gym.vector.AsyncVectorEnv
        envs0 = venv(
            [lambda i=i: counting_env(i) for i in range(2)],
            autoreset_mode=gym.vector.AutoresetMode.DISABLED,
        )
        env = GymWrapper(envs0)
        envs0.reset(seed=0)
        torch.manual_seed(0)
        r0 = env.rollout(20, break_when_any_done=False)
        envs1 = venv(
            [lambda i=i: counting_env(i) for i in range(2)],
            autoreset_mode=gym.vector.AutoresetMode.SAME_STEP,
        )
        env = GymWrapper(envs1)
        envs1.reset(seed=0)
        # env.set_seed(0)
        torch.manual_seed(0)
        r1 = []
        t_ = env.reset()
        for s in r0.unbind(-1):
            t_.set("action", s["action"])
            t, t_ = env.step_and_maybe_reset(t_)
            r1.append(t)
        r1 = torch.stack(r1, -1)
        torch.testing.assert_close(r0["observation"], r1["observation"])
        torch.testing.assert_close(r0["next", "observation"], r1["next", "observation"])
        torch.testing.assert_close(r0["next", "done"], r1["next", "done"])

    @implement_for("gym")
    @pytest.mark.parametrize("heterogeneous", [False, True])
    def test_resetting_strategies(self, heterogeneous):
        return

    @implement_for("gymnasium", None, "1.0.0")
    @pytest.mark.parametrize("heterogeneous", [False, True])
    def test_resetting_strategies(self, heterogeneous):  # noqa
        self._test_resetting_strategies(heterogeneous, {})

    @implement_for("gymnasium", "1.1.0")
    @pytest.mark.parametrize("heterogeneous", [False, True])
    def test_resetting_strategies(self, heterogeneous):  # noqa
        import gymnasium as gym

        self._test_resetting_strategies(
            heterogeneous, {"autoreset_mode": gym.vector.AutoresetMode.SAME_STEP}
        )

    def _test_resetting_strategies(self, heterogeneous, kwargs):
        if _has_gymnasium:
            backend = "gymnasium"
        else:
            backend = "gym"
        with set_gym_backend(backend):
            if version.parse(gym_backend().__version__) < version.parse("0.26"):
                torchrl_logger.info(
                    "Running into unrelated errors with older versions of gym."
                )
                return
            steps = 5
            if not heterogeneous:
                env = GymWrapper(
                    gym_backend().vector.AsyncVectorEnv(
                        [functools.partial(self._get_dummy_gym_env, backend=backend)]
                        * 4,
                        **kwargs,
                    )
                )
            else:
                env = GymWrapper(
                    gym_backend().vector.AsyncVectorEnv(
                        [
                            functools.partial(
                                self._get_dummy_gym_env,
                                max_steps=i + 4,
                                backend=backend,
                            )
                            for i in range(4)
                        ],
                        **kwargs,
                    )
                )
            try:
                check_env_specs(env)
                td = env.rollout(steps, break_when_any_done=False)
                if not heterogeneous:
                    assert not (td["observation"] == 4).any()
                    assert (td["next", "observation"] == 4).sum() == 3 * 4

                # check with manual reset
                torch.manual_seed(0)
                env.set_seed(0)
                reset = env.reset(
                    TensorDict({"_reset": torch.ones(4, 1, dtype=torch.bool)}, [4])
                )
                r0 = env.rollout(
                    10, break_when_any_done=False, auto_reset=False, tensordict=reset
                )
                torch.manual_seed(0)
                env.set_seed(0)
                reset = env.reset()
                r1 = env.rollout(
                    10, break_when_any_done=False, auto_reset=False, tensordict=reset
                )
                torch.manual_seed(0)
                env.set_seed(0)
                r2 = env.rollout(10, break_when_any_done=False)
                assert_allclose_td(r0, r1)
                assert_allclose_td(r1, r2)
                for r in (r0, r1, r2):
                    torch.testing.assert_close(r["field1"], r["observation"].pow(2))
                    torch.testing.assert_close(
                        r["next", "field1"], r["next", "observation"].pow(2)
                    )

            finally:
                if not env.is_closed:
                    env.close()
                del env
                gc.collect()


@pytest.mark.skipif(
    not _has_minigrid or not _has_gymnasium, reason="MiniGrid not found"
)
class TestMiniGrid:
    @pytest.mark.parametrize(
        "id",
        [
            "BabyAI-KeyCorridorS6R3-v0",
            "MiniGrid-Empty-16x16-v0",
            "MiniGrid-BlockedUnlockPickup-v0",
        ],
    )
    def test_minigrid(self, id):
        env_base = gymnasium.make(id)
        env = GymWrapper(env_base)
        check_env_specs(env)


@implement_for("gym", None, "0.26")
def _make_gym_environment(env_name):  # noqa: F811
    gym = gym_backend()
    return gym.make(env_name)


@implement_for("gym", "0.26", None)
def _make_gym_environment(env_name):  # noqa: F811
    gym = gym_backend()
    return gym.make(env_name, render_mode="rgb_array")


@implement_for("gymnasium", None, "1.0.0")
def _make_gym_environment(env_name):  # noqa: F811
    gym = gym_backend()
    return gym.make(env_name, render_mode="rgb_array")


@implement_for("gymnasium", "1.1.0")
def _make_gym_environment(env_name):  # noqa: F811
    gym = gym_backend()
    return gym.make(env_name, render_mode="rgb_array")


@pytest.mark.skipif(not _has_dmc, reason="no dm_control library found")
class TestDMControl:
    @pytest.mark.parametrize("env_name,task", [["cheetah", "run"]])
    @pytest.mark.parametrize("frame_skip", [1, 3])
    @pytest.mark.parametrize(
        "from_pixels,pixels_only", [[True, True], [True, False], [False, False]]
    )
    def test_dmcontrol(self, env_name, task, frame_skip, from_pixels, pixels_only):
        if from_pixels and (not torch.has_cuda or not torch.cuda.device_count()):
            raise pytest.skip("no cuda device")

        tds = []
        tds_reset = []
        final_seed = []
        for _ in range(2):
            env0 = DMControlEnv(
                env_name,
                task,
                frame_skip=frame_skip,
                from_pixels=from_pixels,
                pixels_only=pixels_only,
            )
            torch.manual_seed(0)
            np.random.seed(0)
            final_seed0 = env0.set_seed(0)
            tdreset0 = env0.reset()
            rollout0 = env0.rollout(max_steps=50)
            env0.close()
            del env0
            tds_reset.append(tdreset0)
            tds.append(rollout0)
            final_seed.append(final_seed0)

        tdreset1, tdreset0 = tds_reset
        rollout0, rollout1 = tds
        final_seed0, final_seed1 = final_seed

        assert_allclose_td(tdreset1, tdreset0)
        assert final_seed0 == final_seed1
        assert_allclose_td(rollout0, rollout1)

        env1 = DMControlEnv(
            env_name,
            task,
            frame_skip=frame_skip,
            from_pixels=from_pixels,
            pixels_only=pixels_only,
        )
        torch.manual_seed(1)
        np.random.seed(1)
        final_seed1 = env1.set_seed(1)
        tdreset1 = env1.reset()
        rollout1 = env1.rollout(max_steps=50)
        env1.close()
        del env1

        with pytest.raises(AssertionError):
            assert_allclose_td(tdreset1, tdreset0)
            assert final_seed0 == final_seed1
            assert_allclose_td(rollout0, rollout1)

        from dm_control import suite

        base_env = suite.load(env_name, task)
        if from_pixels:
            from dm_control.suite.wrappers import pixels

            render_kwargs = {"camera_id": 0}
            base_env = pixels.Wrapper(
                base_env, pixels_only=pixels_only, render_kwargs=render_kwargs
            )
        env2 = DMControlWrapper(base_env, frame_skip=frame_skip)
        torch.manual_seed(0)
        np.random.seed(0)
        final_seed2 = env2.set_seed(0)
        tdreset2 = env2.reset()
        rollout2 = env2.rollout(max_steps=50)

        assert_allclose_td(tdreset0, tdreset2)
        assert final_seed0 == final_seed2
        assert_allclose_td(rollout0, rollout2)

    @pytest.mark.skipif(not torch.cuda.is_available(), reason="requires cuda")
    @pytest.mark.parametrize("env_name,task", [["cheetah", "run"]])
    @pytest.mark.parametrize("frame_skip", [1, 3])
    @pytest.mark.parametrize(
        "from_pixels,pixels_only", [[True, True], [True, False], [False, False]]
    )
    def test_dmcontrol_device_consistency(
        self, env_name, task, frame_skip, from_pixels, pixels_only
    ):
        env0 = DMControlEnv(
            env_name,
            task,
            frame_skip=frame_skip,
            from_pixels=from_pixels,
            pixels_only=pixels_only,
            device="cpu",
        )

        env1 = DMControlEnv(
            env_name,
            task,
            frame_skip=frame_skip,
            from_pixels=from_pixels,
            pixels_only=pixels_only,
            device="cuda",
        )

        env0.set_seed(0)
        r0 = env0.rollout(100, break_when_any_done=False)
        assert r0.device == torch.device("cpu")
        actions = collections.deque(r0["action"].unbind(0))
        policy = lambda td: td.set("action", actions.popleft())
        env1.set_seed(0)
        r1 = env1.rollout(100, policy, break_when_any_done=False)
        assert r1.device == torch.device("cuda:0")
        assert_allclose_td(r0, r1.cpu())

    @pytest.mark.parametrize("env_name,task", [["cheetah", "run"]])
    @pytest.mark.parametrize("frame_skip", [1, 3])
    @pytest.mark.parametrize(
        "from_pixels,pixels_only", [[True, True], [True, False], [False, False]]
    )
    def test_faketd(self, env_name, task, frame_skip, from_pixels, pixels_only):
        if from_pixels and not torch.cuda.device_count():
            raise pytest.skip("no cuda device")

        env = DMControlEnv(
            env_name,
            task,
            frame_skip=frame_skip,
            from_pixels=from_pixels,
            pixels_only=pixels_only,
        )
        check_env_specs(env)

    def test_truncated(self):
        env = DMControlEnv("walker", "walk")
        r = env.rollout(1001)
        assert r.shape == (1000,)
        assert r[-1]["next", "truncated"]
        assert r[-1]["next", "done"]
        assert not r[-1]["next", "terminated"]


params = []
if _has_dmc:
    params = [
        # [DMControlEnv, ("cheetah", "run"), {"from_pixels": True}],
        [DMControlEnv, ("cheetah", "run"), {"from_pixels": False}],
    ]
if _has_gym:
    params += [
        # [GymEnv, (HALFCHEETAH_VERSIONED,), {"from_pixels": True}],
        [GymEnv, (HALFCHEETAH_VERSIONED(),), {"from_pixels": False}],
        [GymEnv, (PONG_VERSIONED(),), {}],
    ]


@pytest.mark.skipif(
    IS_OSX,
    reason="rendering unstable on osx, skipping (mujoco.FatalError: gladLoadGL error)",
)
@pytest.mark.parametrize("env_lib,env_args,env_kwargs", params)
def test_td_creation_from_spec(env_lib, env_args, env_kwargs):
    if (
        gym_version < version.parse("0.26.0")
        and env_kwargs.get("from_pixels", False)
        and torch.cuda.device_count() == 0
    ):
        raise pytest.skip(
            "Skipping test as rendering is not supported in tests before gym 0.26."
        )
    env = env_lib(*env_args, **env_kwargs)
    td = env.rollout(max_steps=5)
    td0 = td[0]
    fake_td = env.fake_tensordict()

    assert set(fake_td.keys(include_nested=True, leaves_only=True)) == set(
        td.keys(include_nested=True, leaves_only=True)
    )
    for key in fake_td.keys(include_nested=True, leaves_only=True):
        assert fake_td.get(key).shape == td.get(key)[0].shape
    for key in fake_td.keys(include_nested=True, leaves_only=True):
        assert fake_td.get(key).shape == td0.get(key).shape
        assert fake_td.get(key).dtype == td0.get(key).dtype
        assert fake_td.get(key).device == td0.get(key).device


params = []
if _has_dmc:
    params += [
        # [DMControlEnv, ("cheetah", "run"), {"from_pixels": True}],
        [DMControlEnv, ("cheetah", "run"), {"from_pixels": False}],
    ]
if _has_gym:
    params += [
        # [GymEnv, (HALFCHEETAH_VERSIONED,), {"from_pixels": True}],
        [GymEnv, (HALFCHEETAH_VERSIONED,), {"from_pixels": False}],
        # [GymEnv, (PONG_VERSIONED,), {}],  # 1226: skipping
    ]


# @pytest.mark.skipif(IS_OSX, reason="rendering unstable on osx, skipping")
@pytest.mark.parametrize("env_lib,env_args,env_kwargs", params)
@pytest.mark.parametrize(
    "device",
    [torch.device("cuda:0") if torch.cuda.device_count() else torch.device("cpu")],
)
class TestCollectorLib:
    def test_collector_run(self, env_lib, env_args, env_kwargs, device):
        env_args = tuple(arg() if callable(arg) else arg for arg in env_args)
        if not _has_dmc and env_lib is DMControlEnv:
            raise pytest.skip("no dmc")
        if not _has_gym and env_lib is GymEnv:
            raise pytest.skip("no gym")

        from_pixels = env_kwargs.get("from_pixels", False)
        if from_pixels and (not torch.has_cuda or not torch.cuda.device_count()):
            raise pytest.skip("no cuda device")

        env_fn = EnvCreator(lambda: env_lib(*env_args, **env_kwargs, device=device))
        env = SerialEnv(3, env_fn)
        # env = ParallelEnv(3, env_fn)  # 1226: Serial for efficiency reasons
        # check_env_specs(env)

        # env = ParallelEnv(3, env_fn)
        frames_per_batch = 21
        collector = SyncDataCollector(  # 1226: not using MultiaSync for perf reasons
            create_env_fn=env,
            policy=RandomPolicy(action_spec=env.action_spec),
            total_frames=-1,
            max_frames_per_traj=100,
            frames_per_batch=frames_per_batch,
            init_random_frames=-1,
            reset_at_each_iter=False,
            split_trajs=True,
            device=device,
            storing_device=device,
            exploration_type=ExplorationType.RANDOM,
        )
        for i, _data in enumerate(collector):
            if i == 3:
                break
        collector.shutdown()
        assert _data.shape[1] == -(frames_per_batch // -env.num_workers)
        assert _data.shape[0] == frames_per_batch // _data.shape[1]
        del env


@pytest.mark.skipif(not _has_habitat, reason="habitat not installed")
@pytest.mark.parametrize("envname", ["HabitatRenderPick-v0", "HabitatPick-v0"])
class TestHabitat:
    def test_habitat(self, envname):
        env = HabitatEnv(envname)
        _ = env.rollout(3)
        check_env_specs(env)

    @pytest.mark.parametrize("from_pixels", [True, False])
    def test_habitat_render(self, envname, from_pixels):
        env = HabitatEnv(envname, from_pixels=from_pixels)
        rollout = env.rollout(3)
        check_env_specs(env)
        if from_pixels:
            assert "pixels" in rollout.keys()


def _jumanji_envs():
    if not _has_jumanji:
        return ()
    return JumanjiEnv.available_envs[-10:-5]


@pytest.mark.skipif(not _has_jumanji, reason="jumanji not installed")
@pytest.mark.slow
@pytest.mark.parametrize("envname", _jumanji_envs())
class TestJumanji:
    def test_jumanji_seeding(self, envname):
        final_seed = []
        tdreset = []
        tdrollout = []
        for _ in range(2):
            env = JumanjiEnv(envname)
            torch.manual_seed(0)
            np.random.seed(0)
            final_seed.append(env.set_seed(0))
            tdreset.append(env.reset())
            rollout = env.rollout(max_steps=50)
            tdrollout.append(rollout)
            env.close()
            del env
        assert final_seed[0] == final_seed[1]
        assert_allclose_td(*tdreset)
        assert_allclose_td(*tdrollout)

    @pytest.mark.parametrize("batch_size", [(), (5,), (5, 4)])
    def test_jumanji_batch_size(self, envname, batch_size):
        env = JumanjiEnv(envname, batch_size=batch_size, jit=True)
        env.set_seed(0)
        tdreset = env.reset()
        tdrollout = env.rollout(max_steps=50)
        env.close()
        del env
        assert tdreset.batch_size == batch_size
        assert tdrollout.batch_size[:-1] == batch_size

    @pytest.mark.parametrize("batch_size", [(), (5,), (5, 4)])
    def test_jumanji_spec_rollout(self, envname, batch_size):
        env = JumanjiEnv(envname, batch_size=batch_size, jit=True)
        env.set_seed(0)
        check_env_specs(env)

    @pytest.mark.parametrize("batch_size", [(), (5,), (5, 4)])
    def test_jumanji_consistency(self, envname, batch_size):
        import jax
        import jax.numpy as jnp
        import numpy as onp
        from torchrl.envs.libs.jax_utils import _tree_flatten

        env = JumanjiEnv(envname, batch_size=batch_size, jit=True)
        obs_keys = list(env.observation_spec.keys(True))
        env.set_seed(1)
        rollout = env.rollout(10)

        env.set_seed(1)
        key = env.key
        base_env = env._env
        key, *keys = jax.random.split(key, int(np.prod(batch_size) + 1))
        state, timestep = jax.vmap(base_env.reset)(jnp.stack(keys))
        # state = env._reshape(state)
        # timesteps.append(timestep)
        for i in range(rollout.shape[-1]):
            action = rollout[..., i]["action"]
            # state = env._flatten(state)
            action = _tree_flatten(env.read_action(action), env.batch_size)
            state, timestep = jax.vmap(base_env.step)(state, action)
            # state = env._reshape(state)
            # timesteps.append(timestep)
            for _key in obs_keys:
                if isinstance(_key, str):
                    _key = (_key,)
                try:
                    t2 = getattr(timestep, _key[0])
                except AttributeError:
                    try:
                        t2 = getattr(timestep.observation, _key[0])
                    except AttributeError:
                        continue
                t1 = rollout[..., i][("next", *_key)]
                for __key in _key[1:]:
                    t2 = getattr(t2, _key)
                t2 = torch.tensor(onp.asarray(t2)).view_as(t1)
                torch.testing.assert_close(t1, t2)

    @pytest.mark.parametrize("batch_size", [[3], []])
    def test_jumanji_rendering(self, envname, batch_size):
        # check that this works with a batch-size
        env = JumanjiEnv(envname, from_pixels=True, batch_size=batch_size, jit=True)
        env.set_seed(0)
        env.transform.transform_observation_spec(env.base_env.observation_spec.clone())

        r = env.rollout(10)
        pixels = r["pixels"]
        if not isinstance(pixels, torch.Tensor):
            pixels = torch.as_tensor(np.asarray(pixels))
            assert batch_size
        else:
            assert not batch_size
        assert pixels.unique().numel() > 1
        assert pixels.dtype == torch.uint8

        check_env_specs(env)

    @pytest.mark.parametrize("jit", [True, False])
    def test_jumanji_batch_unlocked(self, envname, jit):
        torch.manual_seed(0)
        env = JumanjiEnv(envname, jit=jit)
        env.set_seed(0)
        assert not env.batch_locked
        reset = env.reset(TensorDict(batch_size=[16]))
        assert reset.batch_size == (16,)
        env.rand_step(reset)
        r = env.rollout(
            2000, auto_reset=False, tensordict=reset, break_when_all_done=True
        )
        assert r.batch_size[0] == 16
        done = r["next", "done"]
        assert done.any(-2).all() or (r.shape[-1] == 2000)


ENVPOOL_CLASSIC_CONTROL_ENVS = [
    PENDULUM_VERSIONED(),
    "MountainCar-v0",
    "MountainCarContinuous-v0",
    "Acrobot-v1",
    CARTPOLE_VERSIONED(),
]
ENVPOOL_ATARI_ENVS = []  # PONG_VERSIONED]
ENVPOOL_GYM_ENVS = ENVPOOL_CLASSIC_CONTROL_ENVS + ENVPOOL_ATARI_ENVS
ENVPOOL_DM_ENVS = ["CheetahRun-v1"]
ENVPOOL_ALL_ENVS = ENVPOOL_GYM_ENVS + ENVPOOL_DM_ENVS


@pytest.mark.skipif(not _has_envpool, reason="No envpool library found")
class TestEnvPool:
    def test_lib(self):
        import envpool

        assert MultiThreadedEnvWrapper.lib is envpool

    @pytest.mark.parametrize("env_name", ENVPOOL_ALL_ENVS)
    def test_env_wrapper_creation(self, env_name):
        env_name = env_name.replace("ALE/", "")  # EnvPool naming convention
        envpool_env = envpool.make(
            task_id=env_name, env_type="gym", num_envs=4, gym_reset_return_info=True
        )
        env = MultiThreadedEnvWrapper(envpool_env)
        env.reset()
        env.rand_step()

    @pytest.mark.skipif(not _has_gym, reason="no gym")
    @pytest.mark.parametrize(
        "env_name", ENVPOOL_GYM_ENVS
    )  # Not working for CheetahRun-v1 yet
    @pytest.mark.parametrize("frame_skip", [4, 1])
    @pytest.mark.parametrize("transformed_out", [False, True])
    def test_specs(self, env_name, frame_skip, transformed_out, T=10, N=3):
        env_multithreaded = _make_multithreaded_env(
            env_name,
            frame_skip,
            transformed_out=transformed_out,
            N=N,
        )
        check_env_specs(env_multithreaded)

    @pytest.mark.skipif(not _has_gym, reason="no gym")
    @pytest.mark.parametrize("env_name", ENVPOOL_ALL_ENVS)
    @pytest.mark.parametrize("frame_skip", [4, 1])
    @pytest.mark.parametrize("transformed_out", [False, True])
    def test_env_basic_operation(
        self, env_name, frame_skip, transformed_out, T=10, N=3
    ):
        torch.manual_seed(0)
        env_multithreaded = _make_multithreaded_env(
            env_name,
            frame_skip,
            transformed_out=transformed_out,
            N=N,
        )
        td = TensorDict(
            source={"action": env_multithreaded.action_spec.rand()},
            batch_size=[
                N,
            ],
        )
        td1 = env_multithreaded.step(td)
        assert not td1.is_shared()
        assert ("next", "done") in td1.keys(True)
        assert ("next", "reward") in td1.keys(True)

        with pytest.raises(RuntimeError):
            # number of actions does not match number of workers
            td = TensorDict(
                source={"action": env_multithreaded.action_spec.rand()},
                batch_size=[N - 1],
            )
            _ = env_multithreaded.step(td)

        _reset = torch.zeros(N, dtype=torch.bool).bernoulli_()
        td_reset = TensorDict(
            source={"_reset": _reset},
            batch_size=[N],
        )
        env_multithreaded.reset(tensordict=td_reset)

        td = env_multithreaded.rollout(
            policy=None, max_steps=T, break_when_any_done=False
        )
        assert (
            td.shape == torch.Size([N, T]) or td.get("done").sum(1).all()
        ), f"{td.shape}, {td.get('done').sum(1)}"

        env_multithreaded.close()

    # Don't run on Atari envs because output is uint8
    @pytest.mark.skipif(not _has_gym, reason="no gym")
    @pytest.mark.parametrize("env_name", ENVPOOL_CLASSIC_CONTROL_ENVS + ENVPOOL_DM_ENVS)
    @pytest.mark.parametrize("frame_skip", [4, 1])
    @pytest.mark.parametrize("transformed_out", [True, False])
    def test_env_with_policy(
        self,
        env_name,
        frame_skip,
        transformed_out,
        T=10,
        N=3,
    ):
        class DiscreteChoice(torch.nn.Module):
            """Dummy module producing discrete output. Necessary when the action space is discrete."""

            def __init__(self, out_dim: int, dtype: torch.dtype | str | None):
                super().__init__()
                self.lin = torch.nn.LazyLinear(out_dim, dtype=dtype)

            def forward(self, x):
                res = torch.argmax(self.lin(x), axis=-1, keepdim=True)
                return res

        env_multithreaded = _make_multithreaded_env(
            env_name,
            frame_skip,
            transformed_out=transformed_out,
            N=N,
        )
        if env_name == "CheetahRun-v1":
            in_keys = [("velocity")]
            dtype = torch.float64
        else:
            in_keys = ["observation"]
            dtype = torch.float32

        if env_multithreaded.action_spec.shape:
            module = torch.nn.LazyLinear(
                env_multithreaded.action_spec.shape[-1], dtype=dtype
            )
        else:
            # Action space is discrete
            module = DiscreteChoice(env_multithreaded.action_spec.space.n, dtype=dtype)

        policy = ActorCriticOperator(
            SafeModule(
                spec=None,
                module=torch.nn.LazyLinear(12, dtype=dtype),
                in_keys=in_keys,
                out_keys=["hidden"],
            ),
            SafeModule(
                spec=None,
                module=module,
                in_keys=["hidden"],
                out_keys=["action"],
            ),
            ValueOperator(
                module=MLP(out_features=1, num_cells=[], layer_kwargs={"dtype": dtype}),
                in_keys=["hidden", "action"],
            ),
        )

        td = TensorDict(
            source={"action": env_multithreaded.action_spec.rand()},
            batch_size=[
                N,
            ],
        )

        td1 = env_multithreaded.step(td)
        assert not td1.is_shared()
        assert ("next", "done") in td1.keys(True)
        assert ("next", "reward") in td1.keys(True)

        with pytest.raises(RuntimeError):
            # number of actions does not match number of workers
            td = TensorDict(
                source={"action": env_multithreaded.action_spec.rand()},
                batch_size=[N - 1],
            )
            _ = env_multithreaded.step(td)

        reset = torch.zeros(N, dtype=torch.bool).bernoulli_()
        td_reset = TensorDict(
            source={"_reset": reset},
            batch_size=[N],
        )
        env_multithreaded.reset(tensordict=td_reset)
        td = env_multithreaded.rollout(
            policy=policy, max_steps=T, break_when_any_done=False
        )
        assert (
            td.shape == torch.Size([N, T]) or td.get("done").sum(1).all()
        ), f"{td.shape}, {td.get('done').sum(1)}"

        env_multithreaded.close()

    @pytest.mark.skipif(not _has_gym, reason="no gym")
    @pytest.mark.parametrize("env_name", ENVPOOL_ALL_ENVS)
    @pytest.mark.parametrize("frame_skip", [4, 1])
    @pytest.mark.parametrize("transformed_out", [True, False])
    def test_multithreaded_env_seed(
        self, env_name, frame_skip, transformed_out, seed=100, N=4
    ):
        # Create the first env, set the seed, and perform a sequence of operations
        env = _make_multithreaded_env(
            env_name,
            frame_skip,
            transformed_out=True,
            N=N,
        )
        action = env.action_spec.rand()
        env.set_seed(seed)
        td0a = env.reset()
        td1a = env.step(td0a.clone().set("action", action))
        td2a = env.rollout(max_steps=10)

        # Create a new env, set the seed, and repeat same operations
        env = _make_multithreaded_env(
            env_name,
            frame_skip,
            transformed_out=True,
            N=N,
        )
        env.set_seed(seed)
        td0b = env.reset()
        td1b = env.step(td0b.clone().set("action", action))
        td2b = env.rollout(max_steps=10)

        # Check that results on two envs are identical
        assert_allclose_td(td0a, td0b.select(*td0a.keys()))
        assert_allclose_td(td1a, td1b)
        assert_allclose_td(td2a, td2b)

        # Check that results are different if seed is different
        # Skip Pong, since there different actions can lead to the same result
        if env_name != PONG_VERSIONED():
            env.set_seed(
                seed=seed + 10,
            )
            td0c = env.reset()
            td1c = env.step(td0c.clone().set("action", action))
            with pytest.raises(AssertionError):
                assert_allclose_td(td0a, td0c.select(*td0a.keys()))
            with pytest.raises(AssertionError):
                assert_allclose_td(td1a, td1c)
        env.close()

    @pytest.mark.skipif(not _has_gym, reason="no gym")
    def test_multithread_env_shutdown(self):
        env = _make_multithreaded_env(
            PENDULUM_VERSIONED(),
            1,
            transformed_out=False,
            N=3,
        )
        env.reset()
        assert not env.is_closed
        env.rand_step()
        assert not env.is_closed
        env.close()
        assert env.is_closed
        env.reset()
        assert not env.is_closed
        env.close()

    @pytest.mark.skipif(not torch.cuda.device_count(), reason="no cuda to test on")
    @pytest.mark.skipif(not _has_gym, reason="no gym")
    @pytest.mark.parametrize("frame_skip", [4])
    @pytest.mark.parametrize("device", [0])
    @pytest.mark.parametrize("env_name", ENVPOOL_ALL_ENVS)
    @pytest.mark.parametrize("transformed_out", [False, True])
    @pytest.mark.parametrize("open_before", [False, True])
    def test_multithreaded_env_cast(
        self,
        env_name,
        frame_skip,
        transformed_out,
        device,
        open_before,
        T=10,
        N=3,
    ):
        # tests casting to device
        env_multithread = _make_multithreaded_env(
            env_name,
            frame_skip,
            transformed_out=transformed_out,
            N=N,
        )
        if open_before:
            td_cpu = env_multithread.rollout(max_steps=10)
            assert td_cpu.device == torch.device("cpu")
        env_multithread = env_multithread.to(device)
        assert env_multithread.observation_spec.device == torch.device(device)
        assert env_multithread.action_spec.device == torch.device(device)
        assert env_multithread.reward_spec.device == torch.device(device)
        assert env_multithread.device == torch.device(device)
        td_device = env_multithread.reset()
        assert td_device.device == torch.device(device), env_multithread
        td_device = env_multithread.rand_step()
        assert td_device.device == torch.device(device), env_multithread
        td_device = env_multithread.rollout(max_steps=10)
        assert td_device.device == torch.device(device), env_multithread
        env_multithread.close()

    @pytest.mark.skipif(not _has_gym, reason="no gym")
    @pytest.mark.skipif(not torch.cuda.device_count(), reason="no cuda device detected")
    @pytest.mark.parametrize("frame_skip", [4])
    @pytest.mark.parametrize("device", [0])
    @pytest.mark.parametrize("env_name", ENVPOOL_ALL_ENVS)
    @pytest.mark.parametrize("transformed_out", [True, False])
    def test_env_device(self, env_name, frame_skip, transformed_out, device):
        # tests creation on device
        torch.manual_seed(0)
        N = 3

        env_multithreaded = _make_multithreaded_env(
            env_name,
            frame_skip,
            transformed_out=transformed_out,
            device=device,
            N=N,
        )

        assert env_multithreaded.device == torch.device(device)
        out = env_multithreaded.rollout(max_steps=20)
        assert out.device == torch.device(device)

        env_multithreaded.close()


@pytest.mark.skipif(not _has_brax, reason="brax not installed")
@pytest.mark.parametrize("device", get_available_devices())
@pytest.mark.parametrize("envname", ["fast"])
class TestBrax:
    @pytest.mark.parametrize("requires_grad", [False, True])
    def test_brax_constructor(self, envname, requires_grad, device):
        env0 = BraxEnv(envname, requires_grad=requires_grad, device=device)
        env1 = BraxWrapper(env0._env, requires_grad=requires_grad, device=device)

        env0.set_seed(0)
        torch.manual_seed(0)
        init = env0.reset()
        if requires_grad:
            init = init.apply(
                lambda x: x.requires_grad_(True) if x.is_floating_point() else x
            )
        r0 = env0.rollout(10, tensordict=init, auto_reset=False)
        assert r0.requires_grad == requires_grad

        env1.set_seed(0)
        torch.manual_seed(0)
        init = env1.reset()
        if requires_grad:
            init = init.apply(
                lambda x: x.requires_grad_(True) if x.is_floating_point() else x
            )
        r1 = env1.rollout(10, tensordict=init, auto_reset=False)
        assert r1.requires_grad == requires_grad
        assert_allclose_td(r0.data, r1.data)

    def test_brax_seeding(self, envname, device):
        final_seed = []
        tdreset = []
        tdrollout = []
        for _ in range(2):
            env = BraxEnv(envname, device=device)
            torch.manual_seed(0)
            np.random.seed(0)
            final_seed.append(env.set_seed(0))
            tdreset.append(env.reset())
            tdrollout.append(env.rollout(max_steps=50))
            env.close()
            del env
        assert final_seed[0] == final_seed[1]
        assert_allclose_td(*tdreset)
        assert_allclose_td(*tdrollout)

    @pytest.mark.parametrize("batch_size", [(), (5,), (5, 4)])
    def test_brax_batch_size(self, envname, batch_size, device):
        env = BraxEnv(envname, batch_size=batch_size, device=device)
        env.set_seed(0)
        tdreset = env.reset()
        tdrollout = env.rollout(max_steps=50)
        env.close()
        del env
        assert tdreset.batch_size == batch_size
        assert tdrollout.batch_size[:-1] == batch_size

    @pytest.mark.parametrize("batch_size", [(), (5,), (5, 4)])
    def test_brax_spec_rollout(self, envname, batch_size, device):
        env = BraxEnv(envname, batch_size=batch_size, device=device)
        env.set_seed(0)
        check_env_specs(env)

    @pytest.mark.parametrize("batch_size", [(), (5,), (5, 4)])
    @pytest.mark.parametrize(
        "requires_grad",
        [
            True,
            False,
        ],
    )
    def test_brax_consistency(self, envname, batch_size, requires_grad, device):
        import jax
        import jax.numpy as jnp
        from torchrl.envs.libs.jax_utils import (
            _ndarray_to_tensor,
            _tensor_to_ndarray,
            _tree_flatten,
        )

        env = BraxEnv(
            envname, batch_size=batch_size, requires_grad=requires_grad, device=device
        )
        env.set_seed(1)
        rollout = env.rollout(10)

        env.set_seed(1)
        key = env._key
        base_env = env._env
        key, *keys = jax.random.split(key, int(np.prod(batch_size) + 1))
        state = jax.vmap(base_env.reset)(jnp.stack(keys))
        for i in range(rollout.shape[-1]):
            action = rollout[..., i]["action"]
            action = _tensor_to_ndarray(action.clone())
            action = _tree_flatten(action, env.batch_size)
            state = jax.vmap(base_env.step)(state, action)
            t1 = rollout[..., i][("next", "observation")]
            t2 = _ndarray_to_tensor(state.obs).view_as(t1)
            torch.testing.assert_close(t1, t2)

    @pytest.mark.parametrize("batch_size", [(), (5,), (5, 4)])
    def test_brax_grad(self, envname, batch_size, device):
        batch_size = (1,)
        env = BraxEnv(envname, batch_size=batch_size, requires_grad=True, device=device)
        env.set_seed(0)
        td1 = env.reset()
        action = torch.randn(env.action_spec.shape)
        action.requires_grad_(True)
        td1["action"] = action
        td2 = env.step(td1)
        td2[("next", "reward")].mean().backward()
        env.close()
        del env

    @pytest.mark.parametrize("batch_size", [(), (5,), (5, 4)])
    @pytest.mark.parametrize("parallel", [False, True])
    def test_brax_parallel(
        self, envname, batch_size, parallel, maybe_fork_ParallelEnv, device, n=1
    ):
        def make_brax():
            env = BraxEnv(
                envname, batch_size=batch_size, requires_grad=False, device=device
            )
            env.set_seed(1)
            return env

        if parallel:
            env = maybe_fork_ParallelEnv(n, make_brax)
        else:
            env = SerialEnv(n, make_brax)
        check_env_specs(env)
        tensordict = env.rollout(3)
        assert tensordict.shape == torch.Size([n, *batch_size, 3])


@pytest.mark.skipif(not _has_vmas, reason="vmas not installed")
class TestVmas:
    @pytest.mark.parametrize("scenario_name", VmasWrapper.available_envs)
    @pytest.mark.parametrize("continuous_actions", [True, False])
    def test_all_vmas_scenarios(self, scenario_name, continuous_actions):
        env = VmasEnv(
            scenario=scenario_name,
            continuous_actions=continuous_actions,
            num_envs=4,
        )
        env.set_seed(0)
        env.reset()
        env.rollout(10)
        env.close()

    @pytest.mark.parametrize(
        "scenario_name", ["simple_reference", "waterfall", "flocking", "discovery"]
    )
    def test_vmas_seeding(self, scenario_name):
        final_seed = []
        tdreset = []
        tdrollout = []
        for _ in range(2):
            env = VmasEnv(
                scenario=scenario_name,
                num_envs=4,
            )
            final_seed.append(env.set_seed(0))
            tdreset.append(env.reset())
            tdrollout.append(env.rollout(max_steps=10))
            env.close()
            del env
        assert final_seed[0] == final_seed[1]
        assert_allclose_td(*tdreset)
        assert_allclose_td(*tdrollout)

    @pytest.mark.parametrize(
        "batch_size", [(), (12,), (12, 2), (12, 3), (12, 3, 1), (12, 3, 4)]
    )
    @pytest.mark.parametrize("scenario_name", VmasWrapper.available_envs)
    def test_vmas_batch_size_error(self, scenario_name, batch_size):
        num_envs = 12
        if len(batch_size) > 1:
            with pytest.raises(
                TypeError,
                match="Batch size used in constructor is not compatible with vmas.",
            ):
                _ = VmasEnv(
                    scenario=scenario_name,
                    num_envs=num_envs,
                    batch_size=batch_size,
                )
        elif len(batch_size) == 1 and batch_size != (num_envs,):
            with pytest.raises(
                TypeError,
                match="Batch size used in constructor does not match vmas batch size.",
            ):
                _ = VmasEnv(
                    scenario=scenario_name,
                    num_envs=num_envs,
                    batch_size=batch_size,
                )
        else:
            env = VmasEnv(
                scenario=scenario_name,
                num_envs=num_envs,
                batch_size=batch_size,
            )
            env.close()

    @pytest.mark.parametrize("num_envs", [1, 20])
    @pytest.mark.parametrize("n_agents", [1, 5])
    @pytest.mark.parametrize(
        "scenario_name",
        ["simple_reference", "simple_tag", "waterfall", "flocking", "discovery"],
    )
    def test_vmas_batch_size(self, scenario_name, num_envs, n_agents):
        torch.manual_seed(0)
        n_rollout_samples = 5
        env = VmasEnv(
            scenario=scenario_name,
            num_envs=num_envs,
            n_agents=n_agents,
            group_map=MarlGroupMapType.ALL_IN_ONE_GROUP,
        )
        env.set_seed(0)
        tdreset = env.reset()
        tdrollout = env.rollout(
            max_steps=n_rollout_samples,
            return_contiguous=False if env.het_specs else True,
        )
        assert (
            env.full_action_spec_unbatched.shape == env.unbatched_action_spec.shape
        ), (
            env.action_spec,
            env.batch_size,
        )

        env.close()

        if env.het_specs:
            assert isinstance(tdreset["agents"], LazyStackedTensorDict)
        else:
            assert isinstance(tdreset["agents"], TensorDict)

        assert tdreset.batch_size == (num_envs,)
        assert tdreset["agents"].batch_size == (num_envs, env.n_agents)
        if not env.het_specs:
            assert tdreset["agents", "observation"].shape[1] == env.n_agents
        assert tdreset["done"].shape[1] == 1

        assert tdrollout.batch_size == (num_envs, n_rollout_samples)
        assert tdrollout["agents"].batch_size == (
            num_envs,
            n_rollout_samples,
            env.n_agents,
        )
        if not env.het_specs:
            assert tdrollout["agents", "observation"].shape[2] == env.n_agents
        assert tdrollout["next", "agents", "reward"].shape[2] == env.n_agents
        assert tdrollout["agents", "action"].shape[2] == env.n_agents
        assert tdrollout["done"].shape[2] == 1
        del env

    @pytest.mark.parametrize("num_envs", [1, 20])
    @pytest.mark.parametrize("n_agents", [1, 5])
    @pytest.mark.parametrize("continuous_actions", [True, False])
    @pytest.mark.parametrize(
        "scenario_name",
        ["simple_reference", "simple_tag", "waterfall", "flocking", "discovery"],
    )
    def test_vmas_spec_rollout(
        self, scenario_name, num_envs, n_agents, continuous_actions
    ):
        import vmas

        vmas_env = VmasEnv(
            scenario=scenario_name,
            num_envs=num_envs,
            n_agents=n_agents,
            continuous_actions=continuous_actions,
        )
        vmas_wrapped_env = VmasWrapper(
            vmas.make_env(
                scenario=scenario_name,
                num_envs=num_envs,
                n_agents=n_agents,
                continuous_actions=continuous_actions,
            )
        )
        for env in [vmas_env, vmas_wrapped_env]:
            env.set_seed(0)
            check_env_specs(env, return_contiguous=False if env.het_specs else True)
            env.close()

    @pytest.mark.parametrize("num_envs", [1, 20])
    @pytest.mark.parametrize("scenario_name", VmasWrapper.available_envs)
    def test_vmas_repr(self, scenario_name, num_envs):
        env = VmasEnv(
            scenario=scenario_name,
            num_envs=num_envs,
        )
        assert str(env) == (
            f"{VmasEnv.__name__}(num_envs={num_envs}, n_agents={env.n_agents},"
            f" batch_size={torch.Size((num_envs,))}, device={env.device}) (scenario={scenario_name})"
        )
        env.close()

    @pytest.mark.parametrize("num_envs", [1, 10])
    @pytest.mark.parametrize("n_workers", [1, 3])
    @pytest.mark.parametrize("continuous_actions", [True, False])
    @pytest.mark.parametrize(
        "scenario_name", ["simple_reference", "waterfall", "flocking", "discovery"]
    )
    def test_vmas_parallel(
        self,
        scenario_name,
        num_envs,
        n_workers,
        continuous_actions,
        maybe_fork_ParallelEnv,
        n_agents=5,
        n_rollout_samples=3,
    ):
        torch.manual_seed(0)

        def make_vmas():
            env = VmasEnv(
                scenario=scenario_name,
                num_envs=num_envs,
                n_agents=n_agents,
                continuous_actions=continuous_actions,
            )
            env.set_seed(0)
            return env

        env = maybe_fork_ParallelEnv(n_workers, make_vmas)
        tensordict = env.rollout(max_steps=n_rollout_samples)

        assert tensordict.shape == torch.Size(
            [n_workers, list(env.num_envs)[0], n_rollout_samples]
        )
        env.close()

    @pytest.mark.parametrize("num_envs", [1, 2])
    @pytest.mark.parametrize("n_workers", [1, 3])
    @pytest.mark.parametrize(
        "scenario_name", ["simple_reference", "waterfall", "flocking", "discovery"]
    )
    def test_vmas_reset(
        self,
        scenario_name,
        num_envs,
        n_workers,
        maybe_fork_ParallelEnv,
        n_agents=5,
        n_rollout_samples=3,
        max_steps=3,
    ):
        torch.manual_seed(0)

        def make_vmas():
            env = VmasEnv(
                scenario=scenario_name,
                num_envs=num_envs,
                n_agents=n_agents,
                max_steps=max_steps,
            )
            env.set_seed(0)
            return env

        env = maybe_fork_ParallelEnv(n_workers, make_vmas)
        tensordict = env.rollout(max_steps=n_rollout_samples)

        assert (
            tensordict["next", "done"]
            .sum(
                tuple(range(tensordict.batch_dims, tensordict["next", "done"].ndim)),
                dtype=torch.bool,
            )[..., -1]
            .all()
        )

        td_reset = TensorDict(
            rand_reset(env), batch_size=env.batch_size, device=env.device
        )
        # it is good practice to have a "complete" input tensordict for reset
        for done_key in env.done_keys:
            td_reset.set(done_key, tensordict[..., -1].get(("next", done_key)))
        reset = td_reset["_reset"]
        tensordict = env.reset(td_reset)

        assert not tensordict.get("done")[reset].any()
        assert tensordict.get("done")[~reset].all()
        env.close()

    @pytest.mark.skipif(len(get_available_devices()) < 2, reason="not enough devices")
    @pytest.mark.parametrize("first", [0, 1])
    @pytest.mark.parametrize(
        "scenario_name", ["simple_reference", "waterfall", "flocking", "discovery"]
    )
    def test_to_device(self, scenario_name: str, first: int):
        torch.manual_seed(0)
        devices = get_available_devices()

        def make_vmas():
            env = VmasEnv(
                scenario=scenario_name,
                num_envs=7,
                n_agents=3,
                seed=0,
                device=devices[first],
            )
            return env

        env = make_vmas()

        assert env.rollout(max_steps=3).device == devices[first]

        env.to(devices[1 - first])

        assert env.rollout(max_steps=3).device == devices[1 - first]
        env.close()

    @pytest.mark.parametrize("n_envs", [1, 4])
    @pytest.mark.parametrize("n_workers", [1, 2])
    @pytest.mark.parametrize("n_agents", [1, 3])
    def test_collector(
        self, n_envs, n_workers, n_agents, maybe_fork_ParallelEnv, frames_per_batch=80
    ):
        torch.manual_seed(1)
        env_fun = lambda: VmasEnv(
            scenario="flocking", num_envs=n_envs, n_agents=n_agents, max_steps=7
        )

        env = maybe_fork_ParallelEnv(n_workers, env_fun)

        n_actions_per_agent = env.action_spec.shape[-1]
        n_observations_per_agent = env.observation_spec["agents", "observation"].shape[
            -1
        ]

        policy = SafeModule(
            nn.Linear(
                n_observations_per_agent,
                n_actions_per_agent,
            ),
            in_keys=[("agents", "observation")],
            out_keys=[env.action_key],
            spec=env.action_spec,
            safe=True,
        )
        ccollector = SyncDataCollector(
            create_env_fn=env,
            policy=policy,
            frames_per_batch=frames_per_batch,
            total_frames=1000,
            device="cpu",
        )

        for i, _td in enumerate(ccollector):
            if i == 1:
                break
        ccollector.shutdown()

        td_batch = (n_workers, n_envs, frames_per_batch // (n_workers * n_envs))
        agents_td_batch = td_batch + (n_agents,)

        assert _td.shape == td_batch
        assert _td["next"].shape == td_batch
        assert _td["agents"].shape == agents_td_batch
        assert _td["agents", "info"].shape == agents_td_batch
        assert _td["next", "agents"].shape == agents_td_batch
        assert _td["next", "agents", "info"].shape == agents_td_batch
        assert _td["collector"].shape == td_batch

        assert _td[env.action_key].shape == agents_td_batch + (n_actions_per_agent,)
        assert _td["agents", "observation"].shape == agents_td_batch + (
            n_observations_per_agent,
        )
        assert _td["next", "agents", "observation"].shape == agents_td_batch + (
            n_observations_per_agent,
        )
        assert _td["next", env.reward_key].shape == agents_td_batch + (1,)
        for done_key in env.done_keys:
            assert _td[done_key].shape == td_batch + (1,)
            assert _td["next", done_key].shape == td_batch + (1,)

        assert env.reward_key not in _td.keys(True, True)
        assert env.action_key not in _td["next"].keys(True, True)

    def test_collector_heterogeneous(self, n_envs=10, frames_per_batch=20):
        env = VmasEnv(
            scenario="simple_tag",
            num_envs=n_envs,
            group_map=MarlGroupMapType.ALL_IN_ONE_GROUP,
        )
        torch.manual_seed(1)

        ccollector = SyncDataCollector(
            create_env_fn=env,
            policy=None,
            frames_per_batch=frames_per_batch,
            total_frames=1000,
            device="cpu",
        )

        for i, _td in enumerate(ccollector):
            if i == 1:
                break
        ccollector.shutdown()

        td_batch = (n_envs, frames_per_batch // n_envs)
        agents_td_batch = td_batch + (env.n_agents,)

        assert _td.shape == td_batch
        assert _td["next"].shape == td_batch
        assert _td["agents"].shape == agents_td_batch
        assert _td["next", "agents"].shape == agents_td_batch
        assert _td["collector"].shape == td_batch
        assert _td["next", env.reward_key].shape == agents_td_batch + (1,)
        for done_key in env.done_keys:
            assert _td[done_key].shape == td_batch + (1,)
            assert _td["next", done_key].shape == td_batch + (1,)

        assert env.reward_key not in _td.keys(True, True)
        assert env.action_key not in _td["next"].keys(True, True)

    @pytest.mark.parametrize("n_agents", [1, 5])
    def test_grouping(self, n_agents, scenario_name="dispersion", n_envs=2):
        env = VmasEnv(
            scenario=scenario_name,
            num_envs=n_envs,
            n_agents=n_agents,
        )
        env = VmasEnv(
            scenario=scenario_name,
            num_envs=n_envs,
            n_agents=n_agents,
            # Put each agent in a group with its name
            group_map={
                agent_name: [agent_name] for agent_name in reversed(env.agent_names)
            },
        )

        # Check that when setting the action for a specific group, it is reflected to the right agent in the backend
        for group in env.group_map.keys():
            env.reset()
            action = env.full_action_spec.zero()
            action.set((group, "action"), action.get((group, "action")) + 1.0)
            prev_pos = {agent.name: agent.state.pos.clone() for agent in env.agents}
            env.step(action)
            pos = {agent.name: agent.state.pos.clone() for agent in env.agents}
            for agent_name in env.agent_names:
                if agent_name == group:
                    assert (pos[agent_name] > prev_pos[agent_name]).all()
                else:
                    assert (pos[agent_name] == prev_pos[agent_name]).all()


@pytest.mark.slow
class TestGenDGRL:
    @staticmethod
    @pytest.fixture
    def _patch_traj_len():
        # avoids processing the entire dataset
        _get_category_len = GenDGRLExperienceReplay._get_category_len

        def new_get_category_len(cls, category_name):
            return 100

        GenDGRLExperienceReplay._get_category_len = classmethod(new_get_category_len)

        yield
        GenDGRLExperienceReplay._get_category_len = _get_category_len

    @pytest.mark.parametrize("dataset_num", [4])
    def test_gen_dgrl_preproc(self, dataset_num, tmpdir, _patch_traj_len):
        dataset_id = GenDGRLExperienceReplay.available_datasets[dataset_num]
        tmpdir = Path(tmpdir)
        dataset = GenDGRLExperienceReplay(
            dataset_id, batch_size=32, root=tmpdir / "1", download="force"
        )
        from torchrl.envs import Compose, GrayScale, Resize

        t = Compose(
            Resize(32, in_keys=["observation", ("next", "observation")]),
            GrayScale(in_keys=["observation", ("next", "observation")]),
        )

        def fn(data):
            return t(data)

        new_storage = dataset.preprocess(
            fn,
            num_workers=max(1, os.cpu_count() - 2),
            num_chunks=1000,
            num_frames=100,
            dest=tmpdir / "2",
        )
        dataset = ReplayBuffer(storage=new_storage, batch_size=32)
        assert len(dataset) == 100
        sample = dataset.sample()
        assert sample["observation"].shape == torch.Size([32, 1, 32, 32])
        assert sample["next", "observation"].shape == torch.Size([32, 1, 32, 32])

    @pytest.mark.parametrize("dataset_num", [0, 4, 8])
    def test_gen_dgrl(self, dataset_num, tmpdir, _patch_traj_len):
        dataset_id = GenDGRLExperienceReplay.available_datasets[dataset_num]
        dataset = GenDGRLExperienceReplay(dataset_id, batch_size=32, root=tmpdir)
        for batch in dataset:  # noqa: B007
            break
        assert batch.get(("next", "observation")).shape[-3] == 3
        for key in (
            ("next", "done"),
            ("next", "truncated"),
            ("next", "terminated"),
            "observation",
            "action",
            ("next", "reward"),
        ):
            assert key in batch.keys(True, True)
        for key in (
            ("next", "done"),
            ("next", "truncated"),
            ("next", "terminated"),
            "terminated",
            "truncated",
            "done",
            ("next", "reward"),
        ):
            val = batch.get(key)
            assert val.shape[:-1] == batch.shape


@pytest.mark.skipif(not _has_d4rl, reason="D4RL not found")
@pytest.mark.slow
class TestD4RL:
    def test_d4rl_preproc(self, tmpdir):
        dataset_id = "walker2d-medium-replay-v2"
        tmpdir = Path(tmpdir)
        dataset = D4RLExperienceReplay(
            dataset_id,
            batch_size=32,
            root=tmpdir / "1",
            download="force",
            direct_download=True,
        )
        from torchrl.envs import CatTensors, Compose

        t = Compose(
            CatTensors(
                in_keys=["observation", ("info", "qpos"), ("info", "qvel")],
                out_key="data",
            ),
            CatTensors(
                in_keys=[
                    ("next", "observation"),
                    ("next", "info", "qpos"),
                    ("next", "info", "qvel"),
                ],
                out_key=("next", "data"),
            ),
        )

        def fn(data):
            return t(data)

        new_storage = dataset.preprocess(
            fn,
            num_workers=max(1, os.cpu_count() - 2),
            num_chunks=1000,
            mp_start_method="fork",
            dest=tmpdir / "2",
            num_frames=100,
        )
        dataset = ReplayBuffer(storage=new_storage, batch_size=32)
        assert len(dataset) == 100
        sample = dataset.sample()
        assert sample["data"].shape == torch.Size([32, 35])
        assert sample["next", "data"].shape == torch.Size([32, 35])

    @pytest.mark.parametrize("task", ["walker2d-medium-replay-v2"])
    @pytest.mark.parametrize("use_truncated_as_done", [True, False])
    @pytest.mark.parametrize("split_trajs", [True, False])
    def test_terminate_on_end(self, task, use_truncated_as_done, split_trajs, tmpdir):
        root1 = tmpdir / "1"
        root2 = tmpdir / "2"
        root3 = tmpdir / "3"

        with pytest.warns(
            UserWarning, match="Using use_truncated_as_done=True"
        ) if use_truncated_as_done else nullcontext():
            data_true = D4RLExperienceReplay(
                task,
                split_trajs=split_trajs,
                from_env=False,
                terminate_on_end=True,
                batch_size=2,
                use_truncated_as_done=use_truncated_as_done,
                download="force",
                root=root1,
            )
        _ = D4RLExperienceReplay(
            task,
            split_trajs=split_trajs,
            from_env=False,
            terminate_on_end=False,
            batch_size=2,
            use_truncated_as_done=use_truncated_as_done,
            download="force",
            root=root2,
        )
        data_from_env = D4RLExperienceReplay(
            task,
            split_trajs=split_trajs,
            from_env=True,
            batch_size=2,
            use_truncated_as_done=use_truncated_as_done,
            download="force",
            root=root3,
        )
        if not use_truncated_as_done:
            keys = set(data_from_env[:].keys(True, True))
            keys = keys.intersection(data_true[:].keys(True, True))
            assert data_true[:].shape == data_from_env[:].shape
            # for some reason, qlearning_dataset overwrites the next obs that is contained in the buffer,
            # resulting in tiny changes in the value contained for that key. Over 99.99% of the values
            # match, but the test still fails because of this.
            # We exclude that entry from the comparison.
            keys.discard(("next", "observation"))
            assert_allclose_td(
                data_true[:].select(*keys),
                data_from_env[:].select(*keys),
            )
        else:
            leaf_names = data_from_env[:].keys(True)
            leaf_names = [
                name[-1] if isinstance(name, tuple) else name for name in leaf_names
            ]
            assert "truncated" in leaf_names
            leaf_names = data_true[:].keys(True)
            leaf_names = [
                name[-1] if isinstance(name, tuple) else name for name in leaf_names
            ]
            assert "truncated" not in leaf_names

    @pytest.mark.parametrize("task", ["walker2d-medium-replay-v2"])
    def test_direct_download(self, task, tmpdir):
        root1 = tmpdir / "1"
        root2 = tmpdir / "2"
        data_direct = D4RLExperienceReplay(
            task,
            split_trajs=False,
            from_env=False,
            batch_size=2,
            use_truncated_as_done=True,
            direct_download=True,
            download="force",
            root=root1,
        )
        data_d4rl = D4RLExperienceReplay(
            task,
            split_trajs=False,
            from_env=True,
            batch_size=2,
            use_truncated_as_done=True,
            direct_download=False,
            terminate_on_end=True,  # keep the last time step
            download="force",
            root=root2,
        )
        keys = set(data_direct[:].keys(True, True))
        keys = keys.intersection(data_d4rl[:].keys(True, True))
        assert len(keys)
        assert_allclose_td(
            data_direct[:].select(*keys).apply(lambda t: t.float()),
            data_d4rl[:].select(*keys).apply(lambda t: t.float()),
        )

    @pytest.mark.parametrize(
        "task",
        [
            # "antmaze-medium-play-v0",
            # "hammer-cloned-v1",
            # "maze2d-open-v0",
            # "maze2d-open-dense-v0",
            # "relocate-human-v1",
            "walker2d-medium-replay-v2",
            # "ant-medium-v2",
            # # "flow-merge-random-v0",
            # "kitchen-partial-v0",
            # # "carla-town-v0",
        ],
    )
    def test_d4rl_dummy(self, task):
        t0 = time.time()
        _ = D4RLExperienceReplay(task, split_trajs=True, from_env=True, batch_size=2)
        torchrl_logger.info(f"terminated test after {time.time()-t0}s")

    @pytest.mark.parametrize("task", ["walker2d-medium-replay-v2"])
    @pytest.mark.parametrize("split_trajs", [True, False])
    @pytest.mark.parametrize("from_env", [True, False])
    def test_dataset_build(self, task, split_trajs, from_env):
        t0 = time.time()
        data = D4RLExperienceReplay(
            task, split_trajs=split_trajs, from_env=from_env, batch_size=2
        )
        sample = data.sample()
        env = GymWrapper(gym.make(task))
        rollout = env.rollout(2)
        for key in rollout.keys(True, True):
            if "truncated" in key:
                # truncated is missing from static datasets
                continue
            sim = rollout.get(key)
            offline = sample.get(key)
            # assert sim.dtype == offline.dtype, key
            assert sim.shape[-1] == offline.shape[-1], key
        torchrl_logger.info(f"terminated test after {time.time()-t0}s")

    @pytest.mark.parametrize("task", ["walker2d-medium-replay-v2"])
    @pytest.mark.parametrize("split_trajs", [True, False])
    def test_d4rl_iteration(self, task, split_trajs):
        t0 = time.time()
        batch_size = 3
        data = D4RLExperienceReplay(
            task,
            split_trajs=split_trajs,
            from_env=False,
            terminate_on_end=True,
            batch_size=batch_size,
            sampler=SamplerWithoutReplacement(drop_last=True),
        )
        i = 0
        for sample in data:  # noqa: B007
            i += 1
        assert len(data) // i == batch_size
        torchrl_logger.info(f"terminated test after {time.time()-t0}s")


_MINARI_DATASETS = []


def _minari_selected_datasets():
    if not _has_minari or not _has_gymnasium:
        return
    global _MINARI_DATASETS
    import minari

    torch.manual_seed(0)

    total_keys = sorted(
        minari.list_remote_datasets(latest_version=True, compatible_minari_version=True)
    )
    indices = torch.randperm(len(total_keys))[:20]
    keys = [total_keys[idx] for idx in indices]

    assert len(keys) > 5, keys
    _MINARI_DATASETS += keys


_minari_selected_datasets()


@pytest.mark.skipif(not _has_minari or not _has_gymnasium, reason="Minari not found")
@pytest.mark.slow
class TestMinari:
    @pytest.mark.parametrize("split", [False, True])
    @pytest.mark.parametrize("selected_dataset", _MINARI_DATASETS)
    def test_load(self, selected_dataset, split):
        torchrl_logger.info(f"dataset {selected_dataset}")
        data = MinariExperienceReplay(
            selected_dataset, batch_size=32, split_trajs=split
        )
        t0 = time.time()
        for i, sample in enumerate(data):
            t1 = time.time()
            torchrl_logger.info(f"sampling time {1000 * (t1-t0): 4.4f}ms")
            assert data.metadata["action_space"].is_in(sample["action"])
            assert data.metadata["observation_space"].is_in(sample["observation"])
            t0 = time.time()
            if i == 10:
                break

    def test_minari_preproc(self, tmpdir):
        dataset = MinariExperienceReplay(
            "D4RL/pointmaze/large-v2",
            batch_size=32,
            split_trajs=False,
            download="force",
        )

        from torchrl.envs import CatTensors, Compose

        t = Compose(
            CatTensors(
                in_keys=[
                    ("observation", "observation"),
                    ("info", "qpos"),
                    ("info", "qvel"),
                ],
                out_key="data",
            ),
            CatTensors(
                in_keys=[
                    ("next", "observation", "observation"),
                    ("next", "info", "qpos"),
                    ("next", "info", "qvel"),
                ],
                out_key=("next", "data"),
            ),
        )

        def fn(data):
            return t(data)

        new_storage = dataset.preprocess(
            fn,
            num_workers=max(1, os.cpu_count() - 2),
            num_chunks=1000,
            mp_start_method="fork",
            num_frames=100,
            dest=tmpdir,
        )
        dataset = ReplayBuffer(storage=new_storage, batch_size=32)
        sample = dataset.sample()
        assert len(dataset) == 100
        assert sample["data"].shape == torch.Size([32, 8])
        assert sample["next", "data"].shape == torch.Size([32, 8])


@pytest.mark.slow
class TestRoboset:
    def test_load(self):
        selected_dataset = RobosetExperienceReplay.available_datasets[0]
        data = RobosetExperienceReplay(
            selected_dataset,
            batch_size=32,
        )
        t0 = time.time()
        for i, _ in enumerate(data):
            t1 = time.time()
            torchrl_logger.info(f"sampling time {1000 * (t1-t0): 4.4f}ms")
            t0 = time.time()
            if i == 10:
                break

    def test_roboset_preproc(self, tmpdir):
        dataset = RobosetExperienceReplay(
            "FK1-v4(expert)/FK1_MicroOpenRandom_v2d-v4", batch_size=32, download="force"
        )

        def func(data):
            return data.set("obs_norm", data.get("observation").norm(dim=-1))

        new_storage = dataset.preprocess(
            func,
            num_workers=max(1, os.cpu_count() - 2),
            num_chunks=1000,
            mp_start_method="fork",
            dest=tmpdir,
            num_frames=100,
        )
        dataset = ReplayBuffer(storage=new_storage, batch_size=32)
        assert len(dataset) == 100
        sample = dataset.sample()
        assert "obs_norm" in sample.keys()


@pytest.mark.slow
class TestVD4RL:
    @pytest.mark.parametrize("image_size", [None, (37, 33)])
    def test_load(self, image_size):
        torch.manual_seed(0)
        datasets = VD4RLExperienceReplay.available_datasets
        for idx in torch.randperm(len(datasets)).tolist()[:4]:
            selected_dataset = datasets[idx]
            data = VD4RLExperienceReplay(
                selected_dataset,
                batch_size=32,
                image_size=image_size,
            )
            t0 = time.time()
            for i, batch in enumerate(data):
                if image_size:
                    assert batch.get("pixels").shape == (32, 3, *image_size)
                    assert batch.get(("next", "pixels")).shape == (32, 3, *image_size)
                else:
                    assert batch.get("pixels").shape[:2] == (32, 3)
                    assert batch.get(("next", "pixels")).shape[:2] == (32, 3)

                assert batch.get("pixels").dtype is torch.float32
                assert batch.get(("next", "pixels")).dtype is torch.float32
                assert (batch.get("pixels") != 0).any()
                assert (batch.get(("next", "pixels")) != 0).any()
                t1 = time.time()
                torchrl_logger.info(f"sampling time {1000 * (t1-t0): 4.4f}ms")
                t0 = time.time()
                if i == 10:
                    break

    def test_vd4rl_preproc(self, tmpdir):
        torch.manual_seed(0)
        datasets = VD4RLExperienceReplay.available_datasets
        dataset_id = list(datasets)[4]
        dataset = VD4RLExperienceReplay(dataset_id, batch_size=32, download="force")
        from torchrl.envs import Compose, GrayScale, ToTensorImage

        func = Compose(
            ToTensorImage(in_keys=["pixels", ("next", "pixels")]),
            GrayScale(in_keys=["pixels", ("next", "pixels")]),
        )
        new_storage = dataset.preprocess(
            func,
            num_workers=max(1, os.cpu_count() - 2),
            num_chunks=1000,
            mp_start_method="fork",
            dest=tmpdir,
            num_frames=100,
        )
        dataset = ReplayBuffer(storage=new_storage, batch_size=32)
        assert len(dataset) == 100
        sample = dataset.sample()
        assert sample["next", "pixels"].shape == torch.Size([32, 1, 64, 64])


@pytest.mark.slow
class TestAtariDQN:
    @pytest.fixture(scope="class")
    def limit_max_runs(self):
        prev_val = AtariDQNExperienceReplay._max_runs
        AtariDQNExperienceReplay._max_runs = 3
        yield
        AtariDQNExperienceReplay._max_runs = prev_val

    @pytest.mark.parametrize("dataset_id", ["Asterix/1", "Pong/4"])
    @pytest.mark.parametrize(
        "num_slices,slice_len", [[None, None], [None, 8], [2, None]]
    )
    def test_single_dataset(self, dataset_id, slice_len, num_slices, limit_max_runs):
        dataset = AtariDQNExperienceReplay(
            dataset_id, slice_len=slice_len, num_slices=num_slices
        )
        sample = dataset.sample(64)
        for key in (
            ("next", "observation"),
            ("next", "truncated"),
            ("next", "terminated"),
            ("next", "done"),
            ("next", "reward"),
            "observation",
            "action",
            "done",
            "truncated",
            "terminated",
        ):
            assert key in sample.keys(True)
        assert sample.shape == (64,)
        assert sample.get_non_tensor("metadata")["dataset_id"] == dataset_id

    @pytest.mark.parametrize(
        "num_slices,slice_len", [[None, None], [None, 8], [2, None]]
    )
    def test_double_dataset(self, slice_len, num_slices, limit_max_runs):
        dataset_pong = AtariDQNExperienceReplay(
            "Pong/4", slice_len=slice_len, num_slices=num_slices
        )
        dataset_asterix = AtariDQNExperienceReplay(
            "Asterix/1", slice_len=slice_len, num_slices=num_slices
        )
        dataset = ReplayBufferEnsemble(
            dataset_pong, dataset_asterix, sample_from_all=True, batch_size=128
        )
        sample = dataset.sample()
        assert sample.shape == (2, 64)
        assert sample[0].get_non_tensor("metadata")["dataset_id"] == "Pong/4"
        assert sample[1].get_non_tensor("metadata")["dataset_id"] == "Asterix/1"

    @pytest.mark.parametrize("dataset_id", ["Pong/4"])
    def test_atari_preproc(self, dataset_id, tmpdir):
        from torchrl.envs import Compose, RenameTransform, Resize, UnsqueezeTransform

        dataset = AtariDQNExperienceReplay(
            dataset_id,
            slice_len=None,
            num_slices=8,
            batch_size=64,
            # num_procs=max(0, os.cpu_count() - 4),
            num_procs=0,
        )

        t = Compose(
            UnsqueezeTransform(
                dim=-3, in_keys=["observation", ("next", "observation")]
            ),
            Resize(32, in_keys=["observation", ("next", "observation")]),
            RenameTransform(in_keys=["action"], out_keys=["other_action"]),
        )

        def preproc(data):
            return t(data)

        new_storage = dataset.preprocess(
            preproc,
            num_workers=max(1, os.cpu_count() - 4),
            num_chunks=1000,
            # mp_start_method="fork",
            pbar=True,
            dest=tmpdir,
            num_frames=100,
        )

        dataset = ReplayBuffer(storage=new_storage, batch_size=32)
        assert len(dataset) == 100


@pytest.mark.slow
class TestOpenX:
    @pytest.mark.parametrize(
        "download,padding",
        [[True, None], [False, None], [False, 0], [False, True], [False, False]],
    )
    @pytest.mark.parametrize("shuffle", [True, False])
    @pytest.mark.parametrize("replacement", [True, False])
    @pytest.mark.parametrize(
        "batch_size,num_slices,slice_len",
        [
            [3000, 2, None],
            [32, 32, None],
            [32, None, 1],
            [3000, None, 1500],
            [None, None, 32],
            [None, None, 1500],
        ],
    )
    def test_openx(
        self, download, shuffle, replacement, padding, batch_size, num_slices, slice_len
    ):
        torch.manual_seed(0)
        np.random.seed(0)

        streaming = not download
        cm = (
            pytest.raises(RuntimeError, match="shuffle=False")
            if not streaming and not shuffle and replacement
            else pytest.raises(
                RuntimeError,
                match="replacement=True is not available with streamed datasets",
            )
            if streaming and replacement
            else nullcontext()
        )
        dataset = None
        with cm:
            dataset = OpenXExperienceReplay(
                "cmu_stretch",
                download=download,
                streaming=streaming,
                batch_size=batch_size,
                shuffle=shuffle,
                num_slices=num_slices,
                slice_len=slice_len,
                pad=padding,
                replacement=replacement,
            )
        if dataset is None:
            return
        # iterating
        if padding is None and (
            (batch_size is not None and batch_size > 1000)
            or (slice_len is not None and slice_len > 1000)
        ):
            raises_cm = pytest.raises(
                RuntimeError,
                match="The trajectory length (.*) is shorter than the slice length|"
                #       "Some stored trajectories have a length shorter than the slice that was asked for|"
                "Did not find a single trajectory with sufficient length",
            )
            with raises_cm:
                for data in dataset:  # noqa: B007
                    break
            if batch_size is None and slice_len is not None:
                with raises_cm:
                    dataset.sample(2 * slice_len)
                return

        else:
            for data in dataset:  # noqa: B007
                break
            # check data shape
            if batch_size is not None:
                assert data.shape[0] == batch_size
            elif slice_len is not None:
                assert data.shape[0] == slice_len
            if batch_size is not None:
                if num_slices is not None:
                    assert data.get(("next", "done")).sum(-2) == num_slices
                elif streaming:
                    assert (
                        data.get(("next", "done")).sum(-2)
                        == data.get("episode").unique().numel()
                    )

        # sampling
        if batch_size is None:
            if slice_len is not None:
                batch_size = 2 * slice_len
            elif num_slices is not None:
                batch_size = num_slices * 32
            sample = dataset.sample(batch_size)
        else:
            if padding is None and (batch_size > 1000):
                with pytest.raises(
                    RuntimeError,
                    match="Did not find a single trajectory with sufficient length"
                    if not streaming
                    else "The trajectory length (.*) is shorter than the slice length",
                ):
                    sample = dataset.sample()
                return
            else:
                sample = dataset.sample()
                assert sample.shape == (batch_size,)
        if slice_len is not None:
            assert sample.get(("next", "done")).sum() == int(
                batch_size // slice_len
            ), sample.get(("next", "done"))
        elif num_slices is not None:
            assert sample.get(("next", "done")).sum() == num_slices

    def test_openx_preproc(self, tmpdir):
        dataset = OpenXExperienceReplay(
            "cmu_stretch",
            download=True,
            streaming=False,
            batch_size=64,
            shuffle=True,
            num_slices=8,
            slice_len=None,
        )
        from torchrl.envs import Compose, RenameTransform, Resize

        t = Compose(
            Resize(
                64,
                64,
                in_keys=[("observation", "image"), ("next", "observation", "image")],
            ),
            RenameTransform(
                in_keys=[
                    ("observation", "image"),
                    ("next", "observation", "image"),
                    ("observation", "state"),
                    ("next", "observation", "state"),
                ],
                out_keys=["pixels", ("next", "pixels"), "state", ("next", "state")],
            ),
        )

        def fn(data: TensorDict):
            data.unlock_()
            data = data.select(
                "action",
                "done",
                "episode",
                ("next", "done"),
                ("next", "observation"),
                ("next", "reward"),
                ("next", "terminated"),
                ("next", "truncated"),
                "observation",
                "terminated",
                "truncated",
            )
            data = t(data)
            data = data.select(*data.keys(True, True))
            return data

        new_storage = dataset.preprocess(
            CloudpickleWrapper(fn),
            num_workers=max(1, os.cpu_count() - 2),
            num_chunks=500,
            # mp_start_method="fork",
            dest=tmpdir,
        )
        dataset = ReplayBuffer(storage=new_storage)
        sample = dataset.sample(32)
        assert "observation" not in sample.keys()
        assert "pixels" in sample.keys()
        assert ("next", "pixels") in sample.keys(True)
        assert "state" in sample.keys()
        assert ("next", "state") in sample.keys(True)
        assert sample["pixels"].shape == torch.Size([32, 3, 64, 64])


@pytest.mark.skipif(not _has_sklearn, reason="Scikit-learn not found")
@pytest.mark.parametrize(
    "dataset",
    [
        # "adult_num", # 1226: Expensive to test
        # "adult_onehot", # 1226: Expensive to test
        "mushroom_num",
        "mushroom_onehot",
        # "covertype",  # 1226: Expensive to test
        "shuttle",
        "magic",
    ],
)
@pytest.mark.slow
class TestOpenML:
    @pytest.mark.parametrize("batch_size", [(), (2,), (2, 3)])
    def test_env(self, dataset, batch_size):
        env = OpenMLEnv(dataset, batch_size=batch_size)
        td = env.reset()
        assert td.shape == torch.Size(batch_size)
        td = env.rand_step(td)
        assert td.shape == torch.Size(batch_size)
        assert "index" not in td.keys()
        check_env_specs(env)

    def test_data(self, dataset):
        data = OpenMLExperienceReplay(
            dataset,
            batch_size=2048,
            transform=Compose(
                RenameTransform(["X"], ["observation"]),
                DoubleToFloat(["observation"]),
            ),
        )
        # check that dataset eventually runs out
        for i, _ in enumerate(data):  # noqa: B007
            continue
        assert len(data) // 2048 in (i, i - 1)


@pytest.mark.skipif(not _has_isaac, reason="IsaacGym not found")
@pytest.mark.parametrize(
    "task",
    [
        "AllegroHand",
        # "AllegroKuka",
        # "AllegroKukaTwoArms",
        # "AllegroHandManualDR",
        # "AllegroHandADR",
        "Ant",
        # "Anymal",
        # "AnymalTerrain",
        # "BallBalance",
        # "Cartpole",
        # "FactoryTaskGears",
        # "FactoryTaskInsertion",
        # "FactoryTaskNutBoltPick",
        # "FactoryTaskNutBoltPlace",
        # "FactoryTaskNutBoltScrew",
        # "FrankaCabinet",
        # "FrankaCubeStack",
        "Humanoid",
        # "HumanoidAMP",
        # "Ingenuity",
        # "Quadcopter",
        # "ShadowHand",
        "Trifinger",
    ],
)
@pytest.mark.parametrize("num_envs", [10, 20])
@pytest.mark.parametrize("device", get_default_devices())
@pytest.mark.parametrize("from_pixels", [False])
class TestIsaacGym:
    @classmethod
    def _run_on_proc(cls, q, task, num_envs, device, from_pixels):
        try:
            env = IsaacGymEnv(
                task=task, num_envs=num_envs, device=device, from_pixels=from_pixels
            )
            check_env_specs(env)
            q.put(("succeeded!", None))
        except Exception as err:
            q.put(("failed!", err))
            raise err

    def test_env(self, task, num_envs, device, from_pixels):
        from torch import multiprocessing as mp

        q = mp.Queue(1)
        self._run_on_proc(q, task, num_envs, device, from_pixels)
        proc = mp.Process(
            target=self._run_on_proc, args=(q, task, num_envs, device, from_pixels)
        )
        try:
            proc.start()
            msg, error = q.get()
            if msg != "succeeded!":
                raise error
        finally:
            q.close()
            proc.join()

    #
    # def test_collector(self, task, num_envs, device):
    #     env = IsaacGymEnv(task=task, num_envs=num_envs, device=device)
    #     collector = SyncDataCollector(
    #         env,
    #         policy=SafeModule(nn.LazyLinear(out_features=env.observation_spec['obs'].shape[-1]), in_keys=["obs"], out_keys=["action"]),
    #         frames_per_batch=20,
    #         total_frames=-1
    #     )
    #     for c in collector:
    #         assert c.shape == torch.Size([num_envs, 20])
    #         break


@pytest.mark.skipif(not _has_pettingzoo, reason="PettingZoo not found")
class TestPettingZoo:
    @pytest.mark.parametrize("parallel", [True, False])
    @pytest.mark.parametrize("continuous_actions", [True, False])
    @pytest.mark.parametrize("use_mask", [True])
    @pytest.mark.parametrize("return_state", [True, False])
    @pytest.mark.parametrize(
        "group_map",
        [None, MarlGroupMapType.ALL_IN_ONE_GROUP, MarlGroupMapType.ONE_GROUP_PER_AGENT],
    )
    def test_pistonball(
        self, parallel, continuous_actions, use_mask, return_state, group_map
    ):
        kwargs = {"n_pistons": 21, "continuous": continuous_actions}

        env = PettingZooEnv(
            task="pistonball_v6",
            parallel=parallel,
            seed=0,
            return_state=return_state,
            use_mask=use_mask,
            group_map=group_map,
            **kwargs,
        )

        check_env_specs(env)

    def test_dead_agents_done(self, seed=0):
        scenario_args = {"n_walkers": 3, "terminate_on_fall": False}

        env = PettingZooEnv(
            task="multiwalker_v9",
            parallel=True,
            seed=seed,
            use_mask=False,
            done_on_any=False,
            **scenario_args,
        )
        td_reset = env.reset(seed=seed)
        with pytest.raises(
            ValueError,
            match="Dead agents found in the environment, "
            "you need to set use_mask=True to allow this.",
        ):
            env.rollout(
                max_steps=500,
                break_when_any_done=True,  # This looks at root done set with done_on_any
                auto_reset=False,
                tensordict=td_reset,
            )

        for done_on_any in [True, False]:
            env = PettingZooEnv(
                task="multiwalker_v9",
                parallel=True,
                seed=seed,
                use_mask=True,
                done_on_any=done_on_any,
                **scenario_args,
            )
            td_reset = env.reset(seed=seed)
            td = env.rollout(
                max_steps=500,
                break_when_any_done=True,  # This looks at root done set with done_on_any
                auto_reset=False,
                tensordict=td_reset,
            )
            done = td.get(("next", "walker", "done"))
            mask = td.get(("next", "walker", "mask"))

            if done_on_any:
                assert not done[-1].all()  # Done triggered on any
            else:
                assert done[-1].all()  # Done triggered on all
            assert not done[
                mask
            ].any()  # When mask is true (alive agent), all agents are not done
            assert done[
                ~mask
            ].all()  # When mask is false (dead agent), all agents are done

    @pytest.mark.parametrize(
        "wins_player_0",
        [True, False],
    )
    def test_tic_tac_toe(self, wins_player_0):
        env = PettingZooEnv(
            task="tictactoe_v3",
            parallel=False,
            group_map={"player": ["player_1", "player_2"]},
            categorical_actions=False,
            seed=0,
            use_mask=True,
        )

        class Policy:
            action = 0
            t = 0

            def __call__(self, td):
                new_td = env.input_spec["full_action_spec"].zero()

                player_acting = 0 if self.t % 2 == 0 else 1
                other_player = 1 if self.t % 2 == 0 else 0
                # The acting player has "mask" True and "action_mask" set to the available actions
                assert td["player", "mask"][player_acting].all()
                assert td["player", "action_mask"][player_acting].any()
                # The non-acting player has "mask" False and "action_mask" set to all Trues
                assert not td["player", "mask"][other_player].any()
                assert td["player", "action_mask"][other_player].all()

                if self.t % 2 == 0:
                    if not wins_player_0 and self.t == 4:
                        new_td["player", "action"][0][self.action + 1] = 1
                    else:
                        new_td["player", "action"][0][self.action] = 1
                else:
                    new_td["player", "action"][1][self.action + 6] = 1
                if td["player", "mask"][1].all():
                    self.action += 1
                self.t += 1
                return td.update(new_td)

        td = env.rollout(100, policy=Policy())

        assert td.batch_size[0] == (5 if wins_player_0 else 6)
        assert (td[:-1]["next", "player", "reward"] == 0).all()
        if wins_player_0:
            assert (
                td[-1]["next", "player", "reward"] == torch.tensor([[1], [-1]])
            ).all()
        else:
            assert (
                td[-1]["next", "player", "reward"] == torch.tensor([[-1], [1]])
            ).all()

    @pytest.mark.parametrize(
        "task",
        [
            "multiwalker_v9",
            "waterworld_v4",
            "pursuit_v4",
            "simple_spread_v3",
            "simple_v3",
            "rps_v2",
            "cooperative_pong_v5",
            "pistonball_v6",
        ],
    )
    def test_envs_one_group_parallel(self, task):
        env = PettingZooEnv(
            task=task,
            parallel=True,
            seed=0,
            use_mask=False,
        )
        check_env_specs(env)
        env.rollout(100, break_when_any_done=False)

    @pytest.mark.parametrize(
        "task",
        [
            "multiwalker_v9",
            "waterworld_v4",
            "pursuit_v4",
            "simple_spread_v3",
            "simple_v3",
            "rps_v2",
            "cooperative_pong_v5",
            "pistonball_v6",
            "connect_four_v3",
            "tictactoe_v3",
            "chess_v6",
            "gin_rummy_v4",
            "tictactoe_v3",
        ],
    )
    def test_envs_one_group_aec(self, task):
        env = PettingZooEnv(
            task=task,
            parallel=False,
            seed=0,
            use_mask=True,
        )
        check_env_specs(env)
        env.rollout(100, break_when_any_done=False)

    @pytest.mark.parametrize(
        "task",
        [
            "simple_adversary_v3",
            "simple_crypto_v3",
            "simple_push_v3",
            "simple_reference_v3",
            "simple_speaker_listener_v4",
            "simple_tag_v3",
            "simple_world_comm_v3",
            "knights_archers_zombies_v10",
            "basketball_pong_v3",
            "boxing_v2",
            "foozpong_v3",
        ],
    )
    def test_envs_more_groups_parallel(self, task):
        env = PettingZooEnv(
            task=task,
            parallel=True,
            seed=0,
            use_mask=False,
        )
        check_env_specs(env)
        env.rollout(100, break_when_any_done=False)

    @pytest.mark.parametrize(
        "task",
        [
            "simple_adversary_v3",
            "simple_crypto_v3",
            "simple_push_v3",
            "simple_reference_v3",
            "simple_speaker_listener_v4",
            "simple_tag_v3",
            "simple_world_comm_v3",
            "knights_archers_zombies_v10",
            "basketball_pong_v3",
            "boxing_v2",
            "foozpong_v3",
            "go_v5",
        ],
    )
    def test_envs_more_groups_aec(self, task):
        env = PettingZooEnv(
            task=task,
            parallel=False,
            seed=0,
            use_mask=True,
        )
        check_env_specs(env)
        env.rollout(100, break_when_any_done=False)

    @pytest.mark.parametrize("task", ["knights_archers_zombies_v10", "pistonball_v6"])
    @pytest.mark.parametrize("parallel", [True, False])
    def test_vec_env(self, task, parallel, maybe_fork_ParallelEnv):
        env_fun = lambda: PettingZooEnv(
            task=task,
            parallel=parallel,
            seed=0,
            use_mask=not parallel,
        )
        vec_env = maybe_fork_ParallelEnv(2, create_env_fn=env_fun)
        vec_env.rollout(100, break_when_any_done=False)

    def test_reset_parallel_env(self, maybe_fork_ParallelEnv):
        def base_env_fn():
            return PettingZooEnv(
                task="multiwalker_v9",
                parallel=True,
                seed=0,
                n_walkers=3,
                max_cycles=1000,
            )

        collector = SyncDataCollector(
            lambda: maybe_fork_ParallelEnv(
                num_workers=2,
                create_env_fn=base_env_fn,
                device="cpu",
            ),
            policy=None,
            frames_per_batch=100,
            max_frames_per_traj=50,
            total_frames=200,
            reset_at_each_iter=False,
        )
        for _ in collector:
            pass
        collector.shutdown()

    @pytest.mark.parametrize("task", ["knights_archers_zombies_v10", "pistonball_v6"])
    @pytest.mark.parametrize("parallel", [True, False])
    def test_collector(self, task, parallel):
        env_fun = lambda: PettingZooEnv(
            task=task,
            parallel=parallel,
            seed=0,
            use_mask=not parallel,
        )
        collector = SyncDataCollector(
            create_env_fn=env_fun, frames_per_batch=30, total_frames=60, policy=None
        )
        for _ in collector:
            break


@pytest.mark.skipif(not _has_robohive, reason="RoboHive not found")
class TestRoboHive:
    # unfortunately we must import robohive to get the available envs
    # and this import will occur whenever pytest is run on this file.
    # The other option would be not to use parametrize but that also
    # means less informative error trace stacks.
    # In the CI, robohive should not coexist with other libs so that's fine.
    # Robohive logging behavior can be controlled via ROBOHIVE_VERBOSITY=ALL/INFO/(WARN)/ERROR/ONCE/ALWAYS/SILENT
    @pytest.mark.parametrize("from_pixels", [False, True])
    @pytest.mark.parametrize("from_depths", [False, True])
    @pytest.mark.parametrize("envname", RoboHiveEnv.available_envs)
    def test_robohive(self, envname, from_pixels, from_depths):
        with set_gym_backend("gymnasium"):
            torchrl_logger.info(f"{envname}-{from_pixels}-{from_depths}")
            if any(
                substr in envname for substr in ("_vr3m", "_vrrl", "_vflat", "_vvc1s")
            ):
                torchrl_logger.info("not testing envs with prebuilt rendering")
                return
            if "Adroit" in envname:
                torchrl_logger.info("tcdm are broken")
                return
            if (
                from_pixels
                and len(RoboHiveEnv.get_available_cams(env_name=envname)) == 0
            ):
                torchrl_logger.info("no camera")
                return
            try:
                env = RoboHiveEnv(
                    envname, from_pixels=from_pixels, from_depths=from_depths
                )
            except AttributeError as err:
                if "'MjData' object has no attribute 'get_body_xipos'" in str(err):
                    torchrl_logger.info("tcdm are broken")
                    return
                else:
                    raise err
            # Make sure that the stack is dense
            for val in env.rollout(4).values(True):
                if is_tensor_collection(val):
                    assert not isinstance(val, LazyStackedTensorDict)
                    assert not val.is_empty()
            check_env_specs(env)


@pytest.mark.skipif(not _has_smacv2, reason="SMACv2 not found")
class TestSmacv2:
    def test_env_procedural(self):
        distribution_config = {
            "n_units": 5,
            "n_enemies": 6,
            "team_gen": {
                "dist_type": "weighted_teams",
                "unit_types": ["marine", "marauder", "medivac"],
                "exception_unit_types": ["medivac"],
                "weights": [0.5, 0.2, 0.3],
                "observe": True,
            },
            "start_positions": {
                "dist_type": "surrounded_and_reflect",
                "p": 0.5,
                "n_enemies": 5,
                "map_x": 32,
                "map_y": 32,
            },
        }
        env = SMACv2Env(
            map_name="10gen_terran",
            capability_config=distribution_config,
            seed=0,
        )
        check_env_specs(env, seed=None)
        env.close()

    @pytest.mark.parametrize("categorical_actions", [True, False])
    @pytest.mark.parametrize("map", ["MMM2", "3s_vs_5z"])
    def test_env(self, map: str, categorical_actions):
        env = SMACv2Env(
            map_name=map,
            categorical_actions=categorical_actions,
            seed=0,
        )
        check_env_specs(env, seed=None)
        env.close()

    def test_parallel_env(self, maybe_fork_ParallelEnv):
        env = TransformedEnv(
            maybe_fork_ParallelEnv(
                num_workers=2,
                create_env_fn=lambda: SMACv2Env(
                    map_name="3s_vs_5z",
                    seed=0,
                ),
            ),
            ActionMask(
                action_key=("agents", "action"), mask_key=("agents", "action_mask")
            ),
        )
        check_env_specs(env, seed=None)
        env.close()

    def test_collector(self):
        env = SMACv2Env(map_name="MMM2", seed=0, categorical_actions=True)
        in_feats = env.observation_spec["agents", "observation"].shape[-1]
        out_feats = env.action_spec.space.n

        module = TensorDictModule(
            nn.Linear(in_feats, out_feats),
            in_keys=[("agents", "observation")],
            out_keys=[("agents", "logits")],
        )
        prob = ProbabilisticTensorDictModule(
            in_keys={"logits": ("agents", "logits"), "mask": ("agents", "action_mask")},
            out_keys=[("agents", "action")],
            distribution_class=MaskedCategorical,
        )
        actor = TensorDictSequential(module, prob)

        collector = SyncDataCollector(
            env, policy=actor, frames_per_batch=20, total_frames=40
        )
        for _ in collector:
            break
        collector.shutdown()


# List of OpenSpiel games to test
# TODO: Some of the games in `OpenSpielWrapper.available_envs` raise errors for
# a few different reasons, mostly because we do not support chance nodes yet. So
# we cannot run tests on all of them yet.
_openspiel_games = [
    # ----------------
    # Sequential games
    # 1-player
    "morpion_solitaire",
    # 2-player
    "amazons",
    "battleship",
    "breakthrough",
    "checkers",
    "chess",
    "cliff_walking",
    "clobber",
    "connect_four",
    "cursor_go",
    "dark_chess",
    "dark_hex",
    "dark_hex_ir",
    "dots_and_boxes",
    "go",
    "havannah",
    "hex",
    "kriegspiel",
    "mancala",
    "nim",
    "nine_mens_morris",
    "othello",
    "oware",
    "pentago",
    "phantom_go",
    "phantom_ttt",
    "phantom_ttt_ir",
    "sheriff",
    "tic_tac_toe",
    "twixt",
    "ultimate_tic_tac_toe",
    "y",
    # --------------
    # Parallel games
    # 2-player
    "blotto",
    "matrix_bos",
    "matrix_brps",
    "matrix_cd",
    "matrix_coordination",
    "matrix_mp",
    "matrix_pd",
    "matrix_rps",
    "matrix_rpsw",
    "matrix_sh",
    "matrix_shapleys_game",
    "oshi_zumo",
    # 3-player
    "matching_pennies_3p",
]


@pytest.mark.skipif(not _has_pyspiel, reason="open_spiel not found")
class TestOpenSpiel:
    @pytest.mark.parametrize("game_string", _openspiel_games)
    @pytest.mark.parametrize("return_state", [False, True])
    @pytest.mark.parametrize("categorical_actions", [False, True])
    def test_all_envs(self, game_string, return_state, categorical_actions):
        env = OpenSpielEnv(
            game_string,
            categorical_actions=categorical_actions,
            return_state=return_state,
        )
        check_env_specs(env)

    @pytest.mark.parametrize("game_string", _openspiel_games)
    @pytest.mark.parametrize("return_state", [False, True])
    @pytest.mark.parametrize("categorical_actions", [False, True])
    def test_wrapper(self, game_string, return_state, categorical_actions):
        import pyspiel

        base_env = pyspiel.load_game(game_string).new_initial_state()
        env_torchrl = OpenSpielWrapper(
            base_env, categorical_actions=categorical_actions, return_state=return_state
        )
        env_torchrl.rollout(max_steps=5)

    @pytest.mark.parametrize("game_string", _openspiel_games)
    @pytest.mark.parametrize("return_state", [False, True])
    @pytest.mark.parametrize("categorical_actions", [False, True])
    def test_reset_state(self, game_string, return_state, categorical_actions):
        env = OpenSpielEnv(
            game_string,
            categorical_actions=categorical_actions,
            return_state=return_state,
        )
        td = env.reset()
        td_init = td.clone()

        # Perform an action
        td = env.step(env.full_action_spec.rand())

        # Save the current td for reset
        td_reset = td["next"].clone()

        # Perform a second action
        td = env.step(env.full_action_spec.rand())

        # Resetting to a specific state can only happen if `return_state` is
        # enabled. Otherwise, it is reset to the initial state.
        if return_state:
            # Check that the state was reset to the specified state
            td = env.reset(td_reset)
            assert (td == td_reset).all()
        else:
            # Check that the state was reset to the initial state
            td = env.reset()
            assert (td == td_init).all()

    def test_chance_not_implemented(self):
        with pytest.raises(
            NotImplementedError,
            match="not yet supported",
        ):
            OpenSpielEnv("bridge")


# NOTE: Each of the registered envs are around 180 MB, so only test a few.
_mlagents_registered_envs = [
    "3DBall",
    "StrikersVsGoalie",
]


@pytest.mark.skipif(not _has_unity_mlagents, reason="mlagents_envs not found")
class TestUnityMLAgents:
    @mock.patch("mlagents_envs.env_utils.launch_executable")
    @mock.patch("mlagents_envs.environment.UnityEnvironment._get_communicator")
    @pytest.mark.parametrize(
        "group_map",
        [None, MarlGroupMapType.ONE_GROUP_PER_AGENT, MarlGroupMapType.ALL_IN_ONE_GROUP],
    )
    def test_env(self, mock_communicator, mock_launcher, group_map):
        from mlagents_envs.mock_communicator import MockCommunicator

        mock_communicator.return_value = MockCommunicator(
            discrete_action=False, visual_inputs=0
        )
        env = UnityMLAgentsEnv(" ", group_map=group_map)
        try:
            check_env_specs(env)
        finally:
            env.close()

    @mock.patch("mlagents_envs.env_utils.launch_executable")
    @mock.patch("mlagents_envs.environment.UnityEnvironment._get_communicator")
    @pytest.mark.parametrize(
        "group_map",
        [None, MarlGroupMapType.ONE_GROUP_PER_AGENT, MarlGroupMapType.ALL_IN_ONE_GROUP],
    )
    def test_wrapper(self, mock_communicator, mock_launcher, group_map):
        from mlagents_envs.environment import UnityEnvironment
        from mlagents_envs.mock_communicator import MockCommunicator

        mock_communicator.return_value = MockCommunicator(
            discrete_action=False, visual_inputs=0
        )
        env = UnityMLAgentsWrapper(UnityEnvironment(" "), group_map=group_map)
        try:
            check_env_specs(env)
        finally:
            env.close()

    @mock.patch("mlagents_envs.env_utils.launch_executable")
    @mock.patch("mlagents_envs.environment.UnityEnvironment._get_communicator")
    @pytest.mark.parametrize(
        "group_map",
        [None, MarlGroupMapType.ONE_GROUP_PER_AGENT, MarlGroupMapType.ALL_IN_ONE_GROUP],
    )
    def test_rollout(self, mock_communicator, mock_launcher, group_map):
        from mlagents_envs.environment import UnityEnvironment
        from mlagents_envs.mock_communicator import MockCommunicator

        mock_communicator.return_value = MockCommunicator(
            discrete_action=False, visual_inputs=0
        )
        env = UnityMLAgentsWrapper(UnityEnvironment(" "), group_map=group_map)
        try:
            env.rollout(
                max_steps=500, break_when_any_done=False, break_when_all_done=False
            )
        finally:
            env.close()

    @pytest.mark.unity_editor
    def test_with_editor(self):
        print("Please press play in the Unity editor")  # noqa: T201
        env = UnityMLAgentsEnv(timeout_wait=30)
        try:
            env.reset()
            check_env_specs(env)

            # Perform a rollout
            td = env.reset()
            env.rollout(
                max_steps=100, break_when_any_done=False, break_when_all_done=False
            )

            # Step manually
            tensordicts = []
            td = env.reset()
            tensordicts.append(td)
            traj_len = 200
            for _ in range(traj_len - 1):
                td = env.step(td.update(env.full_action_spec.rand()))
                tensordicts.append(td)

            traj = torch.stack(tensordicts)
            assert traj.batch_size == torch.Size([traj_len])
        finally:
            env.close()

    @retry(
        (
            urllib.error.HTTPError,
            urllib.error.URLError,
            urllib.error.ContentTooShortError,
        ),
        5,
    )
    @pytest.mark.parametrize("registered_name", _mlagents_registered_envs)
    @pytest.mark.parametrize(
        "group_map",
        [None, MarlGroupMapType.ONE_GROUP_PER_AGENT, MarlGroupMapType.ALL_IN_ONE_GROUP],
    )
    def test_registered_envs(self, registered_name, group_map):
        env = UnityMLAgentsEnv(
            registered_name=registered_name,
            no_graphics=True,
            group_map=group_map,
        )
        try:
            check_env_specs(env)

            # Perform a rollout
            td = env.reset()
            env.rollout(
                max_steps=20, break_when_any_done=False, break_when_all_done=False
            )

            # Step manually
            tensordicts = []
            td = env.reset()
            tensordicts.append(td)
            traj_len = 20
            for _ in range(traj_len - 1):
                td = env.step(td.update(env.full_action_spec.rand()))
                tensordicts.append(td)

            traj = torch.stack(tensordicts)
            assert traj.batch_size == torch.Size([traj_len])
        finally:
            env.close()


@pytest.mark.skipif(not _has_meltingpot, reason="Meltingpot not found")
class TestMeltingpot:
    @pytest.mark.parametrize("substrate", MeltingpotWrapper.available_envs)
    def test_all_envs(self, substrate):
        env = MeltingpotEnv(substrate=substrate)
        check_env_specs(env)

    def test_passing_config(self, substrate="commons_harvest__open"):
        from meltingpot import substrate as mp_substrate

        substrate_config = mp_substrate.get_config(substrate)
        env_torchrl = MeltingpotEnv(substrate_config)
        env_torchrl.rollout(max_steps=5)

    def test_wrapper(self, substrate="commons_harvest__open"):
        from meltingpot import substrate as mp_substrate

        substrate_config = mp_substrate.get_config(substrate)
        mp_env = mp_substrate.build_from_config(
            substrate_config, roles=substrate_config.default_player_roles
        )
        env_torchrl = MeltingpotWrapper(env=mp_env)
        env_torchrl.rollout(max_steps=5)

    @pytest.mark.parametrize("max_steps", [1, 5])
    def test_max_steps(self, max_steps):
        env = MeltingpotEnv(substrate="commons_harvest__open", max_steps=max_steps)
        td = env.rollout(max_steps=100, break_when_any_done=True)
        assert td.batch_size[0] == max_steps

    @pytest.mark.parametrize("categorical_actions", [True, False])
    def test_categorical_actions(self, categorical_actions):
        env = MeltingpotEnv(
            substrate="commons_harvest__open", categorical_actions=categorical_actions
        )
        check_env_specs(env)

    @pytest.mark.parametrize("rollout_steps", [1, 3])
    def test_render(self, rollout_steps):
        env = MeltingpotEnv(substrate="commons_harvest__open")
        td = env.rollout(2)
        rollout_penultimate_image = td[-1].get("RGB")
        rollout_last_image = td[-1].get(("next", "RGB"))
        image_from_env = env.get_rgb_image()
        assert torch.equal(rollout_last_image, image_from_env)
        assert not torch.equal(rollout_penultimate_image, image_from_env)


if __name__ == "__main__":
    args, unknown = argparse.ArgumentParser().parse_known_args()
    pytest.main([__file__, "--capture", "no", "--exitfirst"] + unknown)