diff --git a/rllib/env/wrappers/__init__.py b/rllib/env/wrappers/__init__.py
new file mode 100644
index 000000000..e69de29bb
diff --git a/rllib/env/wrappers/recsim_wrapper.py b/rllib/env/wrappers/recsim_wrapper.py
new file mode 100644
index 000000000..6446d19fb
--- /dev/null
+++ b/rllib/env/wrappers/recsim_wrapper.py
@@ -0,0 +1,116 @@
+"""Wrap Google's RecSim environment for RLlib
+
+RecSim is a configurable recommender systems simulation platform.
+Source: https://github.com/google-research/recsim
+"""
+
+import math
+from typing import List, OrderedDict
+
+import gym
+from gym import spaces
+from recsim.environments import interest_evolution
+
+from ray.rllib.utils.error import UnsupportedSpaceException
+from ray.tune.registry import register_env
+
+
+class RecSimObservationSpaceWrapper(gym.ObservationWrapper):
+    """Fix RecSim environment's observation space
+
+    In RecSim's observation spaces, the "doc" field is a dictionary keyed by
+    document IDs. Those IDs are changing every step, thus generating a
+    different observation space in each time. This causes issues for RLlib
+    because it expects the observation space to remain the same across steps.
+
+    This environment wrapper fixes that by reindexing the documents by their
+    positions in the list.
+    """
+
+    def __init__(self, env: gym.Env):
+        super().__init__(env)
+        obs_space = self.env.observation_space
+        doc_space = spaces.Dict(
+            OrderedDict(
+                [(str(k), doc)
+                 for k, (_,
+                         doc) in enumerate(obs_space["doc"].spaces.items())]))
+        self.observation_space = spaces.Dict(
+            OrderedDict([
+                ("user", obs_space["user"]),
+                ("doc", doc_space),
+                ("response", obs_space["response"]),
+            ]))
+
+    def observation(self, obs):
+        new_obs = OrderedDict()
+        new_obs["user"] = obs["user"]
+        new_obs["doc"] = {
+            str(k): v
+            for k, (_, v) in enumerate(obs["doc"].items())
+        }
+        new_obs["response"] = obs["response"]
+        return new_obs
+
+
+class RecSimResetWrapper(gym.Wrapper):
+    """Fix RecSim environment's reset() function
+
+    RecSim's reset() function returns an observation without the "response"
+    field, breaking RLlib's check. This wrapper fixes that by assigning a
+    random "response".
+    """
+
+    def reset(self):
+        obs = super().reset()
+        obs["response"] = self.env.observation_space["response"].sample()
+        return obs
+
+
+class MultiDiscreteToDiscreteActionWrapper(gym.ActionWrapper):
+    """Convert the action space from MultiDiscrete to Discrete
+
+    At this moment, RLlib's DQN algorithms only work on Discrete action space.
+    This wrapper allows us to apply DQN algorithms to the RecSim environment.
+    """
+
+    def __init__(self, env: gym.Env):
+        super().__init__(env)
+
+        if not isinstance(env.action_space, spaces.MultiDiscrete):
+            raise UnsupportedSpaceException(
+                f"Action space {env.action_space} "
+                f"is not supported by {self.__class__.__name__}")
+        self.action_space_dimensions = env.action_space.nvec
+        self.action_space = spaces.Discrete(
+            math.prod(self.action_space_dimensions))
+
+    def action(self, action: int) -> List[int]:
+        """Convert a Discrete action to a MultiDiscrete action"""
+        multi_action = [None] * len(self.action_space_dimensions)
+        for idx, n in enumerate(self.action_space_dimensions):
+            action, dim_action = divmod(action, n)
+            multi_action[idx] = dim_action
+        return multi_action
+
+
+def make_recsim_env(config):
+    DEFAULT_ENV_CONFIG = {
+        "num_candidates": 10,
+        "slate_size": 2,
+        "resample_documents": True,
+        "seed": 0,
+        "convert_to_discrete_action_space": False,
+    }
+    env_config = DEFAULT_ENV_CONFIG.copy()
+    env_config.update(config)
+    env = interest_evolution.create_environment(env_config)
+    env = RecSimResetWrapper(env)
+    env = RecSimObservationSpaceWrapper(env)
+    if config and config["convert_to_discrete_action_space"]:
+        env = MultiDiscreteToDiscreteActionWrapper(env)
+    return env
+
+
+env_name = "RecSim-v1"
+register_env(name=env_name, env_creator=make_recsim_env)
diff --git a/rllib/env/wrappers/tests/test_recsim_wrapper.py b/rllib/env/wrappers/tests/test_recsim_wrapper.py
new file mode 100644
index 000000000..3211dca71
--- /dev/null
+++ b/rllib/env/wrappers/tests/test_recsim_wrapper.py
@@ -0,0 +1,38 @@
+import unittest
+
+import gym
+
+from ray.rllib.env.wrappers.recsim_wrapper import (
+    MultiDiscreteToDiscreteActionWrapper, make_recsim_env)
+from ray.rllib.utils.error import UnsupportedSpaceException
+
+
+class TestRecSimWrapper(unittest.TestCase):
+    def test_observation_space(self):
+        env = make_recsim_env(config={})
+        obs = env.reset()
+        self.assertTrue(
+            env.observation_space.contains(obs),
+            f"{env.observation_space} doesn't contain {obs}")
+        new_obs, _, _, _ = env.step(env.action_space.sample())
+        self.assertTrue(env.observation_space.contains(new_obs))
+
+    def test_action_space_conversion(self):
+        env = make_recsim_env(
+            config={"convert_to_discrete_action_space": True})
+        self.assertIsInstance(env.action_space, gym.spaces.Discrete)
+        env.reset()
+        action = env.action_space.sample()
+        env.step(action)
+
+    def test_double_action_space_conversion_raises_exception(self):
+        env = make_recsim_env(
+            config={"convert_to_discrete_action_space": True})
+        with self.assertRaises(UnsupportedSpaceException):
+            env = MultiDiscreteToDiscreteActionWrapper(env)
+
+
+if __name__ == "__main__":
+    import sys
+    import pytest
+    sys.exit(pytest.main(["-v", __file__]))