Action Bonus¶

class minigrid.wrappers.ActionBonus(env)[source]¶

Wrapper which adds an exploration bonus. This is a reward to encourage exploration of less visited (state, action) pairs.

Example

>>> import gymnasium as gym
>>> from minigrid.wrappers import ActionBonus
>>> env = gym.make("MiniGrid-Empty-5x5-v0")
>>> _, _ = env.reset(seed=0)
>>> _, reward, _, _, _ = env.step(1)
>>> print(reward)
0
>>> _, reward, _, _, _ = env.step(1)
>>> print(reward)
0
>>> env_bonus = ActionBonus(env)
>>> _, _ = env_bonus.reset(seed=0)
>>> _, reward, _, _, _ = env_bonus.step(1)
>>> print(reward)
1.0
>>> _, reward, _, _, _ = env_bonus.step(1)
>>> print(reward)
1.0

Dict Observation Space¶

class minigrid.wrappers.DictObservationSpaceWrapper(env, max_words_in_mission=50, word_dict=None)[source]¶

Transforms the observation space (that has a textual component) to a fully numerical observation space, where the textual instructions are replaced by arrays representing the indices of each word in a fixed vocabulary.

This wrapper is not applicable to BabyAI environments, given that these have their own language component.

Example

>>> import gymnasium as gym
>>> import matplotlib.pyplot as plt
>>> from minigrid.wrappers import DictObservationSpaceWrapper
>>> env = gym.make("MiniGrid-LavaCrossingS11N5-v0")
>>> obs, _ = env.reset()
>>> obs['mission']
'avoid the lava and get to the green goal square'
>>> env_obs = DictObservationSpaceWrapper(env)
>>> obs, _ = env_obs.reset()
>>> obs['mission'][:10]
[19, 31, 17, 36, 20, 38, 31, 2, 15, 35]

Direction Obs¶

class minigrid.wrappers.DirectionObsWrapper(env, type='slope')[source]¶

Adds the direction from the agent to the goal as goal_direction.

With type="slope", the value is dy / dx, with signed infinity for vertical directions. With type="angle", the legacy value is atan(dy / dx) in radians, in the range [-pi/2, pi/2]. These modes do not distinguish all quadrants and generally give the same value for opposite directions.

Use type="angle_atan2" for a direction that distinguishes all quadrants: atan2(dy, dx) in radians, in the range [-pi, pi]. Here dx and dy are the goal’s coordinates minus the agent’s coordinates in the grid’s coordinate system, where x increases to the right and y increases downward. All modes return a scalar float64 array and use zero when the agent is at the goal.

Requires a Goal object in the grid after each reset; otherwise reset raises ValueError. If there are multiple goals, the first in row-major order is used.

Example

>>> import gymnasium as gym
>>> import matplotlib.pyplot as plt
>>> from minigrid.wrappers import DirectionObsWrapper
>>> env = gym.make("MiniGrid-LavaCrossingS11N5-v0")
>>> env_obs = DirectionObsWrapper(env, type="slope")
>>> obs, _ = env_obs.reset()
>>> obs['goal_direction'].item()
1.0

FlatObs¶

class minigrid.wrappers.FlatObsWrapper(env, maxStrLen: int = 96)[source]¶

Encode mission strings using a one-hot scheme, and combine these with observed images into one flat array.

This wrapper is not applicable to BabyAI environments, given that these have their own language component.

Example

>>> import gymnasium as gym
>>> import matplotlib.pyplot as plt
>>> from minigrid.wrappers import FlatObsWrapper
>>> env = gym.make("MiniGrid-LavaCrossingS11N5-v0")
>>> env_obs = FlatObsWrapper(env)
>>> obs, _ = env_obs.reset()
>>> obs.shape
(2835,)

Fully Obs¶

class minigrid.wrappers.FullyObsWrapper(env)[source]¶

Fully observable gridworld using a compact grid encoding instead of the agent view.

Example

>>> import gymnasium as gym
>>> import matplotlib.pyplot as plt
>>> from minigrid.wrappers import FullyObsWrapper
>>> env = gym.make("MiniGrid-LavaCrossingS11N5-v0")
>>> obs, _ = env.reset()
>>> obs['image'].shape
(7, 7, 3)
>>> env_obs = FullyObsWrapper(env)
>>> obs, _ = env_obs.reset()
>>> obs['image'].shape
(11, 11, 3)

Image Observation¶

class minigrid.wrappers.ImgObsWrapper(env)[source]¶

Use the image as the only observation output, no language/mission.

Example

>>> import gymnasium as gym
>>> from minigrid.wrappers import ImgObsWrapper
>>> env = gym.make("MiniGrid-Empty-5x5-v0")
>>> obs, _ = env.reset()
>>> obs.keys()
dict_keys(['image', 'direction', 'mission'])
>>> env = ImgObsWrapper(env)
>>> obs, _ = env.reset()
>>> obs.shape
(7, 7, 3)

No Death¶

class minigrid.wrappers.NoDeath(env, no_death_types: tuple[str, ...], death_cost: float = -1.0)[source]¶

Wrapper to prevent death in specific cells (e.g., lava cells). Instead of dying, the agent will receive a negative reward.

Example

>>> import gymnasium as gym
>>> from minigrid.wrappers import NoDeath
>>>
>>> env = gym.make("MiniGrid-LavaCrossingS9N1-v0")
>>> _, _ = env.reset(seed=2)
>>> _, _, _, _, _ = env.step(1)
>>> _, reward, term, *_ = env.step(2)
>>> reward, term
(0, True)
>>>
>>> env = NoDeath(env, no_death_types=("lava",), death_cost=-1.0)
>>> _, _ = env.reset(seed=2)
>>> _, _, _, _, _ = env.step(1)
>>> _, reward, term, *_ = env.step(2)
>>> reward, term
(-1.0, False)
>>>
>>>
>>> env = gym.make("MiniGrid-Dynamic-Obstacles-5x5-v0")
>>> _, _ = env.reset(seed=2)
>>> _, reward, term, *_ = env.step(2)
>>> reward, term
(-1, True)
>>>
>>> env = NoDeath(env, no_death_types=("ball",), death_cost=-1.0)
>>> _, _ = env.reset(seed=2)
>>> _, reward, term, *_ = env.step(2)
>>> reward, term
(-2.0, False)

Observation¶

class minigrid.wrappers.ObservationWrapper(env: Env[ObsType, ActType])[source]¶

Modify observations from Env.reset() and Env.step() using observation() function.

If you would like to apply a function to only the observation before passing it to the learning code, you can simply inherit from ObservationWrapper and overwrite the method observation() to implement that transformation. The transformation defined in that method must be reflected by the env observation space. Otherwise, you need to specify the new observation space of the wrapper by setting self.observation_space in the __init__() method of your wrapper.

One Hot Partial Obs¶

class minigrid.wrappers.OneHotPartialObsWrapper(env, tile_size=8)[source]¶

Wrapper to get a one-hot encoding of a partially observable agent view as observation.

Example

>>> import gymnasium as gym
>>> from minigrid.wrappers import OneHotPartialObsWrapper
>>> env = gym.make("MiniGrid-Empty-5x5-v0")
>>> obs, _ = env.reset()
>>> obs["image"][0, :, :]
array([[2, 5, 0],
       [2, 5, 0],
       [2, 5, 0],
       [2, 5, 0],
       [2, 5, 0],
       [2, 5, 0],
       [2, 5, 0]], dtype=uint8)
>>> env = OneHotPartialObsWrapper(env)
>>> obs, _ = env.reset()
>>> obs["image"][0, :, :]
array([[0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 0, 0],
       [0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 0, 0],
       [0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 0, 0],
       [0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 0, 0],
       [0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 0, 0],
       [0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 0, 0],
       [0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 0, 0]],
      dtype=uint8)

Reseed¶

class minigrid.wrappers.ReseedWrapper(env, seeds=(0,), seed_idx=0)[source]¶

Wrapper to always regenerate an environment with the same set of seeds. This can be used to force an environment to always keep the same configuration when reset.

Example

>>> import minigrid
>>> import gymnasium as gym
>>> from minigrid.wrappers import ReseedWrapper
>>> env = gym.make("MiniGrid-Empty-5x5-v0")
>>> _ = env.reset(seed=123)
>>> [env.np_random.integers(10).item() for i in range(10)]
[0, 6, 5, 0, 9, 2, 2, 1, 3, 1]
>>> env = ReseedWrapper(env, seeds=[0, 1], seed_idx=0)
>>> _, _ = env.reset()
>>> [env.np_random.integers(10).item() for i in range(10)]
[8, 6, 5, 2, 3, 0, 0, 0, 1, 8]
>>> _, _ = env.reset()
>>> [env.np_random.integers(10).item() for i in range(10)]
[4, 5, 7, 9, 0, 1, 8, 9, 2, 3]
>>> _, _ = env.reset()
>>> [env.np_random.integers(10).item() for i in range(10)]
[8, 6, 5, 2, 3, 0, 0, 0, 1, 8]
>>> _, _ = env.reset()
>>> [env.np_random.integers(10).item() for i in range(10)]
[4, 5, 7, 9, 0, 1, 8, 9, 2, 3]

RGB Img Obs¶

class minigrid.wrappers.RGBImgObsWrapper(env, tile_size=8)[source]¶

Wrapper to use fully observable RGB image as observation, This can be used to have the agent to solve the gridworld in pixel space.

Example

>>> import gymnasium as gym
>>> import matplotlib.pyplot as plt
>>> from minigrid.wrappers import RGBImgObsWrapper
>>> env = gym.make("MiniGrid-Empty-5x5-v0")
>>> obs, _ = env.reset()
>>> plt.imshow(obs['image'])  
![NoWrapper](../figures/lavacrossing_NoWrapper.png)
>>> env = RGBImgObsWrapper(env)
>>> obs, _ = env.reset()
>>> plt.imshow(obs['image'])  
![RGBImgObsWrapper](../figures/lavacrossing_RGBImgObsWrapper.png)

RGB Partial Img Obs¶

class minigrid.wrappers.RGBImgPartialObsWrapper(env, tile_size=8)[source]¶

Wrapper to use partially observable RGB image as observation. This can be used to have the agent to solve the gridworld in pixel space.

Example

>>> import gymnasium as gym
>>> import matplotlib.pyplot as plt
>>> from minigrid.wrappers import RGBImgObsWrapper, RGBImgPartialObsWrapper
>>> env = gym.make("MiniGrid-LavaCrossingS11N5-v0")
>>> obs, _ = env.reset()
>>> plt.imshow(obs["image"])  
![NoWrapper](../figures/lavacrossing_NoWrapper.png)
>>> env_obs = RGBImgObsWrapper(env)
>>> obs, _ = env_obs.reset()
>>> plt.imshow(obs["image"])  
![RGBImgObsWrapper](../figures/lavacrossing_RGBImgObsWrapper.png)
>>> env_obs = RGBImgPartialObsWrapper(env)
>>> obs, _ = env_obs.reset()
>>> plt.imshow(obs["image"])  
![RGBImgPartialObsWrapper](../figures/lavacrossing_RGBImgPartialObsWrapper.png)

Position Bonus¶

class minigrid.wrappers.PositionBonus(env, scale=1)[source]¶

Adds a scaled exploration bonus based on which positions are visited on the grid.

Note

This wrapper was previously called StateBonus.

Example

>>> import gymnasium as gym
>>> from minigrid.wrappers import PositionBonus
>>> env = gym.make("MiniGrid-Empty-5x5-v0")
>>> _, _ = env.reset(seed=0)
>>> _, reward, _, _, _ = env.step(1)
>>> print(reward)
0
>>> _, reward, _, _, _ = env.step(1)
>>> print(reward)
0
>>> env_bonus = PositionBonus(env, scale=1)
>>> obs, _ = env_bonus.reset(seed=0)
>>> obs, reward, terminated, truncated, info = env_bonus.step(1)
>>> print(reward)
1.0
>>> obs, reward, terminated, truncated, info = env_bonus.step(1)
>>> print(reward)
0.7071067811865475

Stochastic Action¶

class minigrid.wrappers.StochasticActionWrapper(env=None, prob=0.9, random_action=None)[source]¶

Add stochasticity to the actions

If a random action is provided, it is returned with probability 1 - prob. Else, a random action is sampled from the action space.

Symbolic Obs¶

class minigrid.wrappers.SymbolicObsWrapper(env)[source]¶

Fully observable grid with a symbolic state representation. The symbol is a triple of (X, Y, IDX), where X and Y are the coordinates on the grid, and IDX is the id of the object.

Example

>>> import gymnasium as gym
>>> from minigrid.wrappers import SymbolicObsWrapper
>>> env = gym.make("MiniGrid-LavaCrossingS11N5-v0")
>>> obs, _ = env.reset()
>>> obs['image'].shape
(7, 7, 3)
>>> env_obs = SymbolicObsWrapper(env)
>>> obs, _ = env_obs.reset()
>>> obs['image'].shape
(11, 11, 3)

View Size¶

class minigrid.wrappers.ViewSizeWrapper(env, agent_view_size=7)[source]¶

Wrapper to customize the agent field of view size. This cannot be used with fully observable wrappers.

Example

>>> import gymnasium as gym
>>> from minigrid.wrappers import ViewSizeWrapper
>>> env = gym.make("MiniGrid-LavaCrossingS11N5-v0")
>>> obs, _ = env.reset()
>>> obs['image'].shape
(7, 7, 3)
>>> env_obs = ViewSizeWrapper(env, agent_view_size=5)
>>> obs, _ = env_obs.reset()
>>> obs['image'].shape
(5, 5, 3)