multigrid.core.agent module

class multigrid.core.agent.Agent

Bases: object

Class representing an agent in the environment.

Observation Space:

Observations are dictionaries with the following entries:

• imagendarray[int] of shape (view_size, view_size, WorldObj.dim)

  Encoding of the agent’s view of the environment

• directionint

  Agent’s direction (0: right, 1: down, 2: left, 3: up)

• missionMission

  Task string corresponding to the current environment configuration

Action Space:

Actions are discrete integers, as enumerated in Action.

Attributes:

indexint

Index of the agent in the environment

stateAgentState

State of the agent

missionMission

Current mission string for the agent

action_spacegym.spaces.Discrete

Action space for the agent

observation_spacegym.spaces.Dict

Observation space for the agent

__init__(index: int, mission_space: MissionSpace = MissionSpace('maximize reward'), view_size: int = 7, see_through_walls: bool = False)

Parameters:

indexint

Index of the agent in the environment

mission_spaceMissionSpace

The mission space for the agent

view_sizeint

The size of the agent’s view (must be odd)

see_through_wallsbool

Whether the agent can see through walls

property color

Alias for AgentState.color.

property dir

Alias for AgentState.dir.

property pos

Alias for AgentState.pos.

property terminated

Alias for AgentState.terminated.

property carrying

Alias for AgentState.carrying.

property front_pos: tuple[int, int]

Get the position of the cell that is directly in front of the agent.

reset(mission: Mission = Mission('maximize reward'))

Reset the agent to an initial state.

Parameters:

missionMission

Mission string to use for the new episode

encode() → tuple[int, int, int]

Encode a description of this agent as a 3-tuple of integers.

Returns:

type_idxint

The index of the agent type

color_idxint

The index of the agent color

agent_dirint

The direction of the agent (0: right, 1: down, 2: left, 3: up)

render(img: ndarray[Any, dtype[uint8]])

Draw the agent.

Parameters:

imgndarray[int] of shape (width, height, 3)

RGB image array to render agent on

class multigrid.core.agent.AgentState

Bases: ndarray

State for an Agent object.

AgentState objects also support vectorized operations, in which case the AgentState object represents the states of multiple agents.

Examples

Create a vectorized agent state for 3 agents:

>>> agent_state = AgentState(3)
>>> agent_state
AgentState(3)

Access and set state attributes for one agent at a time:

>>> a = agent_state[0]
>>> a
AgentState()
>>> a.color
'red'
>>> a.color = 'yellow'

The underlying vectorized state is automatically updated as well:

>>> agent_state.color
array(['yellow', 'green', 'blue'])

Access and set state attributes all at once:

>>> agent_state.dir
array([-1, -1, -1])
>>> agent_state.dir = np.random.randint(4, size=(len(agent_state)))
>>> agent_state.dir
array([2, 3, 0])
>>> a.dir
2

Attributes:

colorColor or ndarray[str]

Return the agent color.

dirDirection or ndarray[int]

Return the agent direction.

postuple[int, int] or ndarray[int]

Return the agent’s (x, y) position.

terminatedbool or ndarray[bool]

Return whether the agent has terminated.

carryingWorldObj or None or ndarray[object]

Return the object the agent is carrying.

TYPE = 0

COLOR = 1

DIR = 2

ENCODING = slice(0, 3, None)

POS = slice(3, 5, None)

TERMINATED = 5

CARRYING = slice(6, 9, None)

dim = 9

static __new__(cls, *dims: int)

Parameters:

dimsint, optional

Shape of vectorized agent state

property color: Color | ndarray[np.str]

Return the agent color.

property dir: Direction | ndarray[np.int]

Return the agent direction.

property pos: tuple[int, int] | ndarray[np.int]

Return the agent’s (x, y) position.

property terminated: bool | ndarray[np.bool]

Return whether the agent has terminated.

property carrying: WorldObj | None | ndarray[np.object]

Return the object the agent is carrying.