Actions

The action primitives an integer action indexes into.

Each function here is a pure State -> State transformation of the world. An Environment holds an ordered action_set (a tuple of these), and transitions_fn applies action_set[a] for an integer action a; the environment's action_space is Discrete(len(action_set)). So the meaning of action 2 depends on which set the environment uses - see MINIGRID_ACTION_SET (the default) and COMPLETE_ACTION_SET.

Conventions shared by all of them:

• "the player" is state.get_player(idx=0) (navix is single-agent for now); "in front" is the cell one step along player.direction (0 east, 1 south, 2 west, 3 north).
• Movement that would enter a wall or a non-walkable entity is a no-op: the player stays put (and a wall/entity-hit event is recorded).
• Every action is total - it always returns a State of the same structure, so the set can go through jax.lax.switch.

COMPLETE_ACTION_SET = (noop, rotate_cw, rotate_ccw, forward, right, backward, left, pickup, open, done) module-attribute

Every action navix defines, in index order: 0 noop, 1 rotate_cw, 2 rotate_ccw, 3 forward, 4 right, 5 backward, 6 left, 7 pickup, 8 open, 9 done. Wider than MiniGrid's set (adds strafing and noop); use it for environments that need lateral movement.

DEFAULT_ACTION_SET = MINIGRID_ACTION_SET module-attribute

The action_set an Environment uses unless overridden - MINIGRID_ACTION_SET.

MINIGRID_ACTION_SET = (rotate_ccw, rotate_cw, forward, pickup, drop, toggle, done) module-attribute

MiniGrid's seven actions, in the same index order MiniGrid uses: 0 rotate_ccw (MiniGrid "left"), 1 rotate_cw ("right"), 2 forward, 3 pickup, 4 drop, 5 toggle, 6 done. See https://github.com/Farama-Foundation/Minigrid/blob/master/minigrid/core/actions.py

backward(state)

Steps the player one cell backwards (opposite to where it faces) without turning. Blocked-move rules as for forward. Not in the MiniGrid action set.

Parameters:

Name	Type	Description	Default
state	State	the current state.	required

Returns:

Name	Type	Description
State	State	the state with the player's position updated (or not, if
	State	blocked).

done(state)

The agent's "I'm finished" signal. Returns the state unchanged - it does not end the episode by itself. Only environments whose termination_fn inspects for it (e.g. GoToObject, via events.on_target_done) react to it; under the default termination it is a no-op.

Parameters:

Name	Type	Description	Default
state	State	the current state.	required

Returns:

Name	Type	Description
State	State	state, unchanged.

drop(state)

Places the item currently in player.pocket on the cell in front of the player and empties the pocket. No-op if the pocket is empty or the cell in front is not walkable (occupied or a wall).

Parameters:

Name	Type	Description	Default
state	State	the current state.	required

Returns:

Name	Type	Description
State	State	the state with the pocketed item back on the grid in front
	State	of the player and player.pocket cleared.

forward(state)

Moves the player one cell in the direction it faces. No-op if that cell is a wall or a non-walkable entity (a hit event is recorded); direction never changes.

Parameters:

Name	Type	Description	Default
state	State	the current state.	required

Returns:

Name	Type	Description
State	State	the state with the player's position updated (or not, if
	State	blocked).

left(state)

Strafes the player one cell to its left (90 degrees counter-clockwise of where it faces) without turning. Blocked-move rules as for forward. Not in the MiniGrid action set.

Parameters:

Name	Type	Description	Default
state	State	the current state.	required

Returns:

Name	Type	Description
State	State	the state with the player's position updated (or not, if
	State	blocked).

noop(state)

Does nothing - the world is returned unchanged. Present so an action set can include an explicit "wait".

Parameters:

Name	Type	Description	Default
state	State	the current state.	required

Returns:

Name	Type	Description
State	State	state, unchanged.

open(state)

Opens whatever openable thing is directly in front of the player:

• a Box -> removed, and its contents (a Key) revealed at that cell (see open_box);
• a Door -> opened iff it is closed and either unlocked (requires == -1) or the player is carrying the required key (player.pocket == door.requires), in which case that key is consumed from the pocket. Opening records a door-opening event.

A Door that is already open, and any other cell, are left untouched. toggle is an alias for this.

Parameters:

Name	Type	Description	Default
state	State	the current state.	required

Returns:

Name	Type	Description
State	State	the state with the door/box in front opened if the
	State	conditions held.

open_box(state, position_in_front)

Opens the Box in front of the player, if any (verified against MiniGrid's actual Box.toggle: env.grid.set(pos, self.contains) - the box is removed and, if it held something, that object takes its place at the same position). Currently only Key boxes are supported (Box.pocket stores the id of a Key instance, the same way player.pocket does) - ObstructedMaze's only real user of this, keys hidden inside boxes.

Parameters:

Name	Type	Description	Default
state	State	The current state.	required
position_in_front	Array	The position directly in front of the player.	required

Returns:

Name	Type	Description
State	State	The new state, with any opened box removed and its
	State	contents (if any) revealed at its former position.

pickup(state)

Picks up the pickable entity (Key, Box, or Ball) directly in front of the player: the entity is moved off the grid and its id is written to player.pocket, overwriting whatever was there. No-op if the cell in front holds nothing pickable. Records a pickup event.

Parameters:

Name	Type	Description	Default
state	State	the current state.	required

Returns:

Name	Type	Description
State	State	the state with the item removed from the grid and its
	State	id in player.pocket.

right(state)

Strafes the player one cell to its right (90 degrees clockwise of where it faces) without turning. Blocked-move rules as for forward. Not in the MiniGrid action set.

Parameters:

Name	Type	Description	Default
state	State	the current state.	required

Returns:

Name	Type	Description
State	State	the state with the player's position updated (or not, if
	State	blocked).

rotate_ccw(state)

Turns the player 90 degrees counter-clockwise (direction -> direction - 1 mod 4). Position is unchanged.

Parameters:

Name	Type	Description	Default
state	State	the current state.	required

Returns:

Name	Type	Description
State	State	the state with the player's direction updated.

rotate_cw(state)

Turns the player 90 degrees clockwise (direction -> direction + 1 mod 4). Position is unchanged.

Parameters:

Name	Type	Description	Default
state	State	the current state.	required

Returns:

Name	Type	Description
State	State	the state with the player's direction updated.

toggle(state)

MiniGrid's toggle action. An alias for open: in navix a door, once opened, stays open (there is no close), so "toggle" and "open" are the same operation.

Parameters:

Name	Type	Description	Default
state	State	the current state.	required

Returns:

Name	Type	Description
State	State	see open.