def _gen_rewards(self, rewards_pos: List[Tuple[int, int]]):
for i, (x, y) in enumerate(rewards_pos):
g = Goal()
self.grid.set(x, y, g)
g.is_goal = False
if self.reward_position == i % len(rewards_pos):
g.is_goal = True
def _gen_grid(self, width, height):
# Create the grid
self.grid = Grid(width, height)
# Generate the surrounding walls
self.grid.horz_wall(0, 0)
self.grid.horz_wall(0, height - 1)
self.grid.vert_wall(0, 0)
self.grid.vert_wall(width - 1, 0)
room_w = width // 2
room_h = height // 2
# For each row of rooms
for j in range(0, 2):
# For each column
for i in range(0, 2):
xL = i * room_w
yT = j * room_h
xR = xL + room_w
yB = yT + room_h
# Bottom wall and door
if i + 1 < 2:
self.grid.vert_wall(xR, yT, room_h)
# pos = (xR, self._rand_int(yT + 1, yB))
# self.grid.set(*pos, None)
# Bottom wall and door
if j + 1 < 2:
self.grid.horz_wall(xL, yB, room_w)
# pos = (self._rand_int(xL + 1, xR), yB)
# self.grid.set(*pos, None)
for hallway in self.hallways.values():
self.grid.set(*hallway, None)
# Randomize the player start position and orientation
if self._agent_default_pos is not None:
self.agent_pos = self._agent_default_pos
self.grid.set(*self._agent_default_pos, None)
self.agent_dir = self._rand_int(0, 4)
else:
self.place_agent()
if self._goal_default_pos is not None:
goal = Goal()
self.grid.set(*self._goal_default_pos, goal)
goal.init_pos, goal.cur_pos = self._goal_default_pos
else:
self.place_obj(Goal())
self.mission = 'Reach the goal'
def _gen_grid(self, width, height):
# Create an empty grid
self.grid = Grid(width, height)
# Generate the surrounding walls
self.grid.wall_rect(0, 0, width, height)
# Place a goal square in random location
self.goal_pos = (
np.random.randint(1, self.width-1),
np.random.randint(1, self.height-1)
)
while self.goal_pos == (1,1):
self.goal_pos = (
np.random.randint(1, self.width-1),
np.random.randint(1, self.height-1)
)
self.put_obj(
Goal(),
*self.goal_pos
)
# Place the agent
if self.agent_start_pos is not None:
self.agent_pos = self.agent_start_pos
self.agent_dir = self.agent_start_dir
else:
self.place_agent()
self.mission = "get to the green goal square"
def _gen_grid(self, width, height):
assert width % 2 == 1 and height % 2 == 1 # odd size
# Create an empty grid
self.grid = Grid(width, height)
# Generate the surrounding walls
self.grid.wall_rect(0, 0, width, height)
# Place the agent in the top-left corner
self.agent_pos = (1, 1)
self.agent_dir = 0
# Place a goal square in the bottom-right corner
self.put_obj(Goal(), width - 2, height - 2)
# Place obstacles (lava or walls)
v, h = object(), object(
) # singleton `vertical` and `horizontal` objects
# Lava rivers or walls specified by direction and position in grid
rivers = [(v, i) for i in range(2, height - 2, 2)]
rivers += [(h, j) for j in range(2, width - 2, 2)]
self.np_random.shuffle(rivers)
rivers = rivers[:self.num_crossings] # sample random rivers
rivers_v = sorted([pos for direction, pos in rivers if direction is v])
rivers_h = sorted([pos for direction, pos in rivers if direction is h])
obstacle_pos = itt.chain(
itt.product(range(1, width - 1), rivers_h),
itt.product(rivers_v, range(1, height - 1)),
)
for i, j in obstacle_pos:
self.put_obj(self.obstacle_type(), i, j)
# Sample path to goal
path = [h] * len(rivers_v) + [v] * len(rivers_h)
self.np_random.shuffle(path)
# Create openings
limits_v = [0] + rivers_v + [height - 1]
limits_h = [0] + rivers_h + [width - 1]
room_i, room_j = 0, 0
for direction in path:
if direction is h:
i = limits_v[room_i + 1]
j = self.np_random.choice(
range(limits_h[room_j] + 1, limits_h[room_j + 1]))
room_i += 1
elif direction is v:
i = self.np_random.choice(
range(limits_v[room_i] + 1, limits_v[room_i + 1]))
j = limits_h[room_j + 1]
room_j += 1
else:
assert False
self.grid.set(i, j, None)
self.mission = ("avoid the lava and get to the green goal square"
if self.obstacle_type == Lava else
"find the opening and get to the green goal square")
def _gen_grid(self, width, height):
super()._gen_grid(width, height)
# Make sure the two rooms are directly connected by a locked door
door, _ = self.add_door(0, 0, 0, locked=True, color='blue')
# Add a key to unlock the door
self.add_object(0, 0, 'key', door.color)
# self.add_object(1, 0, 'key', door.color)
self.place_agent(0, 0)
self.door = door
# Add balls
self.add_object(0, 0, 'ball', door.color)
self.add_object(1, 0, 'ball', 'red')
# self.add_object(0, 0, 'ball', 'red')
# self.add_object(1, 0, 'ball', 'blue')
# self.mission = "open the door"
# Add boxes
self.add_object(0, 0, 'box', door.color)
self.add_object(1, 0, 'box', 'red')
# self.add_object(0, 0, 'box', 'red')
# self.add_object(1, 0, 'box', 'blue')
# Add lava
self.grid.vert_wall(4, 3, height - 6, Lava)
self.grid.horz_wall(9, 2, height - 6, Lava)
# self.grid.horz_wall(4, width - 3, 3, Lava)
# Place a goal in the bottom-right corner
self.put_obj(Goal(), width - 2, height - 2)
def _gen_grid(self, width, height, val=False, seen=True):
# Create the grid
self.grid = Grid(width, height)
# Generate surrounding walls
self.grid.horz_wall(0, 0)
self.grid.horz_wall(0, height - 1)
self.grid.vert_wall(0, 0)
self.grid.vert_wall(width - 1, 0)
# Even during validation, start state distribution
# should be the same as that during training
if not self.rnd_start:
self._agent_default_pos = (1, self.grid_size - 2)
else:
self._agent_default_pos = None
# Place the agent at the center
if self._agent_default_pos is not None:
self.start_pos = self._agent_default_pos
self.grid.set(*self._agent_default_pos, None)
self.start_dir = self._rand_int(
0, 4) # Agent direction doesn't matter
goal = Goal()
self.grid.set(*self._goal_default_pos, goal)
goal.init_pos = goal.curr_pos = self._goal_default_pos
self.mission = goal.init_pos
def _gen_grid(self, width, height, val=False, seen=True):
assert width >= 10 and height >= 10, "Environment too small to place objects"
# Create the grid
self.grid = Grid(width, height)
# Generate surrounding walls
self.grid.horz_wall(0, 0)
self.grid.horz_wall(0, height - 1)
self.grid.vert_wall(0, 0)
self.grid.vert_wall(width - 1, 0)
np.random.seed(self.grid_seed)
for obj_idx in range(self.num_objects):
while True:
c_x, c_y = np.random.choice(list(range(
2, self.grid_size - 3))), np.random.choice(
list(range(2, self.grid_size - 3)))
#obj_size = np.random.choice(list(range(1, self.obj_size+1)))
obj_size = self.obj_size
if obj_size == 3:
cells = list(
product([c_x - 1, c_x, c_x + 1],
[c_y - 1, c_y, c_y + 1]))
elif obj_size == 2:
cells = list(product([c_x, c_x + 1], [c_y, c_y + 1]))
elif obj_size == 1:
cells = list(product([c_x], [c_y]))
else:
raise ValueError
valid = True
for cell in cells:
cell = self.grid.get(cell[0], cell[1])
if not (cell is None or cell.can_overlap()):
valid = False
break
if valid:
for cell in cells:
self.grid.set(*cell, Wall())
break
# Set the start position and the goal position depending upon where the obstacles are present
goal = Goal()
# [NOTE] : This is a hack, add option to set goal location from arguments.
self.grid.set(*self._goal_default_pos, goal)
goal.init_pos = goal.curr_pos = self._goal_default_pos
self.mission = goal.init_pos
self.start_pos = self._agent_default_pos
def _gen_grid(self, width, height):
# Create empty grid
self.grid = Grid(width, height)
self.grid.wall_rect(0, 0, width, height)
# Agent starts in the center
self.start_pos = (width // 2, height // 2)
self.start_dir = 0
# Goal is anywhere but the center
self.place_obj(Goal())
# Set mission string
self.mission = "GO TO GREEN SQUARE"
def decode(array):
"""
Decode an array grid encoding back into a grid
"""
width, height, channels = array.shape
assert channels == 3
grid = Grid(width, height)
for i in range(width):
for j in range(height):
typeIdx, colorIdx, state = array[i, j]
if typeIdx == OBJECT_TO_IDX['unseen'] or \
typeIdx == OBJECT_TO_IDX['empty']:
continue
objType = IDX_TO_OBJECT[typeIdx]
color = IDX_TO_COLOR[colorIdx]
# State, 0: open, 1: closed, 2: locked
is_open = state == 0
is_locked = state == 2
if objType == 'wall':
v = Wall(color)
elif objType == 'floor':
v = Floor(color)
elif objType == 'ball':
v = Ball(color)
elif objType == 'key':
v = Key(color)
elif objType == 'box':
v = Box(color)
elif objType == 'door':
v = Door(color, is_open, is_locked)
elif objType == 'goal':
v = Goal()
elif objType == 'lava':
v = Lava()
elif objType == 'agent':
v = None
else:
assert False, "unknown obj type in decode '%s'" % objType
grid.set(i, j, v)
return grid
def _make_hallways(self):
""" Creates a sub-room with two hallways and defines one of them
to be the goal for the option within the room """
hallways = {
'topleft->topright': ((9, 4), (3, 9)),
'topleft->botleft': ((3, 9), (9, 4)),
'topright->topleft': ((0, 4), (7, 9)),
'topright->botright': ((7, 9), (0, 4)),
'botleft->topleft': ((3, 0), (9, 5)),
'botleft->botright': ((9, 5), (3, 0)),
'botright->topright': ((7, 0), (0, 5)),
'botright->botleft': ((0, 5), (7, 0)),
}
goal, other_hall = hallways[self.option_name]
self.env.grid.set(*goal, Goal())
self.env.grid.set(*other_hall, Wall())
def _add_reward(self):
"""Add reward in final room."""
i, j = self.lattice.end
xL = i * self.room_w + i
yB = j * self.room_h + j
xR = xL + self.room_w + 1
yT = yB + self.room_h + 1
if self.verbose:
print('Adding reward in room ({},{})'.format(i,j))
print('Choosing tile in square [{},{}] X [{},{}]'.format(xL+2, xR, yB+2, yT))
x = random.choice(range(xL+2, xR-1))
y = random.choice(range(yB+2, yT-1))
self.set(x, y, Goal())
def _gen_grid(self, width: int, height: int) -> None:
"""Generate grid space.
Jobs:
- create grid world
- create wall
- set starting point
- set goal
- set lava
"""
assert width >= 5 and height >= 5
# Current position and direction of the agent
self.agent_pos: Tuple[int, int] = (1, 1) # (0,0) is wall
self.agent_dir: int = 0
# Create an empty grid
self.grid = Grid(width, height)
# Create wall
self.grid.wall_rect(0, 0, width, height)
# Create Goal
for position in self.goal_pos:
goal_with_wall = self.__adjust_pos_consider_walls(position)
self.__set_grid_type(*goal_with_wall, Goal())
# Create Lava
if self.obstacle_pos:
for lava_pos in self.obstacle_pos:
lava_with_wall = self.__adjust_pos_consider_walls(lava_pos)
self.__set_grid_type(*lava_with_wall, self.obstacle_type())
# Settings for reward_grid
for cell in itertools.product(
range(self.valid_height), range(self.valid_width)
):
if cell in self.goal_pos:
self.reward_grid[cell] = self.goal_reward
elif cell in self.obstacle_pos:
self.reward_grid[cell] = self.obstacle_reward
else:
self.reward_grid[cell] = self.default_reward
def _gen_grid(self, width, height):
# Create an empty grid
self.grid = Grid(width, height)
# Generate the surrounding walls
self.grid.wall_rect(0, 0, width, height)
# Place the goals
for _ in range(self.n_goals):
self.place_obj(Goal())
# Place the traps
for _ in range(self.n_traps):
self.place_obj(Lava())
# Place the agent
if self.agent_start_pos is not None:
self.agent_pos = self.agent_start_pos
self.agent_dir = self.agent_start_dir
else:
self.place_agent()
self.mission = "get to the green goal square, avoid the lava"
def _gen_grid(self, width, height, val=False, seen=True):
# Create the grid
self.grid = Grid(width, height)
# Generate surrounding walls
self.grid.horz_wall(0, 0)
self.grid.horz_wall(0, height - 1)
self.grid.vert_wall(0, 0)
self.grid.vert_wall(width - 1, 0)
# Place horizontal walls through the grid
self.grid.horz_wall(0, height // 3)
self.grid.horz_wall(0, (2 * height) // 3)
# Place vertical walls through the grid
self.grid.vert_wall(width // 3, 0)
self.grid.vert_wall((2 * width) // 3, 0)
# Create passages
passage_anchors = [(width // 3, height // 3),
(width // 3, (2 * height) // 3),
((2 * width) // 3, height // 3),
((2 * width) // 3, (2 * height) // 3)]
passage_cells = []
for anchor in passage_anchors:
for delta in range(-1 * self.passage_size, self.passage_size + 1):
passage_cells.append((anchor[0] + delta, anchor[1]))
passage_cells.append((anchor[0], anchor[1] + delta))
for cell in passage_cells:
self.grid.set(*cell, None)
# Even during validation, start state distribution
# should be the same as that during training
if not self.rnd_start:
self._agent_default_pos = ((width - 2) // 2, (height - 2) // 2)
else:
self._agent_default_pos = None
# Place the agent at the center
if self._agent_default_pos is not None:
self.start_pos = self._agent_default_pos
self.grid.set(*self._agent_default_pos, None)
self.start_dir = self._rand_int(
0, 4) # Agent direction doesn't matter
else:
if len(self.start_state_exclude_rooms) == 0:
self.place_agent()
else:
valid_start_pos = []
if seen:
exclude_from = self.start_state_exclude_rooms
else:
exclude_from = [
x for x in range(1, 10)
if x not in self.start_state_exclude_rooms
]
for room in range(1, 10):
if room in exclude_from:
continue
# Ignore that there are walls for now, can handle that with rejection sampling
# Get x coordinates of allowed cells
valid_x = []
if room % 3 == 1:
valid_x = list(range(1, width // 3))
elif room % 3 == 2:
valid_x = list(range(width // 3 + 1, (2 * width) // 3))
else:
valid_x = list(range((2 * width) // 3 + 1, width - 1))
# Get valid y-coordinates of allowed cells
valid_y = []
if (room - 1) // 3 == 0:
valid_y = list(range(1, height // 3))
elif (room - 1) // 3 == 1:
valid_y = list(
range(height // 3 + 1, (2 * height) // 3))
else:
valid_y = list(range((2 * height) // 3 + 1,
height - 1))
room_cells = list(product(valid_x, valid_y))
valid_start_pos += room_cells
# Make sure start position doesn't conflict with other cells
while True:
_start_pos = valid_start_pos[np.random.choice(
len(valid_start_pos))]
row = _start_pos[1]
col = _start_pos[0]
cell = self.grid.get(row, col)
if cell is None or cell.can_overlap():
break
self.start_pos = (col, row)
self.start_dir = self._rand_int(
0, 4) # Agent direction doesn't matter
goal = Goal()
self.grid.set(*self._goal_default_pos, goal)
goal.init_pos = goal.curr_pos = self._goal_default_pos
self.mission = goal.init_pos
def spawn_new_food(self):
empties = [(i, j) for i in range(self.grid.height)
for j in range(self.grid.width)
if self.grid.get(i, j) is None
and self.grid.get(i, j) != tuple(self.agent_pos)]
self.grid.set(*random.choice(empties), Goal())