def _gen_grid(self, width, height):
super()._gen_grid(width=width, height=height)
if self.n_zone > 1:
self.current_zone_num = self.n_reset % self.n_zone
color = IDX_TO_COLOR[self.current_zone_num + 1]
self.change_floor_color(color=color)
if self.use_lava:
directions = self.get_surrounding_agent()
# Turn water into lava, to force the agent to kill himself
for pos in directions:
cell = self.grid.get(*pos)
if cell and cell.type == 'water':
self.grid.set(*pos, Lava())
return
random.shuffle(directions)
# If it doesn't have water, creates it close to the agent
for pos in directions:
cell = self.grid.get(*pos)
assert pos != self.start_pos
if cell is None or cell.type == 'floor':
self.grid.set(*pos, Lava())
return
def step(self, action):
self.step_count += 1
done = False
if action == self.actions.left:
self.agent_dir = (self.agent_dir - 1) % 4
elif action == self.actions.right:
self.agent_dir = (self.agent_dir + 1) % 4
elif action == self.actions.forward:
pass
else:
assert False, "unknown action: %d" % action
fwd_pos = self.agent_pos + self.dir_vec
fwd_cell = self.grid.get(*fwd_pos)
if fwd_cell is None:
self.grid.set(*self.agent_pos, Lava())
self.snake.grow_head(*fwd_pos)
self.grid.set(*self.snake.rm_tail(), None)
self.agent_pos = fwd_pos
reward = -0.001
elif fwd_cell.type == 'goal':
self.grid.set(*self.agent_pos, Lava())
self.snake.grow_head(*fwd_pos)
self.agent_pos = fwd_pos
self.spawn_new_food()
reward = 1.0
elif (fwd_cell.type == 'lava' or fwd_cell.type == 'wall'):
reward = -1.0
done = True
else:
assert False
if self.step_count == 1 and done:
assert False
obs = self.gen_obs()
assert any([
isinstance(self.grid.get(i, j), Goal)
for i in range(self.grid.height) for j in range(self.grid.width)
])
return obs, reward, done, {}
def add_lava(self, i=None, j=None, num_lava=1):
lavas = []
for _ in range(num_lava):
# Add the object to a random room if no room specified
room_i = i
room_j = j
if room_i == None:
room_i = self._rand_int(0, self.num_cols)
if room_j == None:
room_j = self._rand_int(0, self.num_rows)
obj, pos = self.place_in_room(room_i, room_j, Lava())
lavas.append(obj)
return lavas
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, shadeIdx, sizeIdx, 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]
shade = IDX_TO_SHADE[shadeIdx]
size = IDX_TO_SIZE[size]
# 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, shade=shade, size=size)
elif objType == 'key':
v = Key(color, shade=shade, size=size)
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()
else:
assert False, "unknown obj type in decode '%s'" % objType
grid.set(i, j, v)
return grid
def _gen_grid(self, width, height):
self.grid = Grid(width, height)
self.grid.wall_rect(0, 0, width, height)
# self.start_pos = (2, 2)
yl, xl, _ = self.observation_space.spaces["image"].shape
self.start_pos = (random.randint(2, yl - 2), random.randint(2, xl - 2))
self.agent_pos = self.start_pos # TODO: the env holding agent traits is shit!
self.start_dir = random.randint(0, 3)
self.agent_dir = self.start_dir
self.snake = Snake(
[self.start_pos,
tuple(self.start_pos - self.dir_vec)])
[self.grid.set(*pos, Lava()) for pos in self.snake.body]
self.spawn_new_food()
self.mission = None
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"