class SingleTMaze(MiniGridEnv):
is_double = False
reward_values = dict(goal=1, fake_goal=0.1)
view_size: int = None
def __init__(self, corridor_length=3, reward_position=0, max_steps=None, is_double=False, view_size=None,
max_corridor_length=None):
if max_corridor_length is None:
max_corridor_length = corridor_length
self.max_corridor_length = max_corridor_length
self.view_size = view_size if view_size is not None else 7
self.is_double = is_double
self.reward_position = reward_position
self.corridor_length = corridor_length
assert corridor_length > 0
if max_steps is None:
max_steps = 4 + 4 * corridor_length
super().__init__(
grid_size=3 + 2 * self.max_corridor_length,
max_steps=max_steps,
see_through_walls=True, # True for maximum performance
agent_view_size=self.view_size,
)
self.reward_range = (min(self.reward_values["fake_goal"], 0), self.reward_values["goal"])
@property
def mission(self):
goals = ["UPPER LEFT", "UPPER RIGHT", "LOWER RIGHT", "LOWER LEFT"]
return f'Goal is {goals[self.reward_position]}'
def _gen_grid(self, width, height):
# Create an empty grid
self.grid = Grid(width, height)
# Place the agent in the top-left corner
self.start_pos = (int(width / 2), int(height / 2))
self.start_dir = 3
# Create walls
for x in range(0, width):
for y in range(0, height):
self.grid.set(x, y, Wall())
# Create paths
if self.is_double:
for y in range(height // 2 - self.corridor_length, height // 2 + self.corridor_length + 1):
self.grid.set(width // 2, y, None)
for x in range(width // 2 - self.corridor_length, width // 2 + self.corridor_length + 1):
self.grid.set(x, height // 2 - self.corridor_length, None)
self.grid.set(x, height // 2 + self.corridor_length, None)
else:
for y in range(height // 2 - self.corridor_length, height // 2 + 1):
self.grid.set(width // 2, y, None)
for x in range(width // 2 - self.corridor_length, width // 2 + self.corridor_length + 1):
self.grid.set(x, height // 2 - self.corridor_length, None)
# Create rewards
reward_positions = self._reward_positions(width, height)
self._gen_rewards(reward_positions)
def _reward_positions(self, width, height):
reward_positions = [
(width // 2 - self.corridor_length, height // 2 - self.corridor_length),
(width // 2 + self.corridor_length, height // 2 - self.corridor_length),
(width // 2 + self.corridor_length, height // 2 + self.corridor_length),
(width // 2 - self.corridor_length, height // 2 + self.corridor_length),
]
if not self.is_double:
reward_positions = reward_positions[:2]
return reward_positions
def _reward(self):
min_steps = (1 + 2 * self.corridor_length)
if self.is_double and self.reward_position > 1:
min_steps += 2
redundant_steps = max(0, self.step_count - min_steps)
max_steps = self.max_steps - min_steps + 1
cell = self.grid.get(self.agent_pos[0], self.agent_pos[1])
max_reward = self.reward_values["fake_goal"]
if hasattr(cell, "is_goal") and cell.is_goal:
max_reward = self.reward_values["goal"]
return min(max_reward, max_reward * (1 - min(1, (redundant_steps / max_steps))))
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 render(self, mode='human', close=False, **kwargs):
reward_positions = self._reward_positions(width=self.width, height=self.height)
goal = self.grid.get(*reward_positions[self.reward_position])
assert goal.is_goal
start_color = goal.color
goal.color = 'blue'
ret = super().render(mode, close, **kwargs)
goal.color = start_color
return ret
class NineRoomsEnv(MiniGridSimple):
# Only 4 actions needed, left, right, up and down
class NineRoomsCardinalActions(IntEnum):
# Cardinal movement
right = 0
down = 1
left = 2
up = 3
def __len__(self):
return 4
def __init__(
self,
grid_size=20,
passage_size=1,
max_steps=100,
seed=133,
rnd_start=0,
start_state_exclude_rooms=[],
):
self.grid_size = grid_size
self.passage_size = passage_size
self._goal_default_pos = (1, 1)
# set to 1 if agent is to be randomly spawned
self.rnd_start = rnd_start
# If self.rnd_start =1, don't spawn in these rooms
self.start_state_exclude_rooms = start_state_exclude_rooms
super().__init__(grid_size=grid_size,
max_steps=max_steps,
seed=seed,
see_through_walls=False)
self.nActions = len(NineRoomsEnv.NineRoomsCardinalActions)
# Set the action and observation spaces
self.actions = NineRoomsEnv.NineRoomsCardinalActions
self.action_space = spaces.Discrete(self.nActions)
self.max_cells = (grid_size - 1) * (grid_size - 1)
self.observation_space = spaces.Tuple(
[spaces.Discrete(grid_size),
spaces.Discrete(grid_size)])
self.observation_size = self.grid_size * self.grid_size
self.observation_shape = (self.observation_size, )
self.T = max_steps
# Change the observation space to return the position in the grid
@property
def category(self):
# [TODO] Make sure this doesn't break after self.agent_pos is changed to numpy.ndarray
return self.cell_cat_map[self.agent_pos]
def reward(self):
# -1 for every action except if the action leads to the goal state
return 1 if self.success else 0
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 reset(self, val=False, seen=True):
obs, info = super().reset(val=val, seen=seen)
# add state feature to obs
state_feat = self._encode_state(obs['agent_pos'])
obs.update(dict(state_feat=state_feat))
return obs, info
def step(self, action):
self.step_count += 1
'''
Reward doesn't depend on action, but just state.
reward = -1 if not (in_goal_state) else 0
'''
if not self.done:
# check if currently at the goal state
if self.agent_pos == self.mission:
# No penalty, episode done
self.done = True
self.success = True
else:
# Cardinal movement
if action in self.move_actions:
move_pos = self.around_pos(action)
fwd_cell = self.grid.get(*move_pos)
self.agent_dir = (action - 1) % 4
if fwd_cell == None or fwd_cell.can_overlap(
) or self.is_goal(move_pos):
self.agent_pos = move_pos
else:
raise ValueError("Invalid Action: {} ".format(action))
reward = self.reward()
if self.step_count >= self.max_steps - 1:
# print("Max Steps Exceeded.")
self.done = True
obs = self.gen_obs()
# Add state features to the observation
state_feat = self._encode_state(obs['agent_pos'])
obs.update(dict(state_feat=state_feat))
info = {
'done': self.done,
'agent_pos': np.array(self.agent_pos),
}
if self.render_rgb:
info['rgb_grid'] = self.render(mode='rgb_array')
if self.done:
info.update({
'image': self.encode_grid(),
'success': self.success,
'agent_pos': self.agent_pos,
})
return obs, reward, self.done, info
def _encode_state(self, state):
"""
Encode the state to generate observation.
"""
feat = np.ones(self.width * self.height, dtype=float)
curr_x, curr_y = state[1], state[0]
curr_pos = curr_y * self.width + curr_x
feat[curr_pos:] = 0
return feat
class SnakeEnv(MiniGridEnv):
"""
Empty grid environment, no obstacles, sparse reward
"""
# Enumeration of possible actions
class Actions(IntEnum):
# Turn left, turn right, move forward
left = 0
right = 1
forward = 2
def __init__(self, size=9):
super().__init__(grid_size=size,
max_steps=None,
see_through_walls=True)
self.actions = SnakeEnv.Actions
self.action_space = spaces.Discrete(len(self.actions))
# self.observation_space = spaces.Dict({
# 'image': spaces.Box(
# low=0,
# high=255,
# shape=(size,size,3),
# dtype='uint8'
# )
#
# })
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())
def _gen_grid(self, width, height):
# Create an empty grid
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 reset(self):
return super().reset()
# def gen_obs(self):
# image = self.grid.encode()
#
# obs = {
# 'image': image,
# 'direction': self.agent_dir,
# 'mission': self.mission
# }
#
# return obs
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, {}
class Cluttered(MiniGridSimple):
# Only 4 actions needed, left, right, up and down
class ClutteredCardinalActions(IntEnum):
# Cardinal movement
right = 0
down = 1
left = 2
up = 3
def __len__(self):
return 4
def __init__(
self,
grid_size=20,
num_objects=5,
obj_size=3,
max_steps=100,
seed=133,
state_encoding="thermal",
rnd_start=0,
):
self.state_encoding = state_encoding
self.grid_size = grid_size
self.num_objects = num_objects
self.obj_size = obj_size
# set to 1 if agent is to be randomly spawned
self.rnd_start = rnd_start
self.grid_seed = 12
# This only works for 15x15 grid with 6 obstacles
#self._goal_default_pos = (6, 10)
#self._goal_default_pos = (self.grid_size-2, self.grid_size-2)
self._goal_default_pos = (7, 12)
# This is used for some of the experiments.
self._agent_default_pos = (7, 6)
# If self.rnd_start =1, don't spawn in these rooms
super().__init__(grid_size=grid_size,
max_steps=max_steps,
seed=seed,
see_through_walls=False)
self.nActions = len(Cluttered.ClutteredCardinalActions)
# Set the action and observation spaces
self.actions = Cluttered.ClutteredCardinalActions
self.action_space = spaces.Discrete(self.nActions)
self.max_cells = (grid_size - 1) * (grid_size - 1)
self.observation_space = spaces.Tuple(
[spaces.Discrete(grid_size),
spaces.Discrete(grid_size)])
self.observation_size = self.grid_size * self.grid_size
self.observation_shape = (self.observation_size, )
self.T = max_steps
# Change the observation space to return the position in the grid
def reward(self):
# -1 for every action except if the action leads to the goal state
#return 0 if self.success else -1
return 0 if self.success else -1 / self.T
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 reset(self, val=False, seen=True):
obs, info = super().reset(val=val, seen=seen)
# add state feature to obs
state_feat = self._encode_state(obs['agent_pos'])
obs.update(dict(state_feat=state_feat))
return obs, info
def step(self, action):
self.step_count += 1
'''
Reward doesn't depend on action, but just state.
reward = -1 if not (in_goal_state) else 0
'''
if not self.done:
# check if currently at the goal state
if self.agent_pos == self.mission:
# No penalty, episode done
self.done = True
self.success = True
else:
# Cardinal movement
if action in self.move_actions:
move_pos = self.around_pos(action)
fwd_cell = self.grid.get(*move_pos)
self.agent_dir = (action - 1) % 4
if fwd_cell == None or fwd_cell.can_overlap(
) or self.is_goal(move_pos):
self.agent_pos = move_pos
else:
raise ValueError("Invalid Action: {} ".format(action))
reward = self.reward()
if self.step_count >= self.max_steps - 1:
# print("Max Steps Exceeded.")
self.done = True
obs = self.gen_obs()
# Add state features to the observation
state_feat = self._encode_state(obs['agent_pos'])
obs.update(dict(state_feat=state_feat))
info = {
'done': self.done,
'agent_pos': np.array(self.agent_pos),
}
if self.render_rgb:
info['rgb_grid'] = self.render(mode='rgb_array')
if self.done:
info.update({
'image': self.encode_grid(),
'success': self.success,
'agent_pos': self.agent_pos,
})
return obs, reward, self.done, info
def _encode_state(self, state):
"""
Encode the state to generate observation.
"""
feat = np.ones(self.width * self.height, dtype=float)
curr_x, curr_y = state[1], state[0]
curr_pos = curr_y * self.width + curr_x
if self.state_encoding == "thermal":
feat[curr_pos:] = 0
elif self.state_encoding == "one-hot":
feat[:] = 0
feat[curr_pos] = 1
return feat
class SnakeEnv(MiniGridEnv):
class Actions(IntEnum):
left = 0
right = 1
forward = 2
def __init__(self, size=9):
super().__init__(grid_size=size,
max_steps=None,
see_through_walls=True)
self.actions = SnakeEnv.Actions
self.action_space = spaces.Discrete(len(self.actions))
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())
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 reset(self):
return super().reset()
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, {}
class EmptyGridWorld(MiniGridSimple):
# Only 4 actions needed, left, right, up and down
class CardnalActions(IntEnum):
# Cardinal movement
right = 0
down = 1
left = 2
up = 3
def __len__(self):
return 4
def __init__(
self,
grid_size=20,
max_steps=100,
state_encoding="thermal",
seed=133,
rnd_start=0,
):
self.state_encoding = state_encoding
self.grid_size = grid_size
self._goal_default_pos = (self.grid_size - 2, 1)
# set to 1 if agent is to be randomly spawned
self.rnd_start = rnd_start
super().__init__(grid_size=grid_size,
max_steps=max_steps,
seed=seed,
see_through_walls=False)
self.nActions = len(EmptyGridWorld.CardnalActions)
# Set the action and observation spaces
self.actions = EmptyGridWorld.CardnalActions
self.action_space = spaces.Discrete(self.nActions)
self.max_cells = (grid_size - 1) * (grid_size - 1)
self.observation_space = spaces.Tuple(
[spaces.Discrete(grid_size),
spaces.Discrete(grid_size)])
self.observation_size = self.grid_size * self.grid_size
self.observation_shape = (self.observation_size, )
self.T = max_steps
# Change the observation space to return the position in the grid
@property
def category(self):
# [TODO] Make sure this doesn't break after self.agent_pos is changed to numpy.ndarray
return self.cell_cat_map[self.agent_pos]
def reward(self):
# -1 for every action except if the action leads to the goal state
return 1 if self.success else -1 / self.T
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 reset(self, val=False, seen=True):
obs, info = super().reset(val=val, seen=seen)
# add state feature to obs
state_feat = self._encode_state(obs['agent_pos'])
obs.update(dict(state_feat=state_feat))
return obs, info
def step(self, action):
self.step_count += 1
'''
Reward doesn't depend on action, but just state.
reward = -1 if not (in_goal_state) else 0
'''
if not self.done:
# check if currently at the goal state
if self.agent_pos == self.mission:
# No penalty, episode done
self.done = True
self.success = True
else:
# Cardinal movement
if action in self.move_actions:
move_pos = self.around_pos(action)
fwd_cell = self.grid.get(*move_pos)
self.agent_dir = (action - 1) % 4
if fwd_cell == None or fwd_cell.can_overlap(
) or self.is_goal(move_pos):
self.agent_pos = move_pos
else:
raise ValueError("Invalid Action: {} ".format(action))
reward = self.reward()
if self.step_count >= self.max_steps - 1:
# print("Max Steps Exceeded.")
self.done = True
obs = self.gen_obs()
# Add state features to the observation
state_feat = self._encode_state(obs['agent_pos'])
obs.update(dict(state_feat=state_feat))
info = {
'done': self.done,
'agent_pos': np.array(self.agent_pos),
}
if self.render_rgb:
info['rgb_grid'] = self.render(mode='rgb_array')
if self.done:
info.update({
'image': self.encode_grid(),
'success': self.success,
'agent_pos': self.agent_pos,
})
return obs, reward, self.done, info
def _encode_state(self, state):
"""
Encode the state to generate observation.
"""
feat = np.ones(self.width * self.height, dtype=float)
curr_x, curr_y = state[0], state[1]
curr_pos = curr_y * self.width + curr_x
if self.state_encoding == "thermal":
feat[curr_pos:] = 0
elif self.state_encoding == "one-hot":
feat[:] = 0
feat[curr_pos] = 1
return feat