def reset_world(self):
agents = self.create_agents()
pois = self.create_pois()
self._world = World(self.world_width, self.world_height,
self.num_agents, self.num_pois, agents, pois)
self._agents = self._world.get_agents()
self._pois = self._world.get_pois()
def test_apply_cont_actions():
agents = [Agent((0, 0))]
world = World(10, 10, 1, 0, agents)
commands = np.array([np.array([1, 2])])
world.apply_cont_actions(commands)
assert np.array_equal(np.array([1, 2]), world.get_agents()[0].get_loc())
def test_get_obs_image_scale():
agents = [Agent((1.1, 1))]
pois = [POI((2.5, 1.2))]
world = World(10, 10, 1, 1, agents, pois)
img = world.get_obs_image(10)
assert img[10][11][0] == 1
assert img[12][25][1] == 1
def test_get_obs_image():
agents = [Agent((1.1, 1))]
pois = [POI((2.5, 1.2))] # Even numbers round up, odd round down
world = World(10, 10, 1, 1, agents, pois)
img = world.get_obs_image()
assert img[1][1][0] == 1
assert img[1][2][1] == 1
def test_get_local_obs_images():
agents = [Agent((1.1, 1))]
pois = [POI((2.5, 1.2))]
world = World(10, 10, 1, 1, agents, pois)
imgs = world.get_local_obs_images(3)
agent_img = imgs[0]
assert agent_img[1][1][0] == 1 # Observe self
assert agent_img[1][2][1] == 1 # Observe POI
def test_init_creates_valid_pois():
world = World(10, 10, 0, 2)
valid = True
for p in world.get_pois():
loc = p.get_loc()
x, y = loc[0], loc[1]
valid = valid and x >= 0 and x < 10 and y >= 0 and y < 10
assert valid
assert len(world.get_pois()) == 2
def test_init_creates_valid_agents():
world = World(10, 10, 2, 0)
valid = True
for a in world.get_agents():
loc = a.get_loc()
x, y = loc[0], loc[1]
valid = valid and x >= 0 and x < 10 and y >= 0 and y < 10
assert valid
assert len(world.get_agents()) == 2
def test_get_obs_states():
agents = [
Agent((1, 1)),
Agent((0, 1.5)),
Agent((1.5, 0)),
Agent((0, 0)),
Agent((1.5, 1.5))
]
pois = [POI((0, 0))]
world = World(5, 5, 5, 1, agents, pois)
vec = world.get_obs_states()
assert vec.shape[0] == 5
assert vec[0][0] == 1
assert vec[0][6] == 1 / np.sqrt(2)
def create_world(self):
self._world = World(self.world_width, self.world_height,
self.num_agents, self.num_pois)
self._agents = self._world.get_agents()
self._pois = self._world.get_pois()
class RoverContFeature(gym.Env):
metadata = {'render.modes': ['human']}
COLOR = {1: [255, 0, 0], 2: [0, 255, 255], 3: [0, 0, 255]}
def __init__(self):
# Set Default values for env. Because gyms can't have arguments, set
# after construction
self.world_height = 10
self.world_width = 10
self.num_agents = 1
self.num_pois = 1
self.time_limit = 10
# Creates the Space objects used gym
self.set_observation_space()
self.set_action_space()
# Sets internal state for rover domain
self.create_world()
self.time_step = 0
# Env requirements
self._seed()
self.reset()
self.viewer = None
""" Gym env required functions """
def _reset(self):
self.time_step = 0
self.create_world()
return self._get_observation()
def _seed(self, seed=None):
self.np_random, seed = seeding.np_random(seed)
return [seed]
def _step(self, action):
self._world.apply_cont_actions(action)
self.time_step += 1
obs = self._get_observation()
reward = self._world.get_reward()
done = self.time_step > self.time_limit
return obs, reward, done, {}
def _get_observation(self):
return self._world.get_obs_states()
def _render(self, mode='human', close=False):
from gym.envs.classic_control import rendering
if close:
if self.viewer is not None:
self.viewer.close()
self.viewer = None
return
if self.viewer is None:
self.screen_width = 600
self.screen_height = 400
self.scale_w = self.screen_width / self.world_width
self.scale_h = self.screen_height / self.world_height
self.viewer = rendering.Viewer(self.screen_width,
self.screen_height)
self._render_agents()
self._render_pois()
return self.viewer.render(return_rgb_array=mode == 'rgb_array')
""" Helper functions """
def create_world(self):
self._world = World(self.world_width, self.world_height,
self.num_agents, self.num_pois)
self._agents = self._world.get_agents()
self._pois = self._world.get_pois()
def set_observation_space(self):
self._box_low = np.array([0, 0, 0, 0, 0, 0, 0, 0])
self._box_high = np.array([
self.num_agents, self.num_agents, self.num_agents, self.num_agents,
6 * self.num_pois, 6 * self.num_pois, 6 * self.num_pois,
6 * self.num_pois
])
self._box_one_obs = spaces.Box(self._box_low, self._box_high)
self._all_boxes = []
for _ in range(self.num_agents):
self._all_boxes.append(self._box_one_obs)
self.observation_space = spaces.Tuple(self._all_boxes)
def set_action_space(self):
self._agent_action = spaces.Box(np.array([-1, -1]), np.array([1, 1]))
self._all_actions = []
for _ in range(self.num_agents):
self._all_actions.append(self._agent_action)
self.action_space = spaces.Tuple(self._all_actions)
def _render_agents(self):
from gym.envs.classic_control import rendering
for agent in self._agents:
poly = rendering.FilledPolygon(
self._agent_square(agent, 10, self.scale_w, self.scale_h))
poly.set_color(self.COLOR[1][0], self.COLOR[1][1],
self.COLOR[1][2])
self.viewer.add_onetime(poly)
def _render_pois(self):
from gym.envs.classic_control import rendering
for poi in self._pois:
poly = rendering.make_circle()
trans = (poi.get_loc()[0] * self.scale_w,
poi.get_loc()[1] * self.scale_h)
poly.add_attr(rendering.Transform(translation=trans))
if poi.visible():
poly.set_color(self.COLOR[2][0], self.COLOR[2][1],
self.COLOR[2][2])
else:
poly.set_color(self.COLOR[3][0], self.COLOR[3][1],
self.COLOR[3][2])
self.viewer.add_onetime(poly)
def _agent_square(self, agent, side, scale_w=1, scale_h=1):
loc = agent.get_loc()
sloc = np.array([loc[0] * scale_w, loc[1] * scale_h])
half = side // 2
""" Points in CW order """
points = [
np.add(sloc, np.array([half, half])),
np.add(sloc, np.array([half, -half])),
np.add(sloc, np.array([-half, -half])),
np.add(sloc, np.array([-half, half]))
]
return points
def test_apply_cont_actions_mismatch():
world = World(10, 10, 1, 0)
commands = np.array([np.array([1, 2]), np.array([4, 3])])
with pytest.raises(AssertionError):
world.apply_cont_actions(commands)
def test_apply_discrete_actions_mismatch():
world = World(10, 10, 1, 0)
commands = np.array([1, 5])
with pytest.raises(AssertionError):
world.apply_discrete_actions(commands)
class RoverEnv(gym.Env):
metadata = {'render.modes': ['human']}
COLOR = {1: [255, 0, 0], 2: [0, 255, 255], 3: [0, 0, 255]}
def __init__(self,
world_height=10,
world_width=10,
num_agents=1,
num_pois=1,
time_limit=20,
observation_mode='feature',
actions='discrete',
image_width=5,
image_scale=1):
""" Rover Domain environment.
"""
self.world_height = world_height
self.world_width = world_width
self.num_agents = num_agents
self.num_pois = num_pois
self.time_limit = time_limit
self.observation_mode = observation_mode
self.actions = actions
self.image_width = image_width
self.image_scale = image_scale
self.set_observation_space()
self.set_action_space()
#self._world = self.reset_world()
self.create_world()
self.viewer = None
self.time_step = 0
self._seed()
self.reset()
def reset_world(self):
agents = self.create_agents()
pois = self.create_pois()
self._world = World(self.world_width, self.world_height,
self.num_agents, self.num_pois, agents, pois)
self._agents = self._world.get_agents()
self._pois = self._world.get_pois()
def create_agents(self):
return np.array([Agent(np.array([3, 4]))])
def create_pois(self):
return np.array([POI(np.array([7, 9]))])
def _seed(self, seed=None):
self.np_random, seed = seeding.np_random(seed)
return [seed]
def step(self, action):
#assert self.action_space.contains(action), "%r (%s) invalid" % (action, type(action))
if self.actions == 'continuous':
self._world.apply_cont_actions(action)
else:
self._world.apply_discrete_actions(action)
self.time_step += 1
obs = self._get_observation()
reward = self._world.get_reward()
done = self.time_step > self.time_limit # or not self._world.pois_still_left()
return obs, reward, done, {}
def reset(self):
#self.set_observation_space()
#self.set_action_space()
self.time_step = 0
#self.reset_world()
self.create_world()
return self._get_observation()
def create_world(self):
self._world = World(self.world_width, self.world_height,
self.num_agents, self.num_pois)
self._agents = self._world.get_agents()
self._pois = self._world.get_pois()
def set_observation_space(self):
if self.observation_mode == 'feature':
self.set_observation_space_feature()
elif self.observation_mode == 'image':
self.set_observation_space_image()
def set_observation_space_feature(self):
self._box_low = np.array([0, 0, 0, 0, 0, 0, 0, 0])
self._box_high = np.array([
self.num_agents, self.num_agents, self.num_agents, self.num_agents,
6 * self.num_pois, 6 * self.num_pois, 6 * self.num_pois,
6 * self.num_pois
])
self._box_one_obs = spaces.Box(self._box_low, self._box_high)
self._all_boxes = []
for _ in range(self.num_agents):
self._all_boxes.append(self._box_one_obs)
self.observation_space = spaces.Tuple(self._all_boxes)
print(self.observation_space)
def set_observation_space_image(self):
all_images = []
max_count = max(self.num_pois, self.num_agents)
for _ in range(self.num_agents):
single_space = spaces.Box(low=0,
high=max_count,
shape=(self.image_width,
self.image_width, 2))
all_images.append(single_space)
self.observation_space = spaces.Tuple(all_images)
def set_action_space(self):
if self.actions == 'continuous':
self._set_action_space_cont()
elif self.actions == 'discrete':
self._set_action_space_discrete()
def _set_action_space_cont(self):
self._agent_action = spaces.Box(np.array([-1, -1]), np.array([1, 1]))
self._all_actions = []
for _ in range(self.num_agents):
self._all_actions.append(self._agent_action)
self.action_space = spaces.Tuple(self._all_actions)
def _set_action_space_discrete(self):
self._agent_action = spaces.Discrete(8)
self._all_actions = []
for _ in range(self.num_agents):
self._all_actions.append(self._agent_action)
self.action_space = spaces.Tuple(self._all_actions)
def render(self, mode='human', close=False):
from gym.envs.classic_control import rendering
if close:
if self.viewer is not None:
self.viewer.close()
self.viewer = None
return
if self.viewer is None:
self.screen_width = 600
self.screen_height = 400
self.scale_w = self.screen_width / self.world_width
self.scale_h = self.screen_height / self.world_height
self.viewer = rendering.Viewer(self.screen_width,
self.screen_height)
self._render_agents()
self._render_pois()
return self.viewer.render(return_rgb_array=mode == 'rgb_array')
def _render_agents(self):
from gym.envs.classic_control import rendering
for agent in self._agents:
poly = rendering.FilledPolygon(
self._agent_square(agent, 10, self.scale_w, self.scale_h))
poly.set_color(self.COLOR[1][0], self.COLOR[1][1],
self.COLOR[1][2])
self.viewer.add_onetime(poly)
def _render_pois(self):
from gym.envs.classic_control import rendering
for poi in self._pois:
poly = rendering.make_circle()
trans = (poi.get_loc()[0] * self.scale_w,
poi.get_loc()[1] * self.scale_h)
poly.add_attr(rendering.Transform(translation=trans))
if poi.visible():
poly.set_color(self.COLOR[2][0], self.COLOR[2][1],
self.COLOR[2][2])
else:
poly.set_color(self.COLOR[3][0], self.COLOR[3][1],
self.COLOR[3][2])
self.viewer.add_onetime(poly)
def _agent_square(self, agent, side, scale_w=1, scale_h=1):
loc = agent.get_loc()
sloc = np.array([loc[0] * scale_w, loc[1] * scale_h])
half = side // 2
""" Points in CW order """
points = [
np.add(sloc, np.array([half, half])),
np.add(sloc, np.array([half, -half])),
np.add(sloc, np.array([-half, -half])),
np.add(sloc, np.array([-half, half]))
]
return points
def _get_observation(self):
if self.observation_mode == 'feature':
obs = self._world.get_obs_states()
# for i in range(self.num_agents):
# obs[i][8] = self.time_limit - self.time_step
return obs
elif self.observation_mode == 'image':
return self._world.get_local_obs_images(self.image_width,
self.image_scale)