def __init__(self, num_predators, step_scaling_factor,
threshold_trial_len):
"""Constructor.
This class uses the meta-state to keep track of the number of steps
before the agent is caught. See the game rules section near the bottom
of this file for the counter incrementer.
Args:
step_scaling_factor: Float. Fractional decrease of predator mass
after a trial longer than threshold_trial_len. Also used as
fractional increase of predator mass after a trial shorter than
threshold_trial_len. Should be small and positive.
threshold_trial_len: Length of a trial above which the predator
mass is decreased and below which the predator mass is
increased.
"""
self._mass = 1.
self._step_scaling_factor = step_scaling_factor
self._threshold_trial_len = threshold_trial_len
# Agent
agent_factors = distribs.Product(
[
distribs.Continuous('x', 0., 1.),
distribs.Continuous('y', 0., 1.)
],
shape='circle',
scale=0.1,
c0=0.33,
c1=1.,
c2=0.66,
)
self._agent_generator = sprite_generators.generate_sprites(
agent_factors, num_sprites=1)
# Predators
predator_factors = distribs.Product(
[
distribs.Continuous('x', 0., 1.),
distribs.Continuous('y', 0., 1.)
],
shape='circle',
scale=0.1,
c0=0.,
c1=1.,
c2=0.8,
)
self._predator_generator = sprite_generators.generate_sprites(
predator_factors, num_sprites=num_predators)
# Walls
self._walls = shapes.border_walls(visible_thickness=0.,
c0=0.,
c1=0.,
c2=0.5)
self._meta_state = None
def get_config(_):
"""Get environment config"""
############################################################################
# Sprite initialization
############################################################################
# Ball generator
ball_factors = distribs.Product(
[
distribs.Continuous('x', 0.25, 0.75),
distribs.Continuous('y', 0.5, 0.9),
distribs.Continuous('x_vel', -0.01, 0.01)
],
scale=0.1,
shape='circle',
c0=0,
c1=0,
c2=255,
mass=1.,
)
ball_generator = sprite_generators.generate_sprites(ball_factors,
num_sprites=4)
# Walls
bottom_wall = [[-1, 0.1], [2, 0.1], [2, -1], [-1, -1]]
left_wall = [[0.05, -0.1], [0.05, 1.1], [-1, 1.1], [-1, -0.1]]
right_wall = [[0.95, -0.1], [0.95, 1.1], [2, 1.1], [2, -0.1]]
divider = [[0.45, -1], [0.45, 0.3], [0.55, 0.3], [0.55, -1]]
walls = [
sprite_lib.Sprite(shape=np.array(v), x=0, y=0, c0=128, c1=128, c2=128)
for v in [bottom_wall, left_wall, right_wall, divider]
]
def state_initializer():
"""Callable returning new state at each episode reset."""
state = collections.OrderedDict([
('walls', walls),
('balls', ball_generator(disjoint=True)),
('agent', []),
])
return state
############################################################################
# Physics
############################################################################
# Setting max_recursion_depth > 0 can increase stability
# Setting update_angle_vel = False is recommended for stability
collision = physics_lib.Collision(
elasticity=0.6,
symmetric=False,
update_angle_vel=False,
max_recursion_depth=2,
)
physics = physics_lib.Physics(
(collision, 'balls', ['balls', 'walls']),
(physics_lib.DownGravity(g=-0.001), 'balls'),
updates_per_env_step=20,
)
############################################################################
# Task
############################################################################
task = tasks.CompositeTask(timeout_steps=100)
############################################################################
# Action space
############################################################################
# Need an action space, so let it control an empty agent layer
action_space = action_spaces.Grid(action_layers='agent')
############################################################################
# Observer
############################################################################
observer = observers.PILRenderer(image_size=(64, 64), anti_aliasing=1)
############################################################################
# Final config
############################################################################
config = {
'state_initializer': state_initializer,
'physics': physics,
'task': task,
'action_space': action_space,
'observers': {
'image': observer
},
'game_rules': (),
}
return config
def get_config(_):
"""Get environment config."""
############################################################################
# Sprite initialization
############################################################################
# Agent
agent_factors = distribs.Product(
[distribs.Continuous('x', 0.1, 0.9),
distribs.Continuous('y', 0.1, 0.9)],
shape='circle', scale=0.1, c0=0.33, c1=1., c2=0.66,
)
# Predators
shape_0 = 1.8 * np.array(
[[-0.3, -0.3], [0.1, -0.7], [0.4, 0.6], [-0.1, 0.25]])
shape_1 = 1.5 * np.array(
[[-0.5, -0.3], [-0.1, -0.7], [0.7, 0.1], [0., -0.1], [-0.3, 0.25]])
predator_factors = distribs.Product(
[distribs.Continuous('x', 0.2, 0.8),
distribs.Continuous('y', 0.2, 0.8),
distribs.Discrete(
'shape', [shape_0, shape_1, 'star_5', 'triangle', 'spoke_5']),
distribs.Continuous('angle', 0., 2 * np.pi),
distribs.Continuous('aspect_ratio', 0.75, 1.25),
distribs.Continuous('scale', 0.1, 0.15),
distribs.Continuous('x_vel', -0.03, 0.03),
distribs.Continuous('y_vel', -0.03, 0.03),
distribs.Continuous('angle_vel', -0.05, 0.05)],
c0=0., c1=1., c2=0.8,
)
# Walls
walls = shapes.border_walls(visible_thickness=0.05, c0=0., c1=0., c2=0.5)
# Create callable initializer returning entire state
agent_generator = sprite_generators.generate_sprites(
agent_factors, num_sprites=1)
predator_generator = sprite_generators.generate_sprites(
predator_factors, num_sprites=5)
def state_initializer():
predators = predator_generator(
disjoint=True, without_overlapping=walls)
agent = agent_generator(without_overlapping=walls + predators)
state = collections.OrderedDict([
('walls', walls),
('predators', predators),
('agent', agent),
])
return state
############################################################################
# Physics
############################################################################
agent_friction_force = physics_lib.Drag(coeff_friction=0.25)
asymmetric_collision = physics_lib.Collision(
elasticity=1., symmetric=False, update_angle_vel=True)
symmetric_collision = physics_lib.Collision(
elasticity=1., symmetric=True, update_angle_vel=True)
agent_wall_collision = physics_lib.Collision(
elasticity=0., symmetric=False, update_angle_vel=False)
forces = (
(agent_friction_force, 'agent'),
(symmetric_collision, 'predators', 'predators'),
(asymmetric_collision, 'predators', 'walls'),
(agent_wall_collision, 'agent', 'walls'),
)
physics = physics_lib.Physics(*forces, updates_per_env_step=10)
############################################################################
# Task
############################################################################
predator_task = tasks.ContactReward(
-5, layers_0='agent', layers_1='predators')
stay_alive_task = tasks.StayAlive(
reward_period=20,
reward_value=0.2,
)
task = tasks.CompositeTask(
predator_task, stay_alive_task, timeout_steps=200)
############################################################################
# Action space
############################################################################
action_space = action_spaces.Joystick(
scaling_factor=0.01, action_layers='agent')
############################################################################
# Observer
############################################################################
observer = observers.PILRenderer(
image_size=(64, 64), anti_aliasing=1, color_to_rgb='hsv_to_rgb')
############################################################################
# Final config
############################################################################
config = {
'state_initializer': state_initializer,
'physics': physics,
'task': task,
'action_space': action_space,
'observers': {'image': observer},
}
return config
def get_config(_):
"""Get environment config."""
############################################################################
# Sprite initialization
############################################################################
# Agent
agent_factors = distribs.Product(
[
distribs.Continuous('x', 0.1, 0.9),
distribs.Continuous('y', 0.1, 0.9)
],
shape='circle',
scale=0.1,
c0=0.33,
c1=1.,
c2=0.7,
)
# Predators
predator_factors = distribs.Product(
[
distribs.Continuous('x', 0.1, 0.9),
distribs.Continuous('y', 0.1, 0.9)
],
shape='circle',
scale=0.1,
c0=0.,
c1=1.,
c2=0.8,
)
# Prey
prey_factors = distribs.Product(
[
distribs.Continuous('x', 0.1, 0.9),
distribs.Continuous('y', 0.1, 0.9)
],
shape='circle',
scale=0.1,
c0=0.2,
c1=1.,
c2=1.,
)
# Boosters
booster_factors = distribs.Product(
[
distribs.Continuous('x', 0.1, 0.9),
distribs.Continuous('y', 0.1, 0.9)
],
shape='triangle',
scale=0.1,
c0=0.6,
c1=1.,
c2=1.,
)
# Portals
portal_factors = dict(shape='square', scale=0.1, c0=0., c1=0., c2=0.95)
portal_sprites = [
sprite.Sprite(x=0.125, y=0.125, **portal_factors),
sprite.Sprite(x=0.875, y=0.875, **portal_factors),
]
# Walls
wall_color = dict(c0=0., c1=0., c2=0.5)
island_wall_shape_0 = np.array([[0.2, 0.2], [0.4, 0.2], [0.4, 0.4],
[0.2, 0.4]])
island_wall_shapes = [
island_wall_shape_0,
island_wall_shape_0 + np.array([[0., 0.4]]),
island_wall_shape_0 + np.array([[0.4, 0.4]]),
island_wall_shape_0 + np.array([[0.4, 0.]]),
]
island_walls = [
sprite.Sprite(shape=shape, x=0., y=0., **wall_color)
for shape in island_wall_shapes
]
boundary_walls = shapes.border_walls(visible_thickness=0.05, **wall_color)
walls = boundary_walls + island_walls
# Callable sprite generators
agent_generator = sprite_generators.generate_sprites(agent_factors,
num_sprites=1)
predator_generator = sprite_generators.generate_sprites(predator_factors,
num_sprites=1)
prey_generator = sprite_generators.generate_sprites(
prey_factors, num_sprites=lambda: np.random.randint(2, 5))
booster_generator = sprite_generators.generate_sprites(booster_factors,
num_sprites=2)
# Create callable initializer returning entire state
def state_initializer():
portals = portal_sprites
agent = agent_generator(without_overlapping=walls)
predators = predator_generator(without_overlapping=walls + agent)
boosters = booster_generator(without_overlapping=walls + agent)
prey = prey_generator(without_overlapping=walls)
state = collections.OrderedDict([
('walls', walls),
('portals', portals),
('boosters', boosters),
('prey', prey),
('predators', predators),
('agent', agent),
])
return state
############################################################################
# Physics
############################################################################
agent_friction_force = physics_lib.Drag(coeff_friction=0.25)
predator_friction_force = physics_lib.Drag(coeff_friction=0.05)
predator_random_force = physics_lib.RandomForce(max_force_magnitude=0.02)
prey_friction_force = physics_lib.Drag(coeff_friction=0.02)
prey_random_force = physics_lib.RandomForce(max_force_magnitude=0.02)
predator_attraction = physics_lib.DistanceForce(
force_fn=physics_lib.linear_force_fn(zero_intercept=-0.002,
slope=0.001))
asymmetric_collision = physics_lib.Collision(elasticity=0.25,
symmetric=False,
update_angle_vel=False)
forces = (
(agent_friction_force, 'agent'),
(predator_friction_force, 'predators'),
(predator_random_force, 'predators'),
(prey_friction_force, 'prey'),
(prey_random_force, 'prey'),
(predator_attraction, 'agent', 'predators'),
(asymmetric_collision, ['agent', 'predators', 'prey'], 'walls'),
)
physics = physics_lib.Physics(*forces, updates_per_env_step=5)
############################################################################
# Task
############################################################################
predator_task = tasks.ContactReward(-5,
layers_0='agent',
layers_1='predators',
reset_steps_after_contact=0)
prey_task = tasks.ContactReward(1, layers_0='agent', layers_1='prey')
reset_task = tasks.Reset(
condition=lambda state: len(state['prey']) == 0,
steps_after_condition=5,
)
task = tasks.CompositeTask(predator_task,
prey_task,
reset_task,
timeout_steps=400)
############################################################################
# Action space
############################################################################
action_space = action_spaces.Joystick(
scaling_factor=0.01,
action_layers='agent',
)
############################################################################
# Observer
############################################################################
observer = observers.PILRenderer(
image_size=(64, 64),
anti_aliasing=1,
color_to_rgb='hsv_to_rgb',
)
############################################################################
# Game rules
############################################################################
disappear_rule = game_rules.VanishOnContact(vanishing_layer='prey',
contacting_layer='agent')
portal_rule = game_rules.Portal(teleporting_layer='agent',
portal_layer='portals')
rules = (disappear_rule, portal_rule, Booster())
############################################################################
# Final config
############################################################################
config = {
'state_initializer': state_initializer,
'physics': physics,
'task': task,
'action_space': action_space,
'observers': {
'image': observer
},
'game_rules': rules,
}
return config
def get_config(_):
"""Get environment config."""
############################################################################
# Sprite initialization
############################################################################
# Agents
agent_factors = distribs.Product(
[distribs.Continuous('x', 0., 1.),
distribs.Continuous('y', 0.35, 0.65)],
shape='circle', scale=0.1, c1=1., c2=0.7,
)
agent_0_factors = distribs.Product([agent_factors], c0=0.2)
agent_1_factors = distribs.Product([agent_factors], c0=0.1)
agent_2_factors = distribs.Product([agent_factors], c0=0.)
# Walls
walls = shapes.border_walls(visible_thickness=0.05, c0=0., c1=0., c2=0.5)
# Fountains
fountain_factors = {
'shape': 'circle', 'scale': 0.05, 'c0': 0.6, 'c1': 1., 'c2': _BAD_VALUE}
fountains_across = np.linspace(0.1, 0.9, 6)
fountains_up = np.linspace(0.75, 0.9, 2)
fountains_grid_x, fountains_grid_y = np.meshgrid(fountains_across,
fountains_up)
fountains_positions = zip(np.ravel(fountains_grid_x),
np.ravel(fountains_grid_y))
fountain_sprites = [
sprite.Sprite(x=x, y=y, **fountain_factors)
for (x, y) in fountains_positions
]
# Fruits
fruit_factors = {
'shape': 'circle', 'scale': 0.05, 'c0': 0.3, 'c1': 1., 'c2': _BAD_VALUE}
fruits_across = np.linspace(0.1, 0.9, 6)
fruits_up = np.linspace(0.1, 0.25, 2)
fruits_grid_x, fruits_grid_y = np.meshgrid(fruits_across, fruits_up)
fruits_positions = zip(np.ravel(fruits_grid_x), np.ravel(fruits_grid_y))
fruit_sprites = [
sprite.Sprite(x=x, y=y, **fruit_factors)
for (x, y) in fruits_positions
]
# Create callable initializer returning entire state
agent_0_generator = sprite_generators.generate_sprites(
agent_0_factors, num_sprites=1)
agent_1_generator = sprite_generators.generate_sprites(
agent_1_factors, num_sprites=1)
agent_2_generator = sprite_generators.generate_sprites(
agent_2_factors, num_sprites=1)
def state_initializer():
agent_0 = agent_0_generator(without_overlapping=walls)
agent_1 = agent_1_generator(without_overlapping=walls)
agent_2 = agent_2_generator(without_overlapping=walls)
state = collections.OrderedDict([
('walls', walls),
('fountains', fountain_sprites),
('fruits', fruit_sprites),
('agent_2', agent_2),
('agent_1', agent_1),
('agent_0', agent_0),
])
return state
############################################################################
# Physics
############################################################################
agent_friction_force = physics_lib.Drag(coeff_friction=0.25)
asymmetric_collision = physics_lib.Collision(
elasticity=0.25, symmetric=False)
forces = (
(agent_friction_force, ['agent_0', 'agent_1', 'agent_2']),
(asymmetric_collision, ['agent_0', 'agent_1', 'agent_2'], 'walls'),
)
physics = physics_lib.Physics(*forces, updates_per_env_step=5)
############################################################################
# Task
############################################################################
task = tasks.ContactReward(
1, layers_0='agent_0', layers_1='fruits',
condition=lambda s_0, s_1: s_1.c2 > _VALUE_THRESHOLD,
)
############################################################################
# Action space
############################################################################
action_space = action_spaces.Composite(
agent_0=action_spaces.Joystick(
scaling_factor=0.005, action_layers='agent_0'),
agent_1=action_spaces.Joystick(
scaling_factor=0.005, action_layers='agent_1'),
agent_2=action_spaces.Joystick(
scaling_factor=0.005, action_layers='agent_2'),
)
############################################################################
# Observer
############################################################################
image_observer = observers.PILRenderer(
image_size=(64, 64),
anti_aliasing=1,
color_to_rgb='hsv_to_rgb',
)
raw_state_observer = observers.RawState() # needed by hand-crafted agents
############################################################################
# Game rules
############################################################################
def _spoil_fruit(sprite):
sprite.c2 = _BAD_VALUE
def _ripen_fruit(sprite):
sprite.c2 = _GOOD_VALUE
def _poison_fountain(sprite):
sprite.c2 = _BAD_VALUE
def _clean_fountain(sprite):
sprite.c2 = _GOOD_VALUE
def agents_contacting_layer(state, layer, value):
n_contact = 0
for s in state[layer]:
if s.c2 != value:
continue
n_contact += (
s.overlaps_sprite(state['agent_0'][0]) or
s.overlaps_sprite(state['agent_1'][0]) or
s.overlaps_sprite(state['agent_2'][0])
)
return n_contact
poison_fountains = game_rules.ModifySprites(
layers='fountains', modifier=_poison_fountain, sample_one=True,
filter_fn=lambda s: s.c2 > _VALUE_THRESHOLD)
poison_fountains = game_rules.ConditionalRule(
condition=lambda s: agents_contacting_layer(s, 'fruits', _GOOD_VALUE),
rules=poison_fountains,
)
ripen_fruits = game_rules.ModifySprites(
layers='fruits', modifier=_ripen_fruit, sample_one=True,
filter_fn=lambda s: s.c2 < _VALUE_THRESHOLD)
ripen_fruits = game_rules.ConditionalRule(
condition=lambda s: agents_contacting_layer(s, 'fountains', _BAD_VALUE),
rules=ripen_fruits,
)
spoil_fruits = game_rules.ModifyOnContact(
layers_0='fruits',
layers_1=('agent_0', 'agent_1', 'agent_2'),
modifier_0=_spoil_fruit,
filter_0=lambda s: s.c2 > _VALUE_THRESHOLD)
clean_fountains = game_rules.ModifyOnContact(
layers_0='fountains',
layers_1=('agent_0', 'agent_1', 'agent_2'),
modifier_0=_clean_fountain,
filter_0=lambda s: s.c2 < _VALUE_THRESHOLD)
rules = (poison_fountains, spoil_fruits, ripen_fruits, clean_fountains)
############################################################################
# Final config
############################################################################
config = {
'state_initializer': state_initializer,
'physics': physics,
'task': task,
'action_space': action_space,
'observers': {'image': image_observer, 'state': raw_state_observer},
'game_rules': rules,
}
return config
def _get_config(num_prey, num_predators):
"""Get environment config."""
############################################################################
# Sprite initialization
############################################################################
# Agent
agent_factors = distribs.Product(
[distribs.Continuous('x', 0., 1.),
distribs.Continuous('y', 0., 1.)],
scale=0.08, c0=0, c1=255, c2=0,
)
# Predators
predator_factors = distribs.Product(
[distribs.Continuous('x', 0., 1.),
distribs.Continuous('y', 0., 1.),
distribs.Continuous('x_vel', -0.02, 0.02),
distribs.Continuous('y_vel', -0.02, 0.02),],
scale=0.08, shape='circle', opacity=192, c0=255, c1=0, c2=0,
)
# Prey
prey_factors = distribs.Product(
[distribs.Continuous('x', 0., 1.),
distribs.Continuous('y', 0., 1.),
distribs.Continuous('x_vel', -0.02, 0.02),
distribs.Continuous('y_vel', -0.02, 0.02),],
scale=0.08, shape='circle', opacity=192, c0=255, c1=255, c2=0,
)
# Create callable initializer returning entire state
predator_generator = sprite_generators.generate_sprites(
predator_factors, num_sprites=num_predators)
prey_generator = sprite_generators.generate_sprites(
prey_factors, num_sprites=num_prey)
def state_initializer():
"""Callable returning state at every episode reset."""
agent = sprite.Sprite(**agent_factors.sample())
predators = predator_generator(without_overlapping=(agent,))
prey = prey_generator(without_overlapping=(agent,))
state = collections.OrderedDict([
('prey', prey),
('predators', predators),
('agent', [agent]),
])
return state
############################################################################
# Physics
############################################################################
agent_friction_force = physics_lib.Drag(coeff_friction=0.25)
random_force = physics_lib.RandomForce(max_force_magnitude=0.01)
predator_attraction = physics_lib.DistanceForce(
physics_lib.linear_force_fn(zero_intercept=-0.001, slope=0.0005))
prey_avoid = physics_lib.DistanceForce(
physics_lib.linear_force_fn(zero_intercept=0.001, slope=-0.0005))
forces = (
(agent_friction_force, 'agent'),
(random_force, ['predators', 'prey']),
(predator_attraction, 'agent', 'predators'),
(prey_avoid, 'agent', 'prey'),
)
constant_speed = physics_lib.ConstantSpeed(
layer_names=['prey', 'predators'], speed=0.015)
physics = physics_lib.Physics(
*forces,
updates_per_env_step=10,
corrective_physics=[constant_speed],
)
############################################################################
# Task
############################################################################
predator_task = tasks.ContactReward(
-5, layers_0='agent', layers_1='predators', reset_steps_after_contact=0)
prey_task = tasks.ContactReward(1, layers_0='agent', layers_1='prey')
reset_task = tasks.Reset(
condition=lambda state: len(state['prey']) == 0,
steps_after_condition=5,
)
task = tasks.CompositeTask(
reset_task, predator_task, prey_task, timeout_steps=300)
############################################################################
# Action space
############################################################################
action_space = action_spaces.Joystick(
scaling_factor=0.025, action_layers='agent', control_velocity=True)
############################################################################
# Observer
############################################################################
observer = observers.PILRenderer(
image_size=(64, 64),
anti_aliasing=1,
polygon_modifier=polygon_modifiers.TorusGeometry(
['agent', 'predators', 'prey']),
)
############################################################################
# Game rules
############################################################################
prey_vanish = game_rules.VanishOnContact(
vanishing_layer='prey', contacting_layer='agent')
def _torus_position_wrap(s):
s.position = np.remainder(s.position, 1)
torus_position_wrap = game_rules.ModifySprites(
('agent', 'predators', 'prey'), _torus_position_wrap)
rules = (prey_vanish, torus_position_wrap)
############################################################################
# Final config
############################################################################
config = {
'state_initializer': state_initializer,
'physics': physics,
'task': task,
'action_space': action_space,
'observers': {'image': observer},
'game_rules': rules,
}
return config
def _get_config(num_obstacles, valid_step_range):
"""Get environment config.
Args:
num_obstacles: Int. Number of obstacles.
valid_step_range: 2-iterable of ints. (min_num_steps, max_num_steps).
All trials must have duration in this step range.
Returns:
config: Config dictionary to pass to environment constructor.
"""
############################################################################
# Physics
############################################################################
elastic_collision = physics_lib.Collision(elasticity=1.,
symmetric=False,
update_angle_vel=False)
physics = physics_lib.Physics(
(elastic_collision, 'ball', 'walls'),
updates_per_env_step=10,
)
def _predict_trial_end(state):
"""Predict whether a trial will end in step range and true response.
Args:
state: OrderedDict of sprite layers. Initial state of environment.
Returns:
valid_trial: Bool. Whether trial will end with number of steps in
valid_step_range.
contact_color: Binary. 0 if ball will contact red first, 1 if it
will contact green first.
"""
for step in range(valid_step_range[1]):
red_overlap = state['ball'][0].overlaps_sprite(state['red'][0])
green_overlap = state['ball'][0].overlaps_sprite(state['green'][0])
if red_overlap or green_overlap:
if step < valid_step_range[0]:
return False, None
else:
contact_color = 0 if red_overlap else 1
return True, contact_color
physics.step(state)
return False, None
############################################################################
# Sprite initialization
############################################################################
# Ball generator
ball_factors = distribs.Product(
[
distribs.Continuous('x', 0.15, 0.85),
distribs.Continuous('y', 0.15, 0.85),
RadialVelocity(speed=0.03)
],
scale=0.05,
shape='circle',
c0=64,
c1=64,
c2=255,
)
ball_generator = sprite_generators.generate_sprites(
ball_factors,
num_sprites=1,
max_recursion_depth=100,
fail_gracefully=True)
# Obstacle generator
obstacle_factors = distribs.Product(
[
distribs.Continuous('x', 0.2, 0.8),
distribs.Continuous('y', 0.2, 0.8)
],
scale=0.2,
shape='square',
c0=128,
c1=128,
c2=128,
)
obstacle_generator = sprite_generators.generate_sprites(
obstacle_factors,
num_sprites=2 + num_obstacles,
max_recursion_depth=100,
fail_gracefully=True)
# Walls
bottom_wall = [[-1, 0.1], [2, 0.1], [2, -1], [-1, -1]]
top_wall = [[-1, 0.95], [2, 0.95], [2, 2], [-1, 2]]
left_wall = [[0.05, -1], [0.05, 4], [-1, 4], [-1, -1]]
right_wall = [[0.95, -1], [0.95, 4], [2, 4], [2, -1]]
walls = [
sprite.Sprite(shape=np.array(v), x=0, y=0, c0=128, c1=128, c2=128)
for v in [bottom_wall, top_wall, left_wall, right_wall]
]
def state_initializer():
"""Callable returning new state at each episode reset."""
obstacles = obstacle_generator(disjoint=True)
ball = ball_generator(without_overlapping=obstacles)
if len(obstacles) < num_obstacles + 2 or not ball:
# Max recursion depth failed trying to generate without overlapping
return state_initializer()
red = obstacles[0]
green = obstacles[1]
obstacles = obstacles[2:]
# Set the colors of the red and green boxes
red.c0 = 255
red.c1 = 0
red.c2 = 0
green.c0 = 0
green.c1 = 255
green.c2 = 0
# Create agent and response tokens at the bottom of the sreen
agent = sprite.Sprite(x=0.5,
y=0.06,
shape='spoke_4',
scale=0.03,
c0=255,
c1=255,
c2=255)
responses = [
sprite.Sprite(x=0.6,
y=0.06,
shape='square',
scale=0.03,
c0=255,
c1=0,
c2=0),
sprite.Sprite(x=0.4,
y=0.06,
shape='square',
scale=0.03,
c0=0,
c1=255,
c2=0),
]
state = collections.OrderedDict([
('walls', walls + obstacles),
('red', [red]),
('green', [green]),
('ball', ball),
('responses', responses),
('agent', [agent]),
])
# Rejection sampling if trial won't finish in valid step range
original_ball_position = np.copy(ball[0].position)
original_ball_velocity = np.copy(ball[0].velocity)
valid_trial, contact_red = _predict_trial_end(state)
if valid_trial:
ball[0].position = original_ball_position
ball[0].velocity = original_ball_velocity
agent.metadata = {'true_contact_color': contact_red}
else:
return state_initializer()
return state
############################################################################
# Task
############################################################################
def _reward_fn(sprite_agent, sprite_response):
response_green = sprite_response.c0 < 128
if sprite_agent.metadata['true_contact_color'] == response_green:
return 1.
else:
return -1.
contact_reward = tasks.ContactReward(
reward_fn=_reward_fn,
layers_0='agent',
layers_1='responses',
reset_steps_after_contact=10,
)
task = tasks.CompositeTask(contact_reward, timeout_steps=400)
############################################################################
# Action space
############################################################################
action_space = action_spaces.Grid(
scaling_factor=0.015,
action_layers='agent',
control_velocity=True,
)
############################################################################
# Observer
############################################################################
observer = observers.PILRenderer(image_size=(64, 64), anti_aliasing=1)
############################################################################
# Game rules
############################################################################
# Stop ball on contact with red or green box
def _stop_ball(s):
s.velocity = np.zeros(2)
stop_ball = game_rules.ModifyOnContact(layers_0='ball',
layers_1=('red', 'green'),
modifier_0=_stop_ball)
############################################################################
# Final config
############################################################################
config = {
'state_initializer': state_initializer,
'physics': physics,
'task': task,
'action_space': action_space,
'observers': {
'image': observer
},
'game_rules': (stop_ball, ),
}
return config
def get_config(_):
"""Get environment config."""
############################################################################
# Sprite initialization
############################################################################
# Agent
agent = sprite.Sprite(x=0.5,
y=0.5,
shape='circle',
scale=0.04,
c0=0.33,
c1=1.,
c2=0.66)
annulus_vertices = shapes.annulus_vertices(inner_radius=0.08,
outer_radius=0.3)
agent_annulus = sprite.Sprite(x=0.5,
y=0.5,
shape=annulus_vertices,
scale=1.,
c0=0.6,
c1=1.,
c2=1.)
# Predator generator
max_predator_vel = 0.02
predator_pos = _get_boundary_pos_distribution(_FIELD_BUFFER)
predator_vel = _get_vel_distribution(0.5 * max_predator_vel,
max_predator_vel)
predator_factors = distribs.Product(
[predator_pos, predator_vel,
distribs.Continuous('scale', 0.07, 0.13)],
shape='circle',
c0=0.,
c1=1.,
c2=0.8,
)
# Prey generator
max_prey_vel = 0.01
prey_pos = _get_boundary_pos_distribution(_FIELD_BUFFER)
prey_vel = _get_vel_distribution(0.5 * max_prey_vel, max_prey_vel)
prey_factors = distribs.Product(
[prey_pos, prey_vel,
distribs.Continuous('scale', 0.07, 0.13)],
shape='circle',
c0=0.2,
c1=1.,
c2=1.,
)
# Grid
grid = shapes.grid_lines(grid_x=_GRID_SIZE,
grid_y=_GRID_SIZE,
buffer_border=1.,
c0=0.,
c1=0.,
c2=0.5)
def state_initializer():
state = collections.OrderedDict([
('grid', grid),
('prey', []),
('agent', [agent]),
('predators', []),
('agent_annulus', [agent_annulus]),
])
return state
############################################################################
# Physics
############################################################################
agent_friction_force = physics_lib.Drag(coeff_friction=0.25)
physics = physics_lib.Physics(
(agent_friction_force, ['agent', 'agent_annulus']),
updates_per_env_step=10,
)
############################################################################
# Task
############################################################################
def _predator_reward_fn(_, predator_sprite):
return -2. * predator_sprite.scale
predator_task = tasks.ContactReward(
reward_fn=_predator_reward_fn,
layers_0='agent',
layers_1='predators',
reset_steps_after_contact=0,
)
def _prey_reward_fn(_, prey_sprite):
return prey_sprite.scale
prey_task = tasks.ContactReward(
reward_fn=_prey_reward_fn,
layers_0='agent',
layers_1='prey',
)
task = tasks.CompositeTask(predator_task, prey_task)
############################################################################
# Action space
############################################################################
action_space = action_spaces.Joystick(
scaling_factor=0.003,
action_layers=('agent', 'agent_annulus'),
constrained_lr=False,
)
############################################################################
# Observer
############################################################################
_polygon_modifier = observers.polygon_modifiers.FirstPersonAgent(
agent_layer='agent')
observer = observers.PILRenderer(
image_size=(64, 64),
anti_aliasing=1,
color_to_rgb='hsv_to_rgb',
polygon_modifier=_polygon_modifier,
)
############################################################################
# Game rules
############################################################################
# Make predators appear randomly
predator_appear_generator = sprite_generators.generate_sprites(
predator_factors, num_sprites=1)
predator_appear = game_rules.ConditionalRule(
condition=lambda state: np.random.binomial(1, p=0.5),
rules=game_rules.CreateSprites('predators', predator_appear_generator),
)
# Make prey appear randomly
prey_appear_generator = sprite_generators.generate_sprites(prey_factors,
num_sprites=1)
prey_appear = game_rules.ConditionalRule(
condition=lambda state: np.random.binomial(1, p=0.2),
rules=game_rules.CreateSprites('prey', prey_appear_generator),
)
# Make predators and prey vanish when they are distant enough and moving
# away.
vanish_range = [-1. * _VANISH_DIST, 1. + _VANISH_DIST]
def _should_vanish(s):
pos_too_small = (s.position < vanish_range[0]) * (s.velocity < 0.)
pos_too_large = (s.position > vanish_range[1]) * (s.velocity > 0.)
return any(pos_too_small) or any(pos_too_large)
predator_vanish = game_rules.VanishByFilter('predators', _should_vanish)
prey_vanish = game_rules.VanishByFilter('prey', _should_vanish)
# Keep agent near center
keep_near_center = game_rules.KeepNearCenter(
agent_layer='agent',
layers_to_center=['agent_annulus', 'predators', 'prey'],
grid_x=_GRID_SIZE,
)
# Make prey vanish when contacted by agent
prey_caught = game_rules.VanishOnContact(vanishing_layer='prey',
contacting_layer='agent')
rules = (predator_appear, prey_appear, prey_vanish, predator_vanish,
keep_near_center, prey_caught)
############################################################################
# Final config
############################################################################
config = {
'state_initializer': state_initializer,
'physics': physics,
'task': task,
'action_space': action_space,
'observers': {
'image': observer
},
'game_rules': rules,
}
return config