def __init__(
self,
max_speed=0.05,
max_distance=1,
use_low_gear_ratio=True,
speed_weight=0.9,
done_threshold=0.005,
goal_dim_weights=None,
):
Serializable.quick_init(self, locals())
self.max_distance = max_distance
self.max_speed = max_speed
self.speed_weight = speed_weight
self.done_threshold = done_threshold
self.initializing = True
# TODO: fix this hack
if speed_weight is None:
self.speed_weight = 0.9 # just for init to work
MultitaskEnv.__init__(self, goal_dim_weights=goal_dim_weights)
super().__init__(use_low_gear_ratio=use_low_gear_ratio)
self.set_goal(
np.array([
self.max_distance,
self.max_distance,
self.max_speed,
self.max_speed,
]))
self.initializing = False
if speed_weight is None:
assert (self.goal_dim_weights[0] == self.goal_dim_weights[1]) and (
self.goal_dim_weights[2] == self.goal_dim_weights[3])
self.speed_weight = self.goal_dim_weights[2]
assert 0 <= self.speed_weight <= 1
def __init__(
self,
env: MultitaskEnv,
give_goal_difference=False,
):
# self._wrapped_env needs to be called first because
# Serializable.quick_init calls getattr, on this class. And the
# implementation of getattr (see below) calls self._wrapped_env.
# Without setting this first, the call to self._wrapped_env would call
# getattr again (since it's not set yet) and therefore loop forever.
self._wrapped_env = env
# Or else serialization gets delegated to the wrapped_env. Serialize
# this env separately from the wrapped_env.
self._serializable_initialized = False
self._wrapped_obs_dim = env.observation_space.low.size
self.give_goal_difference = give_goal_difference
Serializable.quick_init(self, locals())
ProxyEnv.__init__(self, env)
wrapped_low = self.observation_space.low
low = np.hstack(
(wrapped_low,
min(wrapped_low) * np.ones(self._wrapped_env.goal_dim)))
wrapped_high = self.observation_space.low
high = np.hstack(
(wrapped_high,
max(wrapped_high) * np.ones(self._wrapped_env.goal_dim)))
self.observation_space = Box(low, high)
def __init__(
self,
env,
reward_scale=1.,
obs_mean=None,
obs_std=None,
):
# self._wrapped_env needs to be called first because
# Serializable.quick_init calls getattr, on this class. And the
# implementation of getattr (see below) calls self._wrapped_env.
# Without setting this first, the call to self._wrapped_env would call
# getattr again (since it's not set yet) and therefore loop forever.
self._wrapped_env = env
# Or else serialization gets delegated to the wrapped_env. Serialize
# this env separately from the wrapped_env.
self._serializable_initialized = False
Serializable.quick_init(self, locals())
ProxyEnv.__init__(self, env)
self._should_normalize = not (obs_mean is None and obs_std is None)
if self._should_normalize:
if obs_mean is None:
obs_mean = np.zeros_like(env.observation_space.low)
else:
obs_mean = np.array(obs_mean)
if obs_std is None:
obs_std = np.ones_like(env.observation_space.low)
else:
obs_std = np.array(obs_std)
self._reward_scale = reward_scale
self._obs_mean = obs_mean
self._obs_std = obs_std
ub = np.ones(self._wrapped_env.action_space.shape)
self.action_space = Box(-1 * ub, ub)
def __init__(
self,
action_space,
mu=0,
theta=0.15,
max_sigma=0.3,
min_sigma=None,
decay_period=100000,
):
Serializable.quick_init(self, locals())
if min_sigma is None:
min_sigma = max_sigma
self.mu = mu
self.theta = theta
self.sigma = max_sigma
self._max_sigma = max_sigma
if min_sigma is None:
min_sigma = max_sigma
self._min_sigma = min_sigma
self._decay_period = decay_period
self.dim = np.prod(action_space.low.shape)
self.low = action_space.low
self.high = action_space.high
self.state = np.ones(self.dim) * self.mu
self.reset()
def __init__(self, action_space, max_sigma=1.0, min_sigma=None,
decay_period=1000000):
assert len(action_space.shape) == 1
Serializable.quick_init(self, locals())
self._max_sigma = max_sigma
if min_sigma is None:
min_sigma = max_sigma
self._min_sigma = min_sigma
self._decay_period = decay_period
self._action_space = action_space
def __init__(self,
min_distance=0,
max_distance=2,
use_low_gear_ratio=True):
Serializable.quick_init(self, locals())
self.max_distance = max_distance
self.min_distance = min_distance
MultitaskEnv.__init__(self)
super().__init__(use_low_gear_ratio=use_low_gear_ratio)
self.set_goal(np.array([self.max_distance, self.max_distance]))
def __init__(self, distance_metric_order=None, goal_dim_weights=None):
self._desired_xyz = np.zeros(3)
Serializable.quick_init(self, locals())
MultitaskEnv.__init__(
self,
distance_metric_order=distance_metric_order,
goal_dim_weights=goal_dim_weights,
)
mujoco_env.MujocoEnv.__init__(
self,
get_asset_xml('reacher_7dof.xml'),
5,
)
self.observation_space = Box(
np.array([
-2.28,
-0.52,
-1.4,
-2.32,
-1.5,
-1.094,
-1.5, # joint
-3,
-3,
-3,
-3,
-3,
-3,
-3, # velocity
-0.75,
-1.25,
-0.2, # EE xyz
]),
np.array([
1.71,
1.39,
1.7,
0,
1.5,
0,
1.5, # joints
3,
3,
3,
3,
3,
3,
3, # velocity
0.75,
0.25,
0.6, # EE xyz
]))
def __init__(self, wrapped_env):
Serializable.quick_init(self, locals())
self._wrapped_env = wrapped_env
self.action_space = self._wrapped_env.action_space
self.observation_space = self._wrapped_env.observation_space
def __init__(self):
Serializable.quick_init(self, locals())
self.target_x_vel = np.random.uniform(-MAX_SPEED, MAX_SPEED)
super().__init__()
MultitaskEnv.__init__(self)
self.set_goal(np.array([5]))
def init_serialization(self, locals):
Serializable.quick_init(self, locals)
def __init__(self, action_space, prob_random_action=0.1):
Serializable.quick_init(self, locals())
assert isinstance(action_space, Discrete)
Serializable.quick_init(self, locals())
self.prob_random_action = prob_random_action
self.action_space = action_space