def __init__(self, *args, **kwargs):
self.random_start = kwargs.get("random_start", True)
MujocoEnv.__init__(
self,
os.path.join(os.path.abspath(os.path.dirname(__file__)), "assets",
"inverted_double_pendulum.xml"))
Serializable.quick_init(self, locals())
def __init__(self, env_spec, max_replay_buffer_size):
Serializable.quick_init(self, locals())
super(SimpleReplayBuffer, self).__init__(env_spec)
max_replay_buffer_size = int(max_replay_buffer_size)
# TODO: add support for changing buffer datatypes and attributes
self._env_spec = env_spec
self._observation_dim = env_spec.observation_space.flat_dim
self._action_dim = env_spec.action_space.flat_dim
self._max_buffer_size = max_replay_buffer_size
self._observations = np.zeros(
(max_replay_buffer_size, self._observation_dim), dtype='float32')
# It's a bit memory inefficient to save the observations twice,
# but it makes the code *much* easier since you no longer have to
# worry about termination conditions.
# TODO: add logic for termination condition to save memory
# TODO: at least convert skill to int or use sparse matrix to save mem
self._next_obs = np.zeros(
(max_replay_buffer_size, self._observation_dim), dtype='float32')
self._actions = np.zeros((max_replay_buffer_size, self._action_dim),
dtype='float32')
self._rewards = np.zeros(max_replay_buffer_size, dtype='float32')
# self._terminals[i] = a terminal was received at time i
self._dones = np.zeros(max_replay_buffer_size, dtype='uint8')
self._top = 0
self._size = 0
def __init__(self, observation_space, action_space):
"""
:type observation_space: Space
:type action_space: Space
"""
Serializable.quick_init(self, locals())
self._observation_space = observation_space
self._action_space = action_space
def __init__(
self,
alive_coeff=1,
ctrl_cost_coeff=0.01,
*args, **kwargs):
self.alive_coeff = alive_coeff
self.ctrl_cost_coeff = ctrl_cost_coeff
MujocoEnv.__init__(self, os.path.join(os.path.abspath(os.path.dirname(__file__)), "assets", "hopper.xml"))
Serializable.quick_init(self, locals())
def __init__(self,
env_name,
record_video=False,
video_schedule=None,
log_dir=None,
record_log=False,
force_reset=True,
seed=1):
# if log_dir is None:
# if logger.get_snapshot_dir() is None:
# logger.log("Warning: skipping Gym environment monitoring since snapshot_dir not configured.")
# else:
# log_dir = os.path.join(logger.get_snapshot_dir(), "gym_log")
Serializable.quick_init(self, locals())
env = gym.envs.make(env_name)
env.seed(seed)
# HACK: Gets rid of the TimeLimit wrapper that sets 'done = True' when
# the time limit specified for each environment has been passed and
# therefore the environment is not Markovian (terminal condition depends
# on time rather than state).
env = env.env
self.env = env
self.env_id = env.spec.id
assert not (not record_log and record_video)
if log_dir is None or record_log is False:
self.monitoring = False
else:
if not record_video:
video_schedule = NoVideoSchedule()
else:
if video_schedule is None:
video_schedule = CappedCubicVideoSchedule()
self.env = gym.wrappers.Monitor(self.env,
log_dir,
video_callable=video_schedule,
force=True)
self.monitoring = True
self._observation_space = convert_gym_space(env.observation_space)
# logger.log("observation space: {}".format(self._observation_space))
self._action_space = convert_gym_space(env.action_space)
# logger.log("action space: {}".format(self._action_space))
self._horizon = env.spec.tags[
'wrapper_config.TimeLimit.max_episode_steps']
self._log_dir = log_dir
self._force_reset = force_reset
def __init__(self,
vel_deviation_cost_coeff=1e-2,
alive_bonus=0.2,
ctrl_cost_coeff=1e-3,
impact_cost_coeff=1e-5,
*args,
**kwargs):
self.vel_deviation_cost_coeff = vel_deviation_cost_coeff
self.alive_bonus = alive_bonus
self.ctrl_cost_coeff = ctrl_cost_coeff
self.impact_cost_coeff = impact_cost_coeff
MujocoEnv.__init__(
self,
os.path.join(os.path.abspath(os.path.dirname(__file__)), "assets",
"simple_humanoid.xml"))
Serializable.quick_init(self, locals())
def __init__(self, env, reward_fn):
Serializable.quick_init(self, locals())
self._wrapped_env = env
self.reward_function = reward_fn
self.num_skills = num_skills
self.z = skill
obs_space = self.observation_space
low = np.hstack([obs_space.low, np.full(num_skills, 0)])
high = np.hstack([obs_space.high, np.full(num_skills, 1)])
self.observation_space = spaces.Box(low=low, high=high)
# TODO verfiy below line is doing something useful
self.action_space = self.action_space
print("Veriifying action space: ", self.action_space)
self.spec = EnvSpec(self.observation_space, self.action_space)
self.reset(state=None, skill=self.z)
def __init__(
self,
env,
scale_reward=1.,
normalize_obs=False,
normalize_reward=False,
obs_alpha=0.001,
reward_alpha=0.001,
normalization_scale=1.,
):
Serializable.quick_init(self, locals())
self._scale_reward = 1
self._wrapped_env = env
self._normalize_obs = normalize_obs
self._normalize_reward = normalize_reward
self._obs_alpha = obs_alpha
self._obs_mean = np.zeros(self.observation_space.shape)
self._obs_var = np.ones(self.observation_space.shape)
self._reward_alpha = reward_alpha
self._reward_mean = 0.
self._reward_var = 1.
self._normalization_scale = normalization_scale