def __init__(
self,
render=True,
init_noise=True,
act_noise=False,
obs_noise=False,
control_skip=10,
max_tar_vel=2.5,
energy_weight=0.1,
jl_weight=0.5,
ab=5.0,
q_pen_weight=0.4,
dq_pen_weight=0.001,
vel_r_weight=4.0,
train_dyn=True, # if false, fix dyn and train motor policy
pretrain_dyn=False, # pre-train with deviation to sim
enlarge_act_range=0.25, # make behavior pi more diverse to match collection
behavior_dir="trained_models_laika_bullet_48/ppo",
behavior_env_name="LaikagoBulletEnv-v3",
behavior_iter=None,
dyn_dir="",
dyn_env_name="LaikagoConFEnv-v3", # itself
dyn_iter=None,
cuda_env=True,
):
self.render = render
self.init_noise = init_noise
self.obs_noise = obs_noise
self.act_noise = act_noise
self.control_skip = int(control_skip)
self._ts = 1. / 500.
self.max_tar_vel = max_tar_vel
self.energy_weight = energy_weight
self.jl_weight = jl_weight
self.ab = ab
self.q_pen_weight = q_pen_weight
self.dq_pen_weight = dq_pen_weight
self.vel_r_weight = vel_r_weight
self.train_dyn = train_dyn
self.enlarge_act_range = enlarge_act_range
self.pretrain_dyn = pretrain_dyn
self.cuda_env = cuda_env
if self.render:
self._p = bullet_client.BulletClient(connection_mode=pybullet.GUI)
else:
self._p = bullet_client.BulletClient()
self.np_random = None
self.robot = LaikagoBulletV2(init_noise=self.init_noise,
time_step=self._ts,
np_random=self.np_random)
self.seed(
0
) # used once temporarily, will be overwritten outside though superclass api
self.viewer = None
self.timer = 0
self.feat_select_func = LaikagoBulletEnvV3.feature_selection_G2BD_laika_v3
if self.train_dyn:
if behavior_iter:
behavior_iter = int(behavior_iter)
self.dyn_actor_critic = None
# load fixed behavior policy
self.go_actor_critic, _, \
self.recurrent_hidden_states, \
self.masks = utils.load(
behavior_dir, behavior_env_name, self.cuda_env, behavior_iter
)
else:
if dyn_iter:
dyn_iter = int(dyn_iter)
# train motor policy
self.go_actor_critic = None
# load fixed dynamics model
self.dyn_actor_critic, _, \
self.recurrent_hidden_states, \
self.masks = utils.load(
dyn_dir, dyn_env_name, self.cuda_env, dyn_iter
)
#
# self.discri = utils.load_gail_discriminator(dyn_dir,
# dyn_env_name,
# self.cuda_env,
# dyn_iter)
#
# self.feat_select_func = self.robot.feature_selection_all_laika
self.reset_const = 100
self.reset_counter = self.reset_const # do a hard reset first
self.obs = []
self.behavior_act_len = None
self.init_state = None
# self.obs here is for G which is longer and contains dqs
self.obs_dim = 0 # dummy
self.reset() # and update init obs & self.obs_dim
#
# self.d_scores = []
# # set up imaginary session for pre-train
# self.set_up_imaginary_session()
if self.train_dyn:
assert self.behavior_act_len == len(self.robot.ctrl_dofs) # (12)
self.action_dim = 12 # 12D action scales, see beginning of step() for comment
else:
self.action_dim = len(self.robot.ctrl_dofs)
self.act = [0.0] * self.action_dim
self.action_space = gym.spaces.Box(
low=np.array([-1.] * self.action_dim),
high=np.array([+1.] * self.action_dim))
obs_dummy = np.array([1.12234567] * self.obs_dim)
# print(self.obs_dim)
self.observation_space = gym.spaces.Box(low=-np.inf * obs_dummy,
high=np.inf * obs_dummy)
def __init__(
self,
render=True,
init_noise=True,
act_noise=True,
obs_noise=True,
control_skip=10,
using_torque_ctrl=True,
max_tar_vel=2.5,
energy_weight=0.1,
jl_weight=0.5,
ab=5.0,
q_pen_weight=0.5,
dq_pen_weight=0.02,
vel_r_weight=4.0,
train_dyn=True, # if false, fix dyn and train motor policy
pretrain_dyn=False, # pre-train with deviation to sim
enlarge_act_range=True, # make behavior pi more diverse to match collection
behavior_dir="trained_models_laika_bullet_12/ppo",
behavior_env_name="LaikagoBulletEnv-v1",
behavior_iter=None,
dyn_dir="",
dyn_env_name="LaikagoConFEnv-v1",
dyn_iter=None,
cuda_env=True,
#
# soft_floor_env=False, # for collecting data
):
self.render = render
self.init_noise = init_noise
self.obs_noise = obs_noise
self.act_noise = act_noise
self.control_skip = int(control_skip)
self._ts = 1. / 500.
self.max_tar_vel = max_tar_vel
self.energy_weight = energy_weight
self.jl_weight = jl_weight
self.ab = ab
self.q_pen_weight = q_pen_weight
self.dq_pen_weight = dq_pen_weight
self.vel_r_weight = vel_r_weight
self.train_dyn = train_dyn
self.enlarge_act_range = enlarge_act_range
self.pretrain_dyn = pretrain_dyn
self.cuda_env = cuda_env
# self.soft_floor_env = soft_floor_env
if self.render:
self._p = bullet_client.BulletClient(connection_mode=pybullet.GUI)
else:
self._p = bullet_client.BulletClient()
self.np_random = None
self.robot = LaikagoBullet(init_noise=self.init_noise,
time_step=self._ts,
np_random=self.np_random)
self.seed(
0
) # used once temporarily, will be overwritten outside though superclass api
self.viewer = None
self.timer = 0
if self.train_dyn:
if behavior_iter:
behavior_iter = int(behavior_iter)
self.dyn_actor_critic = None
# load fixed behavior policy
self.go_actor_critic, _, \
self.recurrent_hidden_states, \
self.masks = utils.load(
behavior_dir, behavior_env_name, self.cuda_env, behavior_iter
)
else:
if dyn_iter:
dyn_iter = int(dyn_iter)
# train motor policy
self.go_actor_critic = None
# load fixed dynamics model
self.dyn_actor_critic, _, \
self.recurrent_hidden_states, \
self.masks = utils.load(
dyn_dir, dyn_env_name, self.cuda_env, dyn_iter
)
self.reset_counter = 50 # do a hard reset first
self.obs = []
self.behavior_obs_len = None
self.behavior_act_len = None
self.init_state = None
self.reset() # and update init obs
# set up imaginary session for pre-train
self.set_up_imaginary_session()
if self.train_dyn:
assert self.behavior_act_len == len(self.robot.ctrl_dofs)
self.action_dim = 12 # see beginning of step() for comment
self.obs_dim = self.behavior_act_len + self.behavior_obs_len # see beginning of update_obs() for comment
else:
self.action_dim = len(self.robot.ctrl_dofs)
self.obs_dim = len(self.obs)
self.act = [0.0] * self.action_dim
self.action_space = gym.spaces.Box(
low=np.array([-1.] * self.action_dim),
high=np.array([+1.] * self.action_dim))
obs_dummy = np.array([1.12234567] * self.obs_dim)
self.observation_space = gym.spaces.Box(low=-np.inf * obs_dummy,
high=np.inf * obs_dummy)