def experiment(variant):
env = SawyerXYZEnv(**variant['env_kwargs'])
if variant['normalize']:
env = NormalizedBoxEnv(env)
obs_dim = env.observation_space.low.size
action_dim = env.action_space.low.size
goal_dim = env.goal_dim
qf = FlattenMlp(input_size=obs_dim + action_dim + goal_dim,
output_size=1,
**variant['qf_kwargs'])
vf = FlattenMlp(input_size=obs_dim + goal_dim,
output_size=1,
**variant['vf_kwargs'])
policy = TanhGaussianPolicy(obs_dim=obs_dim + goal_dim,
action_dim=action_dim,
**variant['policy_kwargs'])
replay_buffer = SimpleHerReplayBuffer(env=env,
**variant['replay_buffer_kwargs'])
algorithm = HerSac(env=env,
policy=policy,
qf=qf,
vf=vf,
replay_buffer=replay_buffer,
**variant['algo_kwargs'])
algorithm.to(ptu.device)
algorithm.train()
def experiment(variant):
env = SawyerXYZEnv(**variant['env_kwargs'])
env = MultitaskToFlatEnv(env)
if variant['normalize']:
env = NormalizedBoxEnv(env)
exploration_type = variant['exploration_type']
if exploration_type == 'ou':
es = OUStrategy(action_space=env.action_space)
elif exploration_type == 'gaussian':
es = GaussianStrategy(
action_space=env.action_space,
max_sigma=0.1,
min_sigma=0.1, # Constant sigma
)
elif exploration_type == 'epsilon':
es = EpsilonGreedy(
action_space=env.action_space,
prob_random_action=0.1,
)
else:
raise Exception("Invalid type: " + exploration_type)
obs_dim = env.observation_space.low.size
action_dim = env.action_space.low.size
qf1 = FlattenMlp(
input_size=obs_dim + action_dim,
output_size=1,
hidden_sizes=[400, 300],
)
qf2 = FlattenMlp(
input_size=obs_dim + action_dim,
output_size=1,
hidden_sizes=[400, 300],
)
policy = TanhMlpPolicy(
input_size=obs_dim,
output_size=action_dim,
hidden_sizes=[400, 300],
)
exploration_policy = PolicyWrappedWithExplorationStrategy(
exploration_strategy=es,
policy=policy,
)
algorithm = TD3(
env,
qf1=qf1,
qf2=qf2,
policy=policy,
exploration_policy=exploration_policy,
**variant['algo_kwargs']
)
algorithm.to(ptu.device)
algorithm.train()
def experiment(variant):
env = SawyerXYZEnv(**variant['env_kwargs'])
if variant['normalize']:
env = NormalizedBoxEnv(env)
exploration_type = variant['exploration_type']
if exploration_type == 'ou':
es = OUStrategy(action_space=env.action_space)
elif exploration_type == 'gaussian':
es = GaussianStrategy(
action_space=env.action_space,
max_sigma=0.1,
min_sigma=0.1, # Constant sigma
)
elif exploration_type == 'epsilon':
es = EpsilonGreedy(
action_space=env.action_space,
prob_random_action=0.1,
)
else:
raise Exception("Invalid type: " + exploration_type)
obs_dim = env.observation_space.low.size
action_dim = env.action_space.low.size
goal_dim = env.goal_dim
qf = FlattenMlp(
input_size=obs_dim + action_dim + goal_dim,
output_size=1,
hidden_sizes=[400, 300],
)
policy = TanhMlpPolicy(
input_size=obs_dim + goal_dim,
output_size=action_dim,
hidden_sizes=[400, 300],
)
exploration_policy = PolicyWrappedWithExplorationStrategy(
exploration_strategy=es,
policy=policy,
)
replay_buffer = SimpleHerReplayBuffer(
env=env,
**variant['replay_buffer_kwargs']
)
algorithm = HerDdpg(
env,
qf=qf,
policy=policy,
exploration_policy=exploration_policy,
replay_buffer=replay_buffer,
**variant['algo_kwargs']
)
algorithm.to(ptu.device)
algorithm.train()
def experiment(variant):
env = SawyerXYZEnv(**variant['env_kwargs'])
env = MultitaskToFlatEnv(env)
if variant['normalize']:
env = NormalizedBoxEnv(env)
obs_dim = env.observation_space.low.size
action_dim = env.action_space.low.size
qf = FlattenMlp(input_size=obs_dim + action_dim,
output_size=1,
**variant['qf_kwargs'])
vf = FlattenMlp(input_size=obs_dim, output_size=1, **variant['vf_kwargs'])
policy = TanhGaussianPolicy(obs_dim=obs_dim,
action_dim=action_dim,
**variant['policy_kwargs'])
algorithm = SoftActorCritic(env=env,
policy=policy,
qf=qf,
vf=vf,
**variant['algo_kwargs'])
algorithm.to(ptu.device)
algorithm.train()
from railrl.envs.mujoco.sawyer_gripper_env import SawyerPushXYEnv, \
SawyerPushEnv, SawyerXYZEnv
from railrl.envs.mujoco.sawyer_kitchen import KitchenCabinetEnv
from railrl.envs.wrappers import ImageMujocoEnv
from railrl.exploration_strategies.base import \
PolicyWrappedWithExplorationStrategy
from railrl.exploration_strategies.epsilon_greedy import EpsilonGreedy
from railrl.exploration_strategies.ou_strategy import OUStrategy
from railrl.policies.simple import ZeroPolicy
import numpy as np
print("making env")
# env = SawyerPushXYEnv(randomize_goals=True, frame_skip=50)
env = SawyerPushEnv(randomize_goals=False, frame_skip=50)
env = SawyerXYZEnv(frame_skip=50, pos_action_scale=2./100)
env = SawyerPushXYEnv(frame_skip=50, pos_action_scale=2./100)
# env = KitchenCabinetEnv()
from railrl.images.camera import sawyer_init_camera
# env = ImageMujocoEnv(
# env,
# init_camera=sawyer_init_camera,
# )
# env.enable_render()
policy = ZeroPolicy(env.action_space.low.size)
es = OUStrategy(
env.action_space,
theta=1
)
es = EpsilonGreedy(
def experiment(variant):
feat_points = 16
history = 1
latent_obs_dim = feat_points * 2 * history
imsize = 64
downsampled_size = 32
env = SawyerXYZEnv()
extra_fc_size = env.obs_dim
env = ImageMujocoWithObsEnv(env,
imsize=imsize,
normalize=True,
grayscale=True,
keep_prev=history - 1,
init_camera=camera.sawyer_init_camera)
"""env = ImageMujocoEnv(env,
imsize=imsize,
keep_prev=history-1,
init_camera=camera.sawyer_init_camera)"""
es = GaussianStrategy(action_space=env.action_space, )
obs_dim = env.observation_space.low.size
action_dim = env.action_space.low.size
ae = FeatPointMlp(input_size=imsize,
downsample_size=downsampled_size,
input_channels=1,
num_feat_points=feat_points)
replay_buffer = AEEnvReplayBuffer(int(1e4),
env,
imsize=imsize,
history_length=history,
downsampled_size=downsampled_size)
qf = FlattenMlp(input_size=latent_obs_dim + extra_fc_size + action_dim,
output_size=1,
hidden_sizes=[400, 300])
policy = AETanhPolicy(
input_size=latent_obs_dim + extra_fc_size,
ae=ae,
env=env,
history_length=history,
output_size=action_dim,
hidden_sizes=[400, 300],
)
exploration_policy = PolicyWrappedWithExplorationStrategy(
exploration_strategy=es,
policy=policy,
)
algorithm = FeatPointDDPG(ae,
history,
env=env,
qf=qf,
policy=policy,
exploration_policy=exploration_policy,
replay_buffer=replay_buffer,
extra_fc_size=extra_fc_size,
imsize=imsize,
downsampled_size=downsampled_size,
**variant['algo_params'])
algorithm.to(ptu.device)
algorithm.train()
from railrl.envs.mujoco.sawyer_gripper_env import SawyerXYZEnv
from railrl.envs.wrappers import ImageMujocoEnv
import cv2
import numpy as np
print("making env")
sawyer = SawyerXYZEnv()
env = ImageMujocoEnv(sawyer, imsize=400)
print("starting rollout")
while True:
obs = env.reset()
for t in range(1000):
action = env.action_space.sample()
obs, reward, done, info = env.step(action)
raw_img = env._image_observation()
img = np.concatenate((
raw_img[::-1, :, 2:3],
raw_img[::-1, :, 1:2],
raw_img[::-1, :, 0:1],
),
axis=2)
cv2.imshow('obs', img)
cv2.waitKey(1)
# if done:
# break
print("new episode")
from railrl.envs.mujoco.sawyer_gripper_env import SawyerXYZEnv
from railrl.exploration_strategies.base import \
PolicyWrappedWithExplorationStrategy
from railrl.exploration_strategies.ou_strategy import OUStrategy
from railrl.policies.simple import ZeroPolicy
print("making env")
env = SawyerXYZEnv()
policy = ZeroPolicy(env.action_space.low.size)
es = OUStrategy(env.action_space, theta=1)
policy = exploration_policy = PolicyWrappedWithExplorationStrategy(
exploration_strategy=es,
policy=policy,
)
print("starting rollout")
while True:
obs = env.reset()
last_reward_t = 0
returns = 0
for t in range(1000):
# action = env.action_space.sample()*10
action, _ = policy.get_action(None)
if (t // 100) % 2 == 0:
action[3] = -10
else:
action[3] = 10
obs, reward, done, info = env.step(action)
env.render()
print("action", action)
if done: