def main():
model_class = DDPG # works also with SAC and DDPG
# -j
action_space = 7
# -p
fixed = True
# -o
normalize_observations = False
# -g
gamma = 0.9
# -b
#batch_size = 16
# -m
memory_limit = 1000000
# -r
normalize_returns = True
# -t
timesteps = 1000000
policy_name = "pushing_policy"
discreteAction = 0
rend = False
env = pandaPushGymEnvHERRand(urdfRoot=robot_data.getDataPath(),
renders=rend,
useIK=0,
isDiscrete=discreteAction,
action_space=action_space,
fixedPositionObj=fixed,
includeVelObs=True)
# Available strategies (cf paper): future, final, episode, random
goal_selection_strategy = 'future' # equivalent to GoalSelectionStrategy.FUTURE
n_actions = env.action_space.shape[-1]
action_noise = OrnsteinUhlenbeckActionNoise(mean=np.zeros(n_actions),
sigma=float(0.5) *
np.ones(n_actions))
# Wrap the model
model = HER(
CustomPolicy,
env,
model_class,
n_sampled_goal=4,
goal_selection_strategy=goal_selection_strategy,
verbose=1,
tensorboard_log=
"../pybullet_logs/panda_push_ddpg/stable_baselines/DDPG+HER_FIXED_DYN_RAND",
buffer_size=1000000,
batch_size=256,
random_exploration=0.3,
action_noise=action_noise)
# Train the model starting from a previous policy
model.learn(timesteps)
print("Saving Policy")
model.save("../policies/pushing_fixed_HER_Dyn_Rand")
def main(load_policy=False):
global log_dir, log_dir_policy
if (load_policy):
log_dir_policy = '../policies/PUSHING_TD3+HER_FIXED_POSITION_DYN_RAND_FROM_FIXED_PHYSICS'
model_class = TD3 # works also with SAC and DDPG
action_space = 7
fixed = True
normalize_observations = False
gamma = 0.9
memory_limit = 1000000
normalize_returns = True
timesteps = 1500000
discreteAction = 0
rend = False
env = pandaPushGymEnvHERRand(urdfRoot=robot_data.getDataPath(), renders=rend, useIK=0,
isDiscrete=discreteAction, action_space = action_space,
fixedPositionObj = fixed, includeVelObs = True)
env = Monitor(env, log_dir, allow_early_resets=True)
# Available strategies (cf paper): future, final, episode, random
goal_selection_strategy = 'future' # equivalent to GoalSelectionStrategy.FUTURE
n_actions = env.action_space.shape[-1]
action_noise = OrnsteinUhlenbeckActionNoise(mean=np.zeros(n_actions), sigma=float(0.5) * np.ones(n_actions))
# Wrap the model
model = HER(CustomPolicy, env, model_class, n_sampled_goal=4, goal_selection_strategy=goal_selection_strategy,
verbose=1,tensorboard_log="../pybullet_logs/panda_push_TD3/stable_baselines/TD3+HER_FIXED_DYN_RAND", buffer_size=1000000,batch_size=256,
random_exploration=0.3, action_noise=action_noise)
if (load_policy):
model = HER.load("../policies/USEFUL_POLICIES/PUSHING_TD3+HER_FIXED_POSITIONbest_model.pkl", env=env, n_sampled_goal=4,
goal_selection_strategy=goal_selection_strategy,
tensorboard_log="../pybullet_logs/panda_push_TD3/stable_baselines/TD3+HER_FIXED_DYN_RAND_FROM_FIXED_PHYSICS",
buffer_size=1000000,batch_size=256,random_exploration=0.3, action_noise=action_noise)
# Train the model starting from a previous policy
model.learn(timesteps, callback = callback )
model.save("../policies/PUSHING_FIXED_TD3_DYN_RAND")
print("Finished train1")
# -b
batch_size = 16
# -m
memory_limit = 1000000
# -r
normalize_returns = True
# -t
timesteps = 10000000
policy_name = "pushing_policy"
discreteAction = 0
rend = True
env = pandaPushGymEnvHERRand(urdfRoot=robot_data.getDataPath(),
renders=rend,
useIK=0,
isDiscrete=discreteAction,
numControlledJoints=numControlledJoints,
fixedPositionObj=fixed,
includeVelObs=True)
# Wrap the model
model = HER.load("../policies/pushing_fixed_HER_Dyn_Rand0.pkl", env=env)
obs = env.reset()
for i in range(10000):
action, _ = model.predict(obs)
obs, reward, done, _ = env.step(action)
if done or i == 500:
obs = env.reset()
normalize_observations = False
# -g
gamma = 0.9
# -b
batch_size = 16
# -m
memory_limit = 1000000
# -r
normalize_returns = True
# -t
timesteps = 1000000
policy_name = "pushing_policy"
discreteAction = 0
rend = True
env = pandaPushGymEnvHERRand(urdfRoot=robot_data.getDataPath(), renders=rend, useIK=0,
isDiscrete=discreteAction, action_space = action_space,
fixedPositionObj = fixed, includeVelObs = True, object_position=0, test_phase = True, alg = 'td3_normal_policy_to_different_physics', type_physics=2, max_episode_steps=500)
goal_selection_strategy = 'future' # equivalent to GoalSelectionStrategy.FUTURE
# Wrap the model
model = HER.load("../policies/USEFUL_POLICIES/PUSHING_TD3+HER_FIXED_POSITIONbest_model.pkl", env=env)
obs = env.reset()
for _ in range(10000):
action, _ = model.predict(obs)
obs, reward, done, _ = env.step(action)
if done:
obs = env.reset()