def main():
'''
Main thread:
'''
train = True
environment = FetchReachEnv()
param_file = "./her_bit_env"
ddpg_loop(train=train, environment=environment)
def build_env(self, env_type, reward_type):
if env_type == 'FetchReach-v1':
return FetchReachEnv(reward_type=reward_type)
elif env_type == 'FetchPush-v1':
return FetchPushEnv(reward_type=reward_type)
elif env_type == 'FetchPickAndPlace-v1':
return FetchPickAndPlaceEnv(reward_type=reward_type)
else:
raise NotImplementedError
from gym.envs.robotics.fetch.push import FetchPushEnv
from gym.envs.robotics.fetch.reach import FetchReachEnv
from gym.envs.robotics import fetch_env
import argparse
import collections
import sklearn.pipeline
import sklearn.preprocessing
from sklearn.kernel_approximation import RBFSampler
tf.enable_eager_execution()
# env = gym.make('FetchReach-v1')
# env.reward_type = 'dense'
# env.target_offset=1.0
# _max_episode_steps = 100
env = FetchReachEnv(reward_type='dense')
env.max_episode_steps = 100
observation_examples = np.array(
[env.reset()['observation'] for _ in range(10000)])
scaler = sklearn.preprocessing.StandardScaler()
scaler.fit(observation_examples)
# Used to converte a state to a featurizes represenation.
# We use RBF kernels with different variances to cover different parts of the space
featurizer = sklearn.pipeline.FeatureUnion([
("rbf1", RBFSampler(gamma=5.0, n_components=100)),
("rbf2", RBFSampler(gamma=2.0, n_components=100)),
("rbf3", RBFSampler(gamma=1.0, n_components=100)),
("rbf4", RBFSampler(gamma=0.5, n_components=100))
])
def env_fn():
env = FetchReachEnv(reward_type='dense')
return EnvWrapper(env)
import sys
sys.path.insert(0, "../../gym")
sys.path.insert(0, "../../gym/envs/robotics")
import gym
from gym.envs.robotics.fetch.reach import FetchReachEnv
#from fetch.reach import FetchReachEnv
print(gym)
#env = FetchPickAndPlaceEnv()
env = FetchReachEnv()
env.reset()
#for _ in range(10000):
env.render()
#print(f"action:{env.action_space.sample()}")
# action = env.action_space.sample() # take a random action
# observation, reward, done, info = env.step(action)
#env.close()
import sys sys.path.insert(0,"../../gym") sys.path.insert(0,"../../gym/envs/robotics") import gym from stable_baselines.common.env_checker import check_env from gym.envs.robotics.fetch.reach import FetchReachEnv # It will check your custom environment and output additional warnings if needed print(gym) env=FetchReachEnv() #env=gym.make() print(isinstance(env, gym.GoalEnv)) check_env(env)
action='store_true',
help="Render environment while training")
parser.add_argument('--nb_episodes',
type=int,
default=5000,
help="Number of training episodes")
parser.add_argument('--eval_model', type=str, default='')
parser.add_argument('--eval_random',
action='store_true',
help='evaluate random performance on env')
args = parser.parse_args()
consec_frames = 4
# Continuous Environments Wrapp
# er
# env = gym.make('FetchReach-v1')
env = FetchReachEnv(reward_type='sparse')
env.reset()
state_dim = env.reset()['observation'].shape
action_space = FetchReachEnv().action_space
action_dim = action_space.high.shape[0]
act_range = action_space.high
algo = DDPG(action_dim, state_dim, act_range, consec_frames)
if args.eval_model != '':
print(algo.eval(env, model_name=args.eval_model, render=args.render))
elif args.eval_random:
print(algo.eval(env, random=True, render=args.render))
else:
algo.train(env, args)