def main():
env = KukaDiverseObjectEnv(renders=True, isDiscrete=False)
policy = ContinuousDownwardBiasPolicy()
while True:
obs, done = env.reset(), False
print("===================================")
print("obs")
print(obs)
episode_rew = 0
while not done:
env.render(mode='human')
act = policy.sample_action(obs, .1)
print("Action")
print(act)
obs, rew, done, _ = env.step([0, 0, 0, 0, 0])
episode_rew += rew
print("Episode reward", episode_rew)
#####################
# TENSORBOARD SETTINGS
current_time = datetime.datetime.now().strftime("%Y%m%d-%H%M%S")
train_log_dir = 'logs/train/' + current_time
graph_log_dir = 'logs/func/' + current_time
train_summary_writer = tf.summary.create_file_writer(train_log_dir)
with tf.Graph().as_default():
print(tf.executing_eagerly())
######################
# start open/AI GYM environment
#env = KukaCamGymEnv(renders=False, isDiscrete=False)
env = KukaDiverseObjectEnv(renders=False,
isDiscrete=False,
maxSteps=20,
removeHeightHack=False)
print('shape of Observation space: ', env.observation_space.shape)
print('shape of Action space: ', env.action_space.shape)
print('Reward Range: ', env.reward_range)
print('Action High value: ', env.action_space.high)
print('Action Low Value: ', env.action_space.low)
################
# Hyper-parameters
######################
MAX_EPISODES = 15001
LR_A = 0.001
LR_C = 0.002
GAMMA = 0.99
from PIL import Image from tensorboardX import SummaryWriter # 可视化训练数据 # 导入基本的环境 from pybullet_envs.bullet.kuka_diverse_object_gym_env import KukaDiverseObjectEnv from gym import spaces import pybullet as p #from pybullet_envs.bullet.kukaCamGymEnv import KukaCamGymEnv #只有一个物体的抓取 # env1 = KukaCamGymEnv(renders=True) # env1.render(mode='human') # # env1.cid = p.connect(p.DIRECT) #上面是true之后显示的就已经是GUI版本了 #这里需要加入渲染的部分 env = KukaDiverseObjectEnv(renders=True, isDiscrete=False, removeHeightHack=False, maxSteps=20, numObjects=1) # 渲染True,默认的就是GUI，但是还不显示机器人 # maxsteps是一个episode里面可执行的最多的动作次数 # removeHeightHack 如果是False，每次机器人自动的做下降的动作，否则要从头开始学习 # cameraRandom 可以随机的放置相机的位置，0是位置确定，1是完全随机，大部分的多状态学习都是通过不同的相机位置 // 空间感知？ # width，Height 感知到的相机的画幅 # numObjects 盘子里面物体的数量 # isTest False用于训练，test则是用于测试的数据组 env.cid = p.connect(p.DIRECT) action_space = spaces.Box(low=-1, high=1, shape=(5,1)) #这里插入学习网络的构造 ############## # Actor-Critic implementation import torch import torch.nn as nn import torch.nn.functional as F
from pybullet_envs.bullet.kuka_diverse_object_gym_env import KukaDiverseObjectEnv
from gym import spaces
from itertools import count
import matplotlib.pyplot as plt
env = KukaDiverseObjectEnv(renders=True, # True to see the simulation environment
isDiscrete=False,
removeHeightHack=False,
maxSteps=20)
print('shape of Observation space: ', env.observation_space.shape)
print('shape of Action space: ', env.action_space.shape)
print('Reward Range: ', env.reward_range)
print('Action High value: ', env.action_space.high)
print('Action Low Value: ', env.action_space.low)
for episode in range(30):
print('Episode: ', episode)
obsv = env.reset()
for t in count():
#plt.imshow(obsv)
#plt.show()
#env.render()
action = env.action_space.sample()
print("action=", action)
next_obs, reward, done, info = env.step(action)
print("reward", reward)
if done:
break
obs = next_obs
#print('t =', t)
BATCH_SIZE = 32
REPLAY_SIZE = 10000
LEARNING_RATE = 1e-4
SYNC_TARGET_FRAMES = 500 #150#1000
REPLAY_START_SIZE = BATCH_SIZE
EPSILON_DECAY_LAST_FRAME = 10**4 #10**5
EPSILON_START = 0.9 #1.0
EPSILON_FINAL = 0.1
MODEL_PATH = './models/kuka'
MODEL_NAME = 'kuka_policy_net.pt'
device = 'cuda'
env = KukaDiverseObjectEnv(renders=False,
isDiscrete=True,
removeHeightHack=False,
maxSteps=20)
env.cid = p.connect(p.DIRECT)
env.reset()
## get the camera image from the pybullet environment
## (this is an observation of the enviroment)
def get_observation():
obs = env._get_observation() #.transpose(2, 0, 1)
#print(obs.shape)
return obs
plt.imshow(get_observation())
TB_LOG = False # enable / disable tensorboard logging
if TB_LOG:
current_time = datetime.datetime.now().strftime("%Y%m%d-%H%M%S")
train_log_dir = '../logs/train/' + current_time
train_summary_writer = tf.summary.create_file_writer(train_log_dir)
############################
# for reproducibility
tf.random.set_seed(20)
np.random.seed(20)
#####################
# start open/AI GYM environment
env = KukaDiverseObjectEnv(
renders=False, # True for testing
isDiscrete=False,
maxSteps=20,
removeHeightHack=False)
print('shape of Observation space: ', env.observation_space.shape)
print('shape of Action space: ', env.action_space.shape)
print('Reward Range: ', env.reward_range)
print('Action High value: ', env.action_space.high)
print('Action Low Value: ', env.action_space.low)
############################
upper_bound = env.action_space.high
state_size = env.observation_space.shape # (48, 48, 3)
action_size = env.action_space.shape # (3,)
print('state_size: ', state_size)
print('action_size: ', action_size)
