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()
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)
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)
import os
import psutil
import pybullet
from pybullet_envs.bullet.kuka_diverse_object_gym_env import KukaDiverseObjectEnv # NOQA
env = KukaDiverseObjectEnv()
process = psutil.Process(os.getpid())
pybullet.setPhysicsEngineParameter(enableFileCaching=0)
for i in range(50):
env.reset()
print(i, process.memory_info().rss)
load_param = pickle.load(file)
start_episode = load_param[0]
ep_reward_list = load_param[1]
avg_reward_list = load_param[2]
else:
start_episode = 0
ep_reward_list = []
avg_reward_list = []
actor_loss, critic_loss = 0, 0
best_score = -np.inf
print('Main training loop')
for episode in range(start_episode, MAX_EPISODES):
obsv = env.reset()
state = np.asarray(
obsv, dtype=np.float32) / 255.0 # convert into float array
episodic_reward = 0
frames = []
steps = 0
while True:
if episode > MAX_EPISODES - 3:
frames.append(env.render(mode='rgb_array'))
# take an action as per the policy
if episode < RAND_EPS: # explore for some episodes
action = env.action_space.sample()
else:
action = agent.policy(state)
screen = env._get_observation().transpose((2, 0, 1)) # 将地二轴和第零轴的数据进行转置
#[screen, depth, segement] =
# Convert to float, rescale, convert to torch tensor
# (this doesn't require a copy)
screen = np.ascontiguousarray(screen, dtype=np.float32) / 255 #返回的类型要是float类型
screen = torch.from_numpy(screen) #从numpy获取数据转换成torch用的sensor
# Resize, and add a batch dimension (BCHW)
return resize(screen).unsqueeze(0).to(device) #返回的图像返回到torch里面去
##############
# 训练的过程
env.reset() #环境需要重置
num_agents = 1 #什么意思？
print('Number of agents:', num_agents)
init_screen = get_screen()
_, _, screen_height, screen_width = init_screen.shape #确定屏幕的大小
action_size = env.action_space.shape[0] #动作空间的大小，三个方向吗？由env决定的
print('Size of each action:', action_size)
#展示第一张模拟相机图片 // matplotlib里面的调用和matlab里面相似
plt.figure()
plt.imshow(init_screen.cpu().squeeze(0).permute(1, 2, 0).numpy(),
interpolation='none')
plt.title('Example extracted screen')
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())
#plt.show()