-
Notifications
You must be signed in to change notification settings - Fork 0
/
CPG_DDPG.py
261 lines (216 loc) · 8.69 KB
/
CPG_DDPG.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
import tensorflow as tf
import numpy as np
import os
import shutil
import vrep
import time
import math
import matplotlib.pyplot as plt
import random
from CPG_single import oscillator_nw
from DDPG import DDPG
import csv
clientID=vrep.simxStart('127.0.0.1',19997,True,True,-500000,5)
res, HeadHandle = vrep.simxGetObjectHandle(clientID, "HeadYaw", vrep.simx_opmode_blocking)
res, NAOHandle = vrep.simxGetObjectHandle(clientID, "NAO", vrep.simx_opmode_blocking)
res, NAO_Pos0 = vrep.simxGetObjectPosition(clientID,NAOHandle,-1,vrep.simx_opmode_blocking)
x0 = NAO_Pos0[0]
y0 = NAO_Pos0[1]
max_x = float('-inf')
def getState(dx):
res, NAO_Pos = vrep.simxGetObjectPosition(clientID,NAOHandle,-1,vrep.simx_opmode_blocking)
res, Orientation = vrep.simxGetObjectOrientation(clientID, NAOHandle, -1, vrep.simx_opmode_blocking)
# res, NAO_V, NAO_Angular_V = vrep.simxGetObjectVelocity(clientID, NAOHandle, vrep.simx_opmode_blocking)
#这里的速度与角速度是否可以用上去？
s = []
# state中存储绝对坐标中y的偏移量
dy = NAO_Pos[1]-y0
s.append(dy)
#s.append(dx)
s += Orientation
isFall = 0
isOut = 0
res, HeadPos = vrep.simxGetObjectPosition(clientID, HeadHandle, -1, vrep.simx_opmode_blocking)
if HeadPos[2] < 0.42: isFall = 1
if (abs(dy) > 0.1) or (abs(Orientation[2]) > 0.5): isOut = 1
return s, isFall, isOut
def CalculateReward(dx,y):#计算reward的方式是否合理？
w1 = 2
w2 = 1
reward = w2*dx - y*w1
print('dy:',y)
print('dx:',dx)
return reward
def step(position_vector):
global max_x
res, NAO_Pos1 = vrep.simxGetObjectPosition(clientID, NAOHandle, -1, vrep.simx_opmode_blocking)
# print(clientID)
res_ = oscillator.oscillator_step(position_vector)
i = 0
ls=[]
while True:
res = oscillator.oscillator_step(position_vector)
ls.append((res - bias1) / gain1)
vrep.simxSynchronousTrigger(clientID)
if res * res_ < 0:
i += 1
res_=res
if i==2:
break
# for i in range(20):
# oscillator.oscillator_step(position_vector)
# vrep.simxSynchronousTrigger(clientID)
res, NAO_Pos2 = vrep.simxGetObjectPosition(clientID, NAOHandle, -1, vrep.simx_opmode_blocking)
dx = NAO_Pos2[0] - NAO_Pos1[0]
dy = NAO_Pos2[1] - NAO_Pos1[1] #相对偏移量
#print('x: ', NAO_Pos2[0]+2)
#print('y: ', NAO_Pos2[1])
if NAO_Pos2[0] > max_x:
max_x = NAO_Pos2[0]
print('best_x updated!')
print('best_x: ', max_x+2)
r = CalculateReward(dx, abs(NAO_Pos2[1]))
s_, isFall, isOut = getState(dx)
if isFall: r -= 0.1
if isOut : r -= 0.1
return s_, r, isFall, isOut, dx, NAO_Pos2[1], NAO_Pos2[0],ls
kf = 0.2098797258 # control the frequency
gain1 = 0.4083754801 # hip
gain1= 0.6
gain2 = 0.4646833695 # ankle
gain3 = 0.0545742201 # knee
gain3 = 0.04
gain4 = 0.0197344836 # hip_x
gain5 = 0.5104797402 # ankle_x
gain6 = 0.523562905 # shoulder
bias1 = -0.0636451512 # hip
bias2 = 0.148065662 # knee
bias3 = -0.0410920591 # ankle
bias4 = 1.6105600582 # shoulder (the initial state is 1.57)
k = 1.7334593213 # feedback weight para, optimized with kf
para_vec = [kf, gain1, gain2, gain3, gain4, gain5, gain6, bias1, bias2, bias3, bias4, k]
MAX_EPISODES = 20000
MAX_EP_STEPS = 100
SAVE_MODEL_ITER = 500
MAX_EPISODE_REWARD = -100
STEP_COUNT = 0
# STATE_DIM = 5
STATE_DIM = 4
ACTION_DIM = 8
#ACTION_BOUND = [0.01, 0.01, 0.01, 0.01, 0.01, 0.01, 0.01, 0.005, 0.005, 0.005, 0.01, 0.05]
#ACTION_BOUND = [0.1, 0.3, 0.2, 0.03, 0.03, 0.2, 0.05, 0.005, 0.05, 0.01, 0.05, 0.1]
ACTION_BOUND = [0.01, 0.01, 0.01, 0.01, 0.01,0.005,0.005,0.005]
var = 0.5
var_min = 0.001
epsilon = 1
epsilon_decay = 0.95
osc_ls=[]
if __name__ == "__main__":
ddpg = DDPG(ACTION_DIM, STATE_DIM, ACTION_BOUND)
start = time.time()
print('start.', start)
t = time.strftime("%m%d_%H%M%S", time.localtime(time.time()))
with open("performance_data/{}.csv".format(t), 'a') as f:
f.write('i_epi,epi_r,x,epi_step,y_final\n') #一个episode结束的dy
for i_episode in range(MAX_EPISODES):
vrep.simxSynchronous(clientID, 1)
vrep.simxStartSimulation(clientID, vrep.simx_opmode_oneshot)
#print('#################start##################')
print("Now we are starting a new episode.")
global oscillator
oscillator= oscillator_nw(para_vec, clientID)
episode_reward = 0
x, dx, dy = 0, 0, 0
for i_step in range(MAX_EP_STEPS):
print("-------------------------------------------------------------------------")
s, _ ,_ = getState(dx)
if np.random.random() > epsilon:
a = ddpg.choose_action(np.array(s)) #a is d_para
a = np.clip(a, -1*np.array(ACTION_BOUND), np.array(ACTION_BOUND))
print('use a.')
else:
a = np.random.normal(np.zeros(ACTION_DIM), var)*np.array(ACTION_BOUND)
print('use random.')
#print('a:', a)#4
a_ = [0,a[0],a[1],a[2],a[3],a[4],0,a[5],a[6],a[7],0,0]
para_vec_in = para_vec+ a_
s_, r, isFall, isOut, dx, y, x, osc = step(para_vec)
osc_ls+=osc
episode_reward+=r
ddpg.store_transition(s, a, r, s_)
data = []
data = np.append(data, s)
data = np.append(data, a)
data = np.append(data, [r])
data = np.append(data, s_)
STEP_COUNT += 1
if STEP_COUNT > 500:
print("I'm learning")
for i in range(1): ddpg.learn()
if STEP_COUNT % SAVE_MODEL_ITER == 0:
ddpg.save_model(STEP_COUNT)
print('model saved. Used time: ', int(time.time() - start))
if i_episode > 200:
if i_episode % 3 == 0:
epsilon *= epsilon_decay
if episode_reward > MAX_EPISODE_REWARD:
MAX_EPISODE_REWARD = episode_reward
print('Episode:', i_episode,
'| Step: %i' % i_step,
'| Epi_reward: %f' % episode_reward,
'| Max_epi_r: %f' % MAX_EPISODE_REWARD,
'| Reward for this step: %f' % r,
'| Global steps: %i' % STEP_COUNT,
'| Epsilon: %.4f' % epsilon,
'| ls: ',osc_ls
)
if isFall:
print("The robot falls down.")
break
if isOut:
print("The robot is outside")
break
with open("performance_data/{}.csv".format(t), 'a') as f:
x+=2
f.write(str(i_episode)+','+str(episode_reward)+','+str(x)+','+str(i_step)+','+str(y)+'\n')
vrep.simxStopSimulation(clientID, vrep.simx_opmode_oneshot)
time.sleep(1)
vrep.simxFinish(clientID)
print("Finish.")
# if __name__ == "__main__":
# for i_episode in range(MAX_EPISODES):
# vrep.simxSynchronous(clientID, 1)
# vrep.simxStartSimulation(clientID, vrep.simx_opmode_oneshot)
# #print('#################start##################')
# print("Now we are starting a new episode.")
# episode_reward = 0
# #para_vec_old = para_vec
# dx = 0
# for i_step in range(MAX_EP_STEPS):
# print("-------------------------------------------------------------------------")
# s, _ ,_ = getState(dx)
# para_vec_in = para_vec
# s_, r, isFall, isOut, dx = step(para_vec_in)
# episode_reward+=r
# data = []
# data = np.append(data, s)
# data = np.append(data, [r])
# data = np.append(data, s_)
# STEP_COUNT += 1
# print('Episode:', i_episode,
# '| Step: %i' % i_step,
# '| Epi_reward: %f' % episode_reward,
# '| Exploration: %.3f' % var,
# '| Reward for this step: %f' % r,
# '| Global steps: %i' % STEP_COUNT,
# )
# if isFall:
# print("The robot falls down.")
# break
# if isOut:
# print("The robot is outside")
# break
# vrep.simxStopSimulation(clientID, vrep.simx_opmode_oneshot)
# time.sleep(1)
# vrep.simxFinish(clientID)
# print("Finish.")