-
Notifications
You must be signed in to change notification settings - Fork 8
/
car_kinematic_dqn.py
251 lines (217 loc) · 10.2 KB
/
car_kinematic_dqn.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
import pygame
from pygame.math import Vector2
from ConvDQNAgent import *
import cv2
from car_kinematic_collision import Collision
from math_util import *
from car import Car
from car_kinematic_model import Game
from action_handler import *
from car_kinematic_obstacle import update_object_mask
class DqnGame(Game):
def __init__(self, screen, screen_width, screen_height):
super().__init__(screen, screen_width, screen_height, False, False)
pygame.display.set_caption("Car kinematic dqn model")
self.ticks = 10
self.dqn_x_input_pixel = 220
self.dqn_y_input_pixel = 470
self.hw_surface_box = 320
self.dqn_input = None
self.global_distance = 0.0
# TODO TEST SEAMLESS DQN
self.bkd_color = [255, 255, 0, 255]
self.background = pygame.image.load("resources/backgrounds/maps_overlay.png").convert()
self.background = pygame.transform.scale(self.background, (2500, 1261))
self.bgWidth, self.bgHeight = self.background.get_rect().size
self.object_mask = pygame.Surface((self.screen_width, self.screen_height))
def get_sensor_reward(self, car):
center_rect = Collision.center_rect(self.screen_width, self.screen_height)
mid_of_front_axle = Collision.point_rotation(car, 0, 16, center_rect)
mid_of_rear_axle = Collision.point_rotation(car, 65, 16, center_rect)
# pygame.draw.circle(self.screen, (255, 255, 0), (mid_of_front_axle[0], mid_of_front_axle[1]), 5)
distances = self.enable_sensor(car, self.object_mask, self.screen)
min_distance = np.min(distances)
# rear_distance = self.compute_sensor_distance(car, mid_of_rear_axle, 70, 0)
if min_distance >= 50:
return 1.0
elif 50 > min_distance > 40:
return 0.4
elif 40 > min_distance > 30:
return -0.8
else:
return -1.0
def step(self, action, car, dt):
current_position = [0, 0]
next_position = [0, 0]
current_position[0], current_position[1] = car.position[0], car.position[1]
apply_action(action, car, dt)
self.render(car, dt)
sensor_distance_reward = self.get_sensor_reward(car)
pygame.display.update()
next_state = self.get_current_state(first_frames=False)
next_position[0], next_position[1] = car.position[0], car.position[1]
local_distance = round(
np.sqrt((current_position[0] - next_position[0]) ** 2 + (current_position[1] - next_position[1]) ** 2), 4)
if local_distance > 0.0:
reward = normalize_zero_one(local_distance, 0.9, 0) * 0.15 + normalize_zero_one(car.velocity[0],
car.max_velocity,
0) * 0.15 + sensor_distance_reward * 0.7
else:
reward = -0.5
print("reward " + str(reward))
on_road = self.on_road(car, self.object_mask)
if on_road:
return next_state, reward, False
else:
return next_state, -1, True
def render(self, car, dt):
# Logic
car.update(dt)
# Drawing
stagePosX = car.position[0] * self.ppu
stagePosY = car.position[1] * self.ppu
rel_x = stagePosX % self.bgWidth
rel_y = stagePosY % self.bgHeight
# blit (BLock Image Transfer) the seamless background
self.screen.blit(self.background, (rel_x - self.bgWidth, rel_y - self.bgHeight))
self.screen.blit(self.background, (rel_x, rel_y))
self.screen.blit(self.background, (rel_x - self.bgWidth, rel_y))
self.screen.blit(self.background, (rel_x, rel_y - self.bgHeight))
rotated = pygame.transform.rotate(self.car_image, car.angle)
rect = rotated.get_rect()
center_x = int(self.screen_width / 2) - int(rect.width / 2)
center_y = int(self.screen_height / 2) - int(rect.height / 2)
self.screen.blit(rotated, (center_x, center_y))
self.object_mask.fill((0, 0, 0))
self.object_mask.blit(self.screen, (0, 0))
update_object_mask(self.object_mask, rel_x, rel_y, self.bgWidth, self.bgHeight)
def get_current_state(self, first_frames):
state_resized = self.object_mask.subsurface(
(self.dqn_y_input_pixel, self.dqn_x_input_pixel, self.hw_surface_box, self.hw_surface_box))
state_resized = pygame.transform.scale(state_resized,
(int(self.hw_surface_box / 4), int(self.hw_surface_box / 4)))
state_rgb_arr = pygame.surfarray.array3d(state_resized)
state_bw = cv2.cvtColor(state_rgb_arr, cv2.COLOR_RGB2GRAY)
norm = np.array([])
norm = cv2.normalize(state_bw, dst=norm, alpha=0, beta=1, norm_type=cv2.NORM_MINMAX, dtype=cv2.CV_32F)
if first_frames:
single_state_tensor = np.reshape(norm, (norm.shape[0], norm.shape[1], 1))
multi_state_tensor = np.concatenate(
[single_state_tensor, single_state_tensor, single_state_tensor, single_state_tensor], axis=2)
self.dqn_input = np.reshape(multi_state_tensor, (
1, multi_state_tensor.shape[0], multi_state_tensor.shape[1], multi_state_tensor.shape[2]))
else:
self.dqn_input[0, :, :, 0] = self.dqn_input[0, :, :, 1]
self.dqn_input[0, :, :, 1] = self.dqn_input[0, :, :, 2]
self.dqn_input[0, :, :, 2] = self.dqn_input[0, :, :, 3]
self.dqn_input[0, :, :, 3] = norm
return self.dqn_input
def random_reset(self, car):
reset_posXY_vel_list = [[5, 125, 90.0], [33.5, -130, 0.0], [100.00, -130.0, 0.0],
[123.0, 100.0, 90.0]]
random_state = random.choice(reset_posXY_vel_list)
car.position = Vector2(random_state[0], random_state[1])
car.velocity = Vector2(0.0, 0.0)
car.angle = random_state[2]
def run_conv_dqn(self, episodes):
eps = episodes
state_w = int(self.hw_surface_box / 4)
state_h = int(self.hw_surface_box / 4)
state_size = (state_w, state_h, 4)
action_size = 7
batch_size = 4
# agent = ConvDQNAgent(state_size, action_size)
agent = ConvDQNAgent(state_size, action_size, True, "./checkpoint/gridsim-dqn")
agent.load_memory("./checkpoint/gridsim-dqn")
state = None
# place car on road
# car = Car(pygame.display.get_surface().get_width() / self.ppu / 2,
# pygame.display.get_surface().get_height() / self.ppu / 2 - 26)
car = Car(0, 25)
car.max_steering = 25
# car = Car(0, 0)
# pressed = pygame.key.get_pressed()
for e in range(eps):
init = True
sanity_check = 15
while not self.exit:
dt = self.clock.get_time() / 1000
for event in pygame.event.get():
if event.type == pygame.QUIT:
self.exit = True
if init:
self.render(car, dt)
self.on_road(car, self.object_mask)
pygame.display.update()
init = False
state = self.get_current_state(first_frames=True)
else:
action = agent.act(state)
next_state, reward, done = self.step(action, car, dt)
if reward < -0.2:
sanity_check -= 1
else:
sanity_check = 15
if sanity_check < 0:
done = True
reward = -1
agent.remember(state, action, reward, next_state, done)
state = next_state
if len(agent.memory) > batch_size:
agent.replay(batch_size)
if done:
print("episode: {}/{}, score: {}, e: {:.2}"
.format(e, episodes, time, agent.epsilon))
# self.reset(car)
self.random_reset(car)
break
self.clock.tick(self.ticks)
if e % 100 == 0 and e != 0:
agent.save("./checkpoint/gridsim-dqn")
if e % 1000 == 0 and e != 0:
agent.load_target("./checkpoint/gridsim-dqn")
agent.save_memory("./checkpoint/gridsim-dqn")
pygame.quit()
def predict_conv_dqn(self):
state_w = int(self.hw_surface_box / 4)
state_h = int(self.hw_surface_box / 4)
state_size = (state_w, state_h, 4)
action_size = 7
agent = ConvDQNAgent(state_size, action_size, True, "./checkpoint/gridsim-dqn")
agent.load_memory("./checkpoint/gridsim-dqn")
car = Car(pygame.display.get_surface().get_width() / self.ppu / 2,
pygame.display.get_surface().get_height() / self.ppu / 2 - 26)
state = None
for e in range(1000):
init = True
sanity_check = 15
while not self.exit:
dt = self.clock.get_time() / 1000
for event in pygame.event.get():
if event.type == pygame.QUIT:
self.exit = True
if init:
self.render(car, dt)
self.on_road(car, self.object_mask)
pygame.display.update()
init = False
state = self.get_current_state(first_frames=True)
else:
action = agent.act(state)
next_state, reward, done = self.step(action, car, dt)
state = next_state
if reward < -0.2:
sanity_check -= 1
else:
sanity_check = 15
if done:
self.random_reset(car)
break
self.clock.tick(self.ticks)
pygame.quit()
if __name__ == '__main__':
screen = pygame.display.set_mode((1280, 720))
game = DqnGame(screen, 1280, 720)
episodes = 100000
game.run_conv_dqn(episodes)
# game.predict_conv_dqn()