def collect_data():
robot = Robot()
cubusm = CubesManager()
for i in range(MAX_PICTURE_NUM):
cubusm.reset_cube(rand=True)
Box_position = cubusm.read_cube_pose("demo_cube")
# print "cube position:", str(Box_position)
joint, view = robot.get_state()
rgb, dep = robot.get_rgb_dep()
# b, g, r = cv2.split(rgb)
# print view[0,0,0]
# print dep
# rgb = cv2.merge([r, g, b])
# print dep
# plt.imshow(dep)
# plt.show()
rgb = cv2.resize(rgb, (224, 224))
dep = cv2.resize(dep, (224, 224))
# print dep
cv2.imwrite(
"/home/ljt/Desktop/images/rgb/" + str(Box_position) + ".png", rgb)
# cv2.imwrite("/home/ljt/Desktop/ws/src/fetch_moveit_config/images/dep/" + str(Box_position) + ".png", dep)
# a = np.array(rgb).shape
# print a
# print "camera image shape:", view.shape
np.save("/home/ljt/Desktop/images/dep/" + str(Box_position), dep)
# start training
for i in range(MAX_EPISODES):
cubm.reset_cube(rand=True)
Box_position = cubm.read_cube_pose("demo_cube")
print "cube position:", Box_position
robot.Box_position = copy.deepcopy(Box_position)
now_position = robot.gripper.get_current_pose(
"gripper_link").pose.position
now_dis = math.sqrt(
math.pow(now_position.x - robot.Box_position[0], 2) +
math.pow(now_position.y - robot.Box_position[1], 2) +
math.pow(now_position.z - robot.Box_position[2], 2))
robot.reward = -10 * now_dis
robot.reset()
s = robot.get_state()
ep_r = 0. # reward of each epoch
for j in range(MAX_EP_STEPS):
a = rl.choose_action(s)
s_, r, done = robot.step(a)
number += 1
print "-------the %i step-------" % number
rl.store_transition(s, a, r, s_)
# print s_[0]
ep_r += r
if rl.memory_full:
# start to learn once has fulfilled the memory
rl.learn()
# rl.learn()