cubm = CubesManager()
observation_dim = 3
action_dim = 3
action_bound = -1, 1
# set RL method (continuous)
rl = DDPG(action_dim, observation_dim, action_bound)
number = 0
steps = []
# 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)
if __name__ == "__main__":
robot = Robot()
s_dim = robot.state_dim
a_dim = robot.action_dim
a_bound = robot.action_bound
cubm = CubesManager()
rl = DQN()
for i in range(MAX_EPISODES):
robot.reset()
cubm.reset_cube(rand=True)
Box_position = cubm.read_cube_pose("cube1")
Box_position[0] -= 0.2
Box_position[2] -= 0.1
robot.Box_position = Box_position
# print(cubm.read_cube_pose("cube1"))
# print(robot.Box_position)
s = robot.get_state()
st = 0
rw = 0
while True:
st += 1
a = rl.choose_action(s)
s_, r, done = robot.step(a)
rw += r
r = -r
rl.store_transition(s, a, r, s_)
if rl.memory_counter > 50:
rl.learn()