-
Notifications
You must be signed in to change notification settings - Fork 1
/
solve_mdp.py
208 lines (183 loc) · 6.61 KB
/
solve_mdp.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
import os
import numpy as np
import argparse
from distutils.util import strtobool
from multiprocessing import Pool, Manager
from numpy.core.numeric import normalize_axis_tuple
from mdp import (
States,
Actions,
Rewards,
Policy,
StateTransitionProbability,
MarkovDecisionProcess,
PolicyIteration,
ValueIteration
)
from environment import DifferentialDriveKinematic
from planner import NmpcCartesian, NmpcPolar
from utils import Verbose, print_test, normalize_angle
parser = argparse.ArgumentParser()
env_group = parser.add_argument_group('env')
env_group.add_argument('--env-dt', type=float, default=0.05)
env_group.add_argument('--env-pose_tol', type=float, default=1e-2)
env_group.add_argument('--env-control_tol', type=float, default=1e-2)
env_group.add_argument('--env-timeout', type=float, default=1000)
mpc_group = parser.add_argument_group('mpc')
mpc_group.add_argument('--mpc-dt', type=float, default=0.05)
mpc_group.add_argument('--mpc-time_horizon', type=int, default=10)
mpc_group.add_argument('--mpc-regularization_control', type=float, default=1e-3)
mpc_group.add_argument('--mpc-regularization_v', type=float, default=0)
mpc_group.add_argument('--mpc-regularization_omega', type=float, default=0)
mpc_group.add_argument('--mpc-r_tol', type=float, default=5e-3)
mdp_group = parser.add_argument_group('mdp')
mdp_group.add_argument('--n_r', type=int, default=40)
mdp_group.add_argument('--n_alpha', type=int, default=24)
mdp_group.add_argument('--n_psi', type=int, default=25)
mdp_group.add_argument('--n_theta', type=int, default=7)
mdp_group.add_argument('--repeats_per_sample_point', type=int, default=1)
mdp_group.add_argument('--mdp-discount', type=float, default=0.99)
mdp_group.add_argument('--mdp-max_iter', type=int, default=100000)
mdp_group.add_argument('--mdp-tol', type=float, default=1e-6)
mdp_group.add_argument('--mdp-solver', type=str, default='ValueIteration',
choices=['PolicyIteration', 'ValueIteration']
)
parser.add_argument('--data_dir', type=str, default='data/mdp')
parser.add_argument('--sample', type=strtobool, default=True)
parser.add_argument('--solve', type=strtobool, default=True)
parser.add_argument('--test', type=strtobool, default=True)
args = parser.parse_args()
if not os.path.isdir(args.data_dir):
print('{} does not exist. Create target directory...'.format(args.data_dir))
os.mkdir(args.data_dir)
file_mdp = os.path.join(args.data_dir, 'mdp.npz')
file_result = os.path.join(args.data_dir, 'result.npz')
file_test = os.path.join(args.data_dir, 'test.npy')
if not args.sample:
if not os.path.exists(file_mdp):
raise FileNotFoundError('{} does not exist.'.format(file_mdp))
if not args.solve:
if not os.path.exists(file_result):
raise FileNotFoundError('{} does not exist.'.format(file_result))
r_list = np.abs(-np.log(np.linspace(1, 0, args.n_r + 1, dtype=np.float32)[:-1]))
alpha_list = np.linspace(-np.pi+np.pi/args.n_alpha, np.pi-np.pi/args.n_alpha, args.n_alpha, dtype=np.float32)
psi_list = np.linspace(-np.pi+np.pi/args.n_psi, np.pi-np.pi/args.n_psi, args.n_psi, dtype=np.float32)
states = States(
r_list,
alpha_list,
psi_list,
cycles=[None, np.pi * 2, np.pi * 2],
terminal_states=[np.array([0., alpha, 0.], dtype=np.float32) for alpha in alpha_list]
)
action_list = []
for theta1 in np.cos(np.linspace(0, np.pi/2, args.n_theta)):
for theta2 in np.cos(np.linspace(0, np.pi/2, args.n_theta)):
for theta3 in np.cos(np.linspace(0, np.pi/2, args.n_theta)):
action_list.append(
np.array([2 - 2*theta1, 2 - 2*theta2, 2 - 2*theta3], dtype=np.float32) + np.finfo(np.float32).eps
)
actions = Actions(
action_list
)
env = DifferentialDriveKinematic(
dt=args.env_dt, tol=np.finfo(np.float32).eps, timeout=args.env_timeout,
observation_noise=np.zeros((3), dtype=np.float32),
transition_noise=np.zeros((3), dtype=np.float32)
)
nmpc = NmpcCartesian(
T=args.mpc_dt,
N=args.mpc_time_horizon,
R_u=args.mpc_regularization_control*np.eye(2),
R_vel=np.diag(
np.array([args.mpc_regularization_v, args.mpc_regularization_omega])
),
w_max=2,
a_max=1,
v_max=0.2,
omega_max=0.2 * 0.8 * 2 / 0.653,
r_tol=args.mpc_r_tol
)
def simulate(state, action):
reward = -1
state = env.reset(state)
u = nmpc.solve(state, action, warm_start=0)
env.step(u)
state = env.get_state()
return state, reward
state_transition_prob = StateTransitionProbability(
states = states, actions = actions,
)
rewards = Rewards(states, actions)
policy = Policy(states = states, actions = actions)
mdp = MarkovDecisionProcess(
states = states,
actions = actions,
rewards = rewards,
state_transition_probability = state_transition_prob,
policy = policy,
discount = args.mdp_discount
)
if args.sample:
mdp.sample(simulate, n_repeat=args.repeats_per_sample_point, sample_reward=True)
mdp.save(file_mdp)
print('The MDP is recorded in {}'.format(file_mdp))
else:
mdp.load(args.data_dir+'/mdp.npz')
mdp.discount = args.mdp_discount
if args.solve:
solver = eval(args.mdp_solver)(mdp)
solver.solve(max_iteration=args.mdp_max_iter, tolerance=args.mdp_tol)
solver.save(file_result)
print('The results are recorded in {}'.format(file_result))
else:
res = np.load(file_result)
mdp.policy.update(res['policy'])
if args.test:
policy = mdp.policy
env.tol = args.env_pose_tol
nmpc.r_tol = args.mpc_r_tol
def simulate(queue, state):
state = env.reset(state)
done = False
traj = []
traj.append(env.get_state())
u_prev = np.zeros((2))
info = {'success':None}
while not done:
theta = policy.get_action(state)
u = nmpc.solve(state, theta, u_prev=u_prev, warm_start=0)
state, _, done, info = env.step(u)
if np.linalg.norm(u) + np.linalg.norm(u - u_prev) < args.env_control_tol:
done = True
u_prev = u
traj.append(env.get_state())
queue.put(1)
return traj, info['success']
data = []
for r in np.linspace(0.5, 1.5, 3):
for alpha in np.linspace(-np.pi, np.pi, 19)[1:]:
for psi in np.linspace(-np.pi, np.pi, 19)[1:]:
data.append([r, alpha, psi])
verbose = Verbose(True)
queue = Manager().Queue()
with Pool(os.cpu_count()) as p:
trajs = p.starmap_async(simulate, [(queue, state) for state in data])
counter = 0
while counter < len(data):
counter += queue.get()
progress = counter / len(data)
progress *= 100
verbose('Simulation progress: %5.1f %%...'%(progress))
verbose('Simulation is done.\n')
trajs = trajs.get()
err = []
arrival_time = []
for target, (traj, success) in zip(data, trajs):
target.append(traj)
target.append(success)
np.save(
file_test,
np.array(data, dtype=np.object),
allow_pickle=True
)
print_test(data, 3)