-
Notifications
You must be signed in to change notification settings - Fork 0
/
mw.py
219 lines (172 loc) · 6.79 KB
/
mw.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
# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved.
#
# This source code is licensed under the MIT license found in the
# LICENSE file in the root directory of this source tree.
import metaworld
import random
import gymnasium as gym
# import gym
import mujoco
from metaworld.envs.mujoco.sawyer_xyz.v2 import SawyerPushEnvV2 as Env
from metaworld.envs.mujoco.env_dict import ALL_V2_ENVIRONMENTS
def make_env(env_name, seed=42, render_mode=None):
from metaworld.envs import (
ALL_V2_ENVIRONMENTS_GOAL_OBSERVABLE,
ALL_V2_ENVIRONMENTS_GOAL_HIDDEN,
)
task = f"{env_name}-goal-observable"
env_cls = ALL_V2_ENVIRONMENTS_GOAL_OBSERVABLE[task]
env = env_cls(seed=seed,render_mode=render_mode)
env._freeze_rand_vec = False
# 相机视角设置
env.camera_name = "corner2"
env.model.cam_pos[2][:] = [0.75, 0.075, 0.7]
return env
def make_env_randomized(env_name, seed=0, render_mode=None):
Env = ALL_V2_ENVIRONMENTS[env_name]
env = Env(render_mode=render_mode)
env._freeze_rand_vec = False
env._set_task_called = True
return env
class MWEnvWrapper:
def __init__(self, seed=0, env_name='push-v2', task_id=None):
self.env_name = env_name
self.env = make_env(env_name,seed,render_mode='rgb_array')
self.env.model.vis.global_.offwidth = 84
self.env.model.vis.global_.offheight = 84 # 设置 render
self.act_dim = int(np.prod(self.env.action_space.shape))
self.state_dim = int(np.prod(self.env.observation_space.shape))
self.env.mujoco_renderer.width = 84
self.env.mujoco_renderer.height = 84
self.env.model.cam_pos[2][:] = [0.75, 0.075, 0.7]
# set the camera id
self.env.mujoco_renderer.camera_id = mujoco.mj_name2id(
self.env.model,
mujoco.mjtObj.mjOBJ_CAMERA,
"corner2",
)
# print("action dim:{},state dim:{}".format(self.act_dim, self.state_dim))
def reset(self):
state , _ = self.env.reset()
return state
def get_random_action(self):
action = self.env.action_space.sample()
return action
def step(self, action):
next_state, reward, terminate, truncated,info = self.env.step(action.ravel())
done = terminate or truncated
return next_state, reward, done, info
def set_random_seed(self, seed):
self.env.seed(seed)
def render(self):
frame = self.env.render()
return frame
def render_for_video(self):
frame = self.env.render_for_video()
return frame
def close(self):
self.env.close()
def get_action_space(self):
return self.env.action_space
def get_obs_space(self):
return self.env.observation_space
def normalise_state(self, state):
return state
def normalise_reward(self, reward):
return reward
import numpy as np
from collections import deque
from gymnasium.spaces import Box
class ActionDTypeWrapper(gym.Wrapper):
def __init__(self, env: MWEnvWrapper, dtype):
self._env = env
action_space = env.get_action_space()
self.action_space = Box(
low=action_space.low.astype(dtype),
high=action_space.high.astype(dtype),
shape=action_space.shape,
dtype=dtype,
)
def step(self, action):
action = action.astype(self.action_space.dtype)
return self._env.step(action)
def reset(self):
return self._env.reset()
def render(self):
return self._env.render()
def render_for_video(self):
return self._env.render_for_video()
def __getattr__(self, name):
return getattr(self._env, name)
class ActionRepeatWrapper(gym.Wrapper):
def __init__(self, env: ActionDTypeWrapper, num_repeats,discount):
self._env = env
self._num_repeats = num_repeats
self.discount = discount
def step(self, action):
total_reward = 0.0
gamma = 1.0
for _ in range(self._num_repeats):
observation, reward, done,info = self._env.step(action)
total_reward += reward * gamma
gamma *= self.discount
if done:
break
return observation, total_reward, done, gamma , info
def reset(self):
return self._env.reset()
def render(self):
return self._env.render()
def render_for_video(self):
return self._env.render_for_video()
def __getattr__(self, name):
return getattr(self._env, name)
def stack_frames(frames):
return np.concatenate(list(frames), axis=0)
class MetaWorldFrameStackWrapper(gym.Wrapper):
def __init__(self, env: ActionRepeatWrapper, num_frames):
super().__init__(env)
self._num_frames = num_frames
self._frames = deque([], maxlen=num_frames)
self._env = env
obs = env.render()
obs_shape = obs.shape
# remove batch dim
if len(obs_shape) == 4:
obs_shape = obs_shape[1:]
new_shape = (obs_shape[2] * num_frames, obs_shape[0],obs_shape[1])
# print("new_shape is {}".format(new_shape)) 9,480,480
self.observation_space = Box(low=0, high=255, shape=new_shape, dtype=np.uint8)
def reset(self):
self._env.reset()
pixels = get_pixels(self._env)
# print("pixels shape : {}".format(pixels.shape)) 3,480,480
for _ in range(self._num_frames):
self._frames.append(pixels)
return stack_frames(self._frames)
def step(self, action):
obs, reward, done, gamma,info = self._env.step(action)
pixels = get_pixels(self._env)
# print("after step, pixel shape is {}".format(pixels.shape))
self._frames.append(pixels)
# print("after step, len(self._frames) is {}".format(len(self._frames)))
assert len(self._frames) == self._num_frames
return stack_frames(self._frames), reward, done, gamma,info
def render(self):
return self._env.render()
def render_for_video(self):
return self._env.render_for_video()
def get_pixels(env: MetaWorldFrameStackWrapper):
obs = env.render()
# remove batch dim
if len(obs.shape) == 4:
obs = obs[0]
return obs.transpose(2, 0, 1).copy()
def make(env_name, frame_stack, action_repeat, seed,discount):
# env = make_env(seed=seed,render_mode='rgb_array')
env = MWEnvWrapper(seed=seed,env_name=env_name) # 此时 step 返回值是 state,reward,done,info
env = ActionDTypeWrapper(env, np.float32)
env = ActionRepeatWrapper(env, action_repeat, discount) # 此时 step 返回值是 state,reward,done, gamma,info
# # stack several frames
env = MetaWorldFrameStackWrapper(env, frame_stack)
return env