-
Notifications
You must be signed in to change notification settings - Fork 3
/
Copy pathenv_aug.py
179 lines (150 loc) · 6.08 KB
/
env_aug.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
from gym.envs.mujoco import HalfCheetahEnv, AntEnv, HopperEnv
import numpy as np
from gym import utils
import os
import torch
from gym.envs.mujoco import mujoco_env
from gym.envs.registration import EnvSpec
class HalfCheetahEnvAug(HalfCheetahEnv):
def __init__(self):
self.aug_vel = 0
super(HalfCheetahEnvAug, self).__init__()
self.spec = EnvSpec('HalfCheetah-v2')
def _get_obs(self):
return np.concatenate([
self.sim.data.qpos.flat[1:],
self.sim.data.qvel.flat,
np.array([self.aug_vel])
])
def reset_model(self):
qpos = self.init_qpos + self.np_random.uniform(low=-.1, high=.1, size=self.model.nq)
qvel = self.init_qvel + self.np_random.randn(self.model.nv) * .1
self.set_state(qpos, qvel)
self.aug_vel = 0
return self._get_obs()
def step(self, action):
xposbefore = self.sim.data.qpos[0]
self.do_simulation(action, self.frame_skip)
xposafter = self.sim.data.qpos[0]
self.aug_vel = (xposafter - xposbefore)/self.dt
ob = self._get_obs()
reward_ctrl = - 0.1 * np.square(action).sum()
reward_run = (xposafter - xposbefore)/self.dt
reward = reward_ctrl + reward_run
done = False
return ob, reward, done, dict(reward_run=reward_run, reward_ctrl=reward_ctrl)
class AntEnvAug(AntEnv):
def __init__(self):
super(AntEnvAug, self).__init__()
self.aug_vel = 0
self.spec = EnvSpec('Ant-v2')
def _get_obs(self):
return np.concatenate([
self.sim.data.qpos.flat[2:],
self.sim.data.qvel.flat,
np.array([self.aug_vel])
])
def reset_model(self):
qpos = self.init_qpos + self.np_random.uniform(size=self.model.nq, low=-.1, high=.1)
qvel = self.init_qvel + self.np_random.randn(self.model.nv) * .1
self.set_state(qpos, qvel)
self.aug_vel = 0
return self._get_obs()
def step(self, a):
xposbefore = self.get_body_com("torso")[0]
self.do_simulation(a, self.frame_skip)
xposafter = self.get_body_com("torso")[0]
self.aug_vel = (xposafter - xposbefore)/self.dt
forward_reward = (xposafter - xposbefore)/self.dt
ctrl_cost = .5 * np.square(a).sum()
contact_cost = 0.5 * 1e-3 * np.sum(
np.square(np.clip(self.sim.data.cfrc_ext, -1, 1)))
survive_reward = 1.0
# In the true env we keep the contact cost so we can compare to other algorithms
reward = forward_reward - ctrl_cost - contact_cost + survive_reward
state = self.state_vector()
notdone = np.isfinite(state).all() \
and state[2] >= 0.2 and state[2] <= 1.0
done = not notdone
ob = self._get_obs()
return ob, reward, done, dict(
reward_forward=forward_reward,
reward_ctrl=-ctrl_cost,
reward_contact=-contact_cost,
reward_survive=survive_reward)
@staticmethod
def is_done_func(states):
finite_check = torch.isfinite(states).all(dim=1)
bounds_check = (states[:,0] >= 0.2) & (states[:,0] <= 1.0)
notdone = finite_check & bounds_check
return ~notdone
class HopperEnvAug(HopperEnv):
def __init__(self):
self.aug_vel = 0
super(HopperEnvAug, self).__init__()
self.spec = EnvSpec('Hopper-v2')
def step(self, a):
posbefore = self.sim.data.qpos[0]
self.do_simulation(a, self.frame_skip)
posafter, height, ang = self.sim.data.qpos[0:3]
alive_bonus = 1.0
reward = (posafter - posbefore) / self.dt
self.aug_vel = reward
reward += alive_bonus
reward -= 1e-1 * np.square(a).sum()
height_penalty = -3.0 * (height - 1.3)**2
reward += height_penalty
s = self.state_vector()
done = False
ob = self._get_obs()
return ob, reward, done, {}
def _get_obs(self):
return np.concatenate([
self.sim.data.qpos.flat[1:],
np.clip(self.sim.data.qvel.flat, -10, 10),
np.array([self.aug_vel])
])
def reset_model(self):
qpos = self.init_qpos + self.np_random.uniform(low=-.005, high=.005, size=self.model.nq)
qvel = self.init_qvel + self.np_random.uniform(low=-.005, high=.005, size=self.model.nv)
self.set_state(qpos, qvel)
self.aug_vel = 0
return self._get_obs()
def mass_center(model, sim):
mass = np.expand_dims(model.body_mass, 1)
xpos = sim.data.xipos
return (np.sum(mass * xpos, 0) / np.sum(mass))[0]
class fixedSwimmerEnv(mujoco_env.MujocoEnv, utils.EzPickle):
def __init__(self):
dir_path = os.path.dirname(os.path.realpath(__file__))
mujoco_env.MujocoEnv.__init__(self, '%s/assets/fixed_swimmer.xml' % dir_path, 4)
utils.EzPickle.__init__(self)
self.spec = EnvSpec('Swimmer-v2')
self.pos_diff = 0
def step(self, a):
ctrl_cost_coeff = 0.0001
"""
xposbefore = self.model.data.qpos[0, 0]
self.do_simulation(a, self.frame_skip)
xposafter = self.model.data.qpos[0, 0]
"""
self.xposbefore = self.sim.data.site_xpos[0][0] / self.dt
self.do_simulation(a, self.frame_skip)
self.xposafter = self.sim.data.site_xpos[0][0] / self.dt
self.pos_diff = self.xposafter - self.xposbefore
reward_fwd = self.xposafter - self.xposbefore
reward_ctrl = - ctrl_cost_coeff * np.square(a).sum()
reward = reward_fwd + reward_ctrl
ob = self._get_obs()
return ob, reward, False, dict(reward_fwd=reward_fwd, reward_ctrl=reward_ctrl)
def _get_obs(self):
qpos = self.sim.data.qpos
qvel = self.sim.data.qvel
return np.concatenate([qpos.flat[2:], qvel.flat, np.array([self.pos_diff])])
def reset_model(self):
self.set_state(
self.init_qpos + self.np_random.uniform(low=-.1, high=.1, size=self.model.nq),
self.init_qvel + self.np_random.uniform(low=-.1, high=.1, size=self.model.nv)
)
self.pos_diff = 0
return self._get_obs()