-
Notifications
You must be signed in to change notification settings - Fork 16
/
run.py
282 lines (225 loc) · 10.7 KB
/
run.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
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
import opensim
import math
import numpy as np
import os
import random
import string
from itertools import chain
from .osim import OsimEnv
def flatten(listOfLists):
"Flatten one level of nesting"
return chain.from_iterable(listOfLists)
class RunEnv(OsimEnv):
STATE_PELVIS_X = 1
STATE_PELVIS_Y = 2
STATE_PELVIS_V_X = 4
MUSCLES_PSOAS_R = 3
MUSCLES_PSOAS_L = 11
num_obstacles = 0
max_obstacles = None
model_path = os.path.join(os.path.dirname(__file__), '../models/gait9dof18musc.osim')
ligamentSet = []
verbose = False
pelvis = None
env_desc = {"obstacles": [], "muscles": [1]*18}
ninput = 41
noutput = 18
def __init__(self, visualize=True, max_obstacles=3, original_reward=False):
self.original_reward = original_reward
self.max_obstacles = max_obstacles
super(RunEnv, self).__init__(visualize = False, noutput = self.noutput)
self.osim_model.model.setUseVisualizer(visualize)
self.create_obstacles()
state = self.osim_model.model.initSystem()
if visualize:
manager = opensim.Manager(self.osim_model.model)
manager.setInitialTime(-0.00001)
manager.setFinalTime(0.0)
manager.integrate(state)
def setup(self, difficulty, seed=None):
# create the new env
# set up obstacles
self.env_desc = self.generate_env(difficulty, seed, self.max_obstacles)
self.clear_obstacles(self.osim_model.state)
for x,y,r in self.env_desc['obstacles']:
self.add_obstacle(self.osim_model.state,x,y,r)
# set up muscle strength
self.osim_model.set_strength(self.env_desc['muscles'])
def reset(self, difficulty=2, seed=None):
super(RunEnv, self).reset()
self.istep = 0
self.setup(difficulty, seed)
self.last_state = self.get_observation()
self.current_state = self.last_state
return self.last_state
def compute_reward(self):
# Compute ligaments penalty
lig_pen = 0
# Get ligaments
for j in range(20, 26):
lig = opensim.CoordinateLimitForce.safeDownCast(self.osim_model.forceSet.get(j))
lig_pen += lig.calcLimitForce(self.osim_model.state) ** 2
if self.original_reward:
reward = self.current_state[self.STATE_PELVIS_X] - self.last_state[self.STATE_PELVIS_X]
else:
reward = self.current_state[self.STATE_PELVIS_V_X] * 0.01
reward += 0.01 # small reward for still standing
reward += min(0, self.current_state[22] - self.current_state[self.STATE_PELVIS_X]) * 0.1 # penalty for head behind pelvis
reward -= sum([max(0.0, k - 0.1) for k in [self.current_state[7], self.current_state[10]]]) * 0.02 # penalty for straight legs
return reward - math.sqrt(lig_pen) * 10e-8
def is_pelvis_too_low(self):
return (self.current_state[self.STATE_PELVIS_Y] < 0.65)
def is_done(self):
return self.is_pelvis_too_low() or (self.istep >= self.spec.timestep_limit)
def configure(self):
super(RunEnv, self).configure()
if self.verbose:
print("JOINTS")
for i in range(11):
print(i,self.osim_model.jointSet.get(i).getName())
print("\nBODIES")
for i in range(13):
print(i,self.osim_model.bodySet.get(i).getName())
print("\nMUSCLES")
for i in range(18):
print(i,self.osim_model.muscleSet.get(i).getName())
print("\nFORCES")
for i in range(26):
print(i,self.osim_model.forceSet.get(i).getName())
print("")
# for i in range(18):
# m = opensim.Thelen2003Muscle.safeDownCast(self.osim_model.muscleSet.get(i))
# m.setActivationTimeConstant(0.0001) # default 0.01
# m.setDeactivationTimeConstant(0.0001) # default 0.04
# The only joint that has to be cast
self.pelvis = opensim.PlanarJoint.safeDownCast(self.osim_model.get_joint("ground_pelvis"))
def next_obstacle(self):
obstacles = self.env_desc['obstacles']
x = self.pelvis.getCoordinate(self.STATE_PELVIS_X).getValue(self.osim_model.state)
for obstacle in obstacles:
if obstacle[0] + obstacle[2] < x:
continue
else:
ret = list(obstacle)
ret[0] = ret[0] - x
return ret
return [100,0,0]
def _step(self, action):
self.last_state = self.current_state
return super(RunEnv, self)._step(action)
def get_observation(self):
bodies = ['head', 'pelvis', 'torso', 'toes_l', 'toes_r', 'talus_l', 'talus_r']
pelvis_pos = [self.pelvis.getCoordinate(i).getValue(self.osim_model.state) for i in range(3)]
pelvis_vel = [self.pelvis.getCoordinate(i).getSpeedValue(self.osim_model.state) for i in range(3)]
jnts = ['hip_r','knee_r','ankle_r','hip_l','knee_l','ankle_l']
joint_angles = [self.osim_model.get_joint(jnts[i]).getCoordinate().getValue(self.osim_model.state) for i in range(6)]
joint_vel = [self.osim_model.get_joint(jnts[i]).getCoordinate().getSpeedValue(self.osim_model.state) for i in range(6)]
mass_pos = [self.osim_model.model.calcMassCenterPosition(self.osim_model.state)[i] for i in range(2)]
mass_vel = [self.osim_model.model.calcMassCenterVelocity(self.osim_model.state)[i] for i in range(2)]
body_transforms = [[self.osim_model.get_body(body).getTransformInGround(self.osim_model.state).p()[i] for i in range(2)] for body in bodies]
muscles = [ self.env_desc['muscles'][self.MUSCLES_PSOAS_L], self.env_desc['muscles'][self.MUSCLES_PSOAS_R] ]
# see the next obstacle
obstacle = self.next_obstacle()
# feet = [opensim.HuntCrossleyForce.safeDownCast(self.osim_model.forceSet.get(j)) for j in range(20,22)]
self.current_state = pelvis_pos + pelvis_vel + joint_angles + joint_vel + mass_pos + mass_vel + list(flatten(body_transforms)) + muscles + obstacle
return self.current_state
def create_obstacles(self):
x = 0
y = 0
r = 0.1
for i in range(self.max_obstacles):
name = i.__str__()
blockos = opensim.Body(name + '-block', 0.0001 , opensim.Vec3(0), opensim.Inertia(1,1,.0001,0,0,0) );
pj = opensim.PlanarJoint(name + '-joint',
self.osim_model.model.getGround(), # PhysicalFrame
opensim.Vec3(0, 0, 0),
opensim.Vec3(0, 0, 0),
blockos, # PhysicalFrame
opensim.Vec3(0, 0, 0),
opensim.Vec3(0, 0, 0))
self.osim_model.model.addJoint(pj)
self.osim_model.model.addBody(blockos)
block = opensim.ContactSphere(r, opensim.Vec3(0,0,0), blockos)
block.setName(name + '-contact')
self.osim_model.model.addContactGeometry(block)
force = opensim.HuntCrossleyForce()
force.setName(name + '-force')
force.addGeometry(name + '-contact')
force.addGeometry("r_heel")
force.addGeometry("l_heel")
force.addGeometry("r_toe")
force.addGeometry("l_toe")
force.setStiffness(1.0e6/r)
force.setDissipation(1e-5)
force.setStaticFriction(0.0)
force.setDynamicFriction(0.0)
force.setViscousFriction(0.0)
self.osim_model.model.addForce(force);
def clear_obstacles(self, state):
for j in range(0, self.max_obstacles):
joint_generic = self.osim_model.get_joint("%d-joint" % j)
joint = opensim.PlanarJoint.safeDownCast(joint_generic)
joint.getCoordinate(1).setValue(state, 0)
joint.getCoordinate(2).setValue(state, -0.1)
contact_generic = self.osim_model.get_contact_geometry("%d-contact" % j)
contact = opensim.ContactSphere.safeDownCast(contact_generic)
contact.setRadius(0.0001)
for i in range(3):
joint.getCoordinate(i).setLocked(state, True)
self.num_obstacles = 0
pass
def add_obstacle(self, state, x, y, r):
# set obstacle number num_obstacles
contact_generic = self.osim_model.get_contact_geometry("%d-contact" % self.num_obstacles)
contact = opensim.ContactSphere.safeDownCast(contact_generic)
contact.setRadius(r)
force_generic = self.osim_model.get_force("%d-force" % self.num_obstacles)
force = opensim.HuntCrossleyForce.safeDownCast(force_generic)
force.setStiffness(1.0e6/r)
joint_generic = self.osim_model.get_joint("%d-joint" % self.num_obstacles)
joint = opensim.PlanarJoint.safeDownCast(joint_generic)
newpos = [x,y]
for i in range(2):
joint.getCoordinate(1 + i).setLocked(state, False)
joint.getCoordinate(1 + i).setValue(state, newpos[i], False)
joint.getCoordinate(1 + i).setLocked(state, True)
self.num_obstacles += 1
pass
def generate_env(self, difficulty, seed, max_obstacles):
if seed is not None:
np.random.seed(seed) # seed the RNG if seed is provided
# obstacles
num_obstacles = 0
xs = []
ys = []
rs = []
if 0 < difficulty:
num_obstacles = min(3, max_obstacles)
xs = np.random.uniform(1.0, 5.0, num_obstacles)
ys = np.random.uniform(-0.25, 0.25, num_obstacles)
rs = [0.05 + r for r in np.random.exponential(0.05, num_obstacles)]
if 0 < difficulty and 3 < max_obstacles:
extra_obstacles = max(min(20, max_obstacles) - num_obstacles, 0)
xs = np.concatenate([xs,(np.cumsum(np.random.uniform(2.0, 4.0, extra_obstacles)) + 5)])
ys = np.concatenate([ys,np.random.uniform(-0.05, 0.25, extra_obstacles)])
rs = rs + [0.05 + r for r in np.random.exponential(0.05, extra_obstacles)]
num_obstacles = len(xs)
ys = map(lambda xy: xy[0]*xy[1], list(zip(ys, rs)))
# muscle strength
rpsoas = 1
lpsoas = 1
if difficulty >= 2:
rpsoas = 1 - np.random.normal(0, 0.1)
lpsoas = 1 - np.random.normal(0, 0.1)
rpsoas = max(0.5, rpsoas)
lpsoas = max(0.5, lpsoas)
muscles = [1] * 18
# modify only psoas
muscles[self.MUSCLES_PSOAS_R] = rpsoas
muscles[self.MUSCLES_PSOAS_L] = lpsoas
obstacles = list(zip(xs,ys,rs))
obstacles.sort()
return {
'muscles': muscles,
'obstacles': obstacles
}