FazBrowse GitHub Viewer
|
Trending
|
URL:
|
Home
Tools:
[Download Repo ZIP]
[View Raw Code]
[Original HTTPS Page]
dm_control/dm_control/rl/control.py at master · StephennFernandes/dm_control · GitHub
StephennFernandes
/
dm_control
Public
forked from
google-deepmind/dm_control
Notifications
You must be signed in to change notification settings
Fork
0
Star
1
Code
Pull requests
0
Actions
Projects
Security and quality
0
Insights
Additional navigation options
Code
Pull requests
Actions
Projects
Security and quality
Insights
Expand file tree
Breadcrumbs
dm_control
/
dm_control
/
rl
/
control.py
Copy path
More file actions
More file actions
Latest commit
History
History
History
402 lines (312 loc) · 12.3 KB
Breadcrumbs
dm_control
/
dm_control
/
rl
/
control.py
Copy path
File metadata and controls
402 lines (312 loc) · 12.3 KB
Raw
Copy raw file
Download raw file
Open symbols panel
Edit and raw actions
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
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
# Copyright 2017 The dm_control Authors.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# ============================================================================
"""A dm_env.Environment subclass for control-specific environments."""
from
__future__
import
absolute_import
from
__future__
import
division
from
__future__
import
print_function
import
abc
import
collections
import
contextlib
import
dm_env
from
dm_env
import
specs
import
numpy
as
np
import
six
from
six
.
moves
import
range
FLAT_OBSERVATION_KEY
=
'observations'
class
Environment
(
dm_env
.
Environment
):
"""Class for physics-based reinforcement learning environments."""
def
__init__
(
self
,
physics
,
task
,
time_limit
=
float
(
'inf'
),
control_timestep
=
None
,
n_sub_steps
=
None
,
flat_observation
=
False
):
"""Initializes a new `Environment`.
Args:
physics: Instance of `Physics`.
task: Instance of `Task`.
time_limit: Optional `int`, maximum time for each episode in seconds. By
default this is set to infinite.
control_timestep: Optional control time-step, in seconds.
n_sub_steps: Optional number of physical time-steps in one control
time-step, aka "action repeats". Can only be supplied if
`control_timestep` is not specified.
flat_observation: If True, observations will be flattened and concatenated
into a single numpy array.
Raises:
ValueError: If both `n_sub_steps` and `control_timestep` are supplied.
"""
self
.
_task
=
task
self
.
_physics
=
physics
self
.
_flat_observation
=
flat_observation
if
n_sub_steps
is
not
None
and
control_timestep
is
not
None
:
raise
ValueError
(
'Both n_sub_steps and control_timestep were supplied.'
)
elif
n_sub_steps
is
not
None
:
self
.
_n_sub_steps
=
n_sub_steps
elif
control_timestep
is
not
None
:
self
.
_n_sub_steps
=
compute_n_steps
(
control_timestep
,
self
.
_physics
.
timestep
())
else
:
self
.
_n_sub_steps
=
1
if
time_limit
==
float
(
'inf'
):
self
.
_step_limit
=
float
(
'inf'
)
else
:
self
.
_step_limit
=
time_limit
/
(
self
.
_physics
.
timestep
()
*
self
.
_n_sub_steps
)
self
.
_step_count
=
0
self
.
_reset_next_step
=
True
def
reset
(
self
):
"""Starts a new episode and returns the first `TimeStep`."""
self
.
_reset_next_step
=
False
self
.
_step_count
=
0
with
self
.
_physics
.
reset_context
():
self
.
_task
.
initialize_episode
(
self
.
_physics
)
observation
=
self
.
_task
.
get_observation
(
self
.
_physics
)
if
self
.
_flat_observation
:
observation
=
flatten_observation
(
observation
)
return
dm_env
.
TimeStep
(
step_type
=
dm_env
.
StepType
.
FIRST
,
reward
=
None
,
discount
=
None
,
observation
=
observation
)
def
step
(
self
,
action
):
"""Updates the environment using the action and returns a `TimeStep`."""
if
self
.
_reset_next_step
:
return
self
.
reset
()
self
.
_task
.
before_step
(
action
,
self
.
_physics
)
for
_
in
range
(
self
.
_n_sub_steps
):
self
.
_physics
.
step
()
self
.
_task
.
after_step
(
self
.
_physics
)
reward
=
self
.
_task
.
get_reward
(
self
.
_physics
)
observation
=
self
.
_task
.
get_observation
(
self
.
_physics
)
if
self
.
_flat_observation
:
observation
=
flatten_observation
(
observation
)
self
.
_step_count
+=
1
if
self
.
_step_count
>=
self
.
_step_limit
:
discount
=
1.0
else
:
discount
=
self
.
_task
.
get_termination
(
self
.
_physics
)
episode_over
=
discount
is
not
None
if
episode_over
:
self
.
_reset_next_step
=
True
return
dm_env
.
TimeStep
(
dm_env
.
StepType
.
LAST
,
reward
,
discount
,
observation
)
else
:
return
dm_env
.
TimeStep
(
dm_env
.
StepType
.
MID
,
reward
,
1.0
,
observation
)
def
action_spec
(
self
):
"""Returns the action specification for this environment."""
return
self
.
_task
.
action_spec
(
self
.
_physics
)
def
step_spec
(
self
):
"""May return a specification for the values returned by `step`."""
return
self
.
_task
.
step_spec
(
self
.
_physics
)
def
observation_spec
(
self
):
"""Returns the observation specification for this environment.
Infers the spec from the observation, unless the Task implements the
`observation_spec` method.
Returns:
An dict mapping observation name to `ArraySpec` containing observation
shape and dtype.
"""
try
:
return
self
.
_task
.
observation_spec
(
self
.
_physics
)
except
NotImplementedError
:
observation
=
self
.
_task
.
get_observation
(
self
.
_physics
)
if
self
.
_flat_observation
:
observation
=
flatten_observation
(
observation
)
return
_spec_from_observation
(
observation
)
@
property
def
physics
(
self
):
return
self
.
_physics
@
property
def
task
(
self
):
return
self
.
_task
def
control_timestep
(
self
):
"""Returns the interval between agent actions in seconds."""
return
self
.
physics
.
timestep
()
*
self
.
_n_sub_steps
def
compute_n_steps
(
control_timestep
,
physics_timestep
,
tolerance
=
1e-8
):
"""Returns the number of physics timesteps in a single control timestep.
Args:
control_timestep: Control time-step, should be an integer multiple of the
physics timestep.
physics_timestep: The time-step of the physics simulation.
tolerance: Optional tolerance value for checking if `physics_timestep`
divides `control_timestep`.
Returns:
The number of physics timesteps in a single control timestep.
Raises:
ValueError: If `control_timestep` is smaller than `physics_timestep` or if
`control_timestep` is not an integer multiple of `physics_timestep`.
"""
if
control_timestep
<
physics_timestep
:
raise
ValueError
(
'Control timestep ({}) cannot be smaller than physics timestep ({}).'
.
format
(
control_timestep
,
physics_timestep
))
if
abs
((
control_timestep
/
physics_timestep
-
round
(
control_timestep
/
physics_timestep
)))
>
tolerance
:
raise
ValueError
(
'Control timestep ({}) must be an integer multiple of physics timestep '
'({})'
.
format
(
control_timestep
,
physics_timestep
))
return
int
(
round
(
control_timestep
/
physics_timestep
))
def
_spec_from_observation
(
observation
):
result
=
collections
.
OrderedDict
()
for
key
,
value
in
six
.
iteritems
(
observation
):
result
[
key
]
=
specs
.
Array
(
value
.
shape
,
value
.
dtype
,
name
=
key
)
return
result
# Base class definitions for objects supplied to Environment.
@
six
.
add_metaclass
(
abc
.
ABCMeta
)
class
Physics
(
object
):
"""Simulates a physical environment."""
@
abc
.
abstractmethod
def
step
(
self
,
n_sub_steps
=
1
):
"""Updates the simulation state.
Args:
n_sub_steps: Optional number of times to repeatedly update the simulation
state. Defaults to 1.
"""
@
abc
.
abstractmethod
def
time
(
self
):
"""Returns the elapsed simulation time in seconds."""
@
abc
.
abstractmethod
def
timestep
(
self
):
"""Returns the simulation timestep."""
def
set_control
(
self
,
control
):
"""Sets the control signal for the actuators."""
raise
NotImplementedError
(
'set_control is not supported.'
)
@
contextlib
.
contextmanager
def
reset_context
(
self
):
"""Context manager for resetting the simulation state.
Sets the internal simulation to a default state when entering the block.
```python
with physics.reset_context():
# Set joint and object positions.
physics.step()
```
Yields:
The `Physics` instance.
"""
try
:
self
.
reset
()
except
PhysicsError
:
pass
yield
self
self
.
after_reset
()
@
abc
.
abstractmethod
def
reset
(
self
):
"""Resets internal variables of the physics simulation."""
@
abc
.
abstractmethod
def
after_reset
(
self
):
"""Runs after resetting internal variables of the physics simulation."""
def
check_divergence
(
self
):
"""Raises a `PhysicsError` if the simulation state is divergent.
The default implementation is a no-op.
"""
class
PhysicsError
(
RuntimeError
):
"""Raised if the state of the physics simulation becomes divergent."""
@
six
.
add_metaclass
(
abc
.
ABCMeta
)
class
Task
(
object
):
"""Defines a task in a `control.Environment`."""
@
abc
.
abstractmethod
def
initialize_episode
(
self
,
physics
):
"""Sets the state of the environment at the start of each episode.
Called by `control.Environment` at the start of each episode *within*
`physics.reset_context()` (see the documentation for `base.Physics`).
Args:
physics: Instance of `Physics`.
"""
@
abc
.
abstractmethod
def
before_step
(
self
,
action
,
physics
):
"""Updates the task from the provided action.
Called by `control.Environment` before stepping the physics engine.
Args:
action: numpy array or array-like action values, or a nested structure of
such arrays. Should conform to the specification returned by
`self.action_spec(physics)`.
physics: Instance of `Physics`.
"""
def
after_step
(
self
,
physics
):
"""Optional method to update the task after the physics engine has stepped.
Called by `control.Environment` after stepping the physics engine and before
`control.Environment` calls `get_observation, `get_reward` and
`get_termination`.
The default implementation is a no-op.
Args:
physics: Instance of `Physics`.
"""
@
abc
.
abstractmethod
def
action_spec
(
self
,
physics
):
"""Returns a specification describing the valid actions for this task.
Args:
physics: Instance of `Physics`.
Returns:
A `BoundedArraySpec`, or a nested structure containing `BoundedArraySpec`s
that describe the shapes, dtypes and elementwise lower and upper bounds
for the action array(s) passed to `self.step`.
"""
def
step_spec
(
self
,
physics
):
"""Returns a specification describing the time_step for this task.
Args:
physics: Instance of `Physics`.
Returns:
A `BoundedArraySpec`, or a nested structure containing `BoundedArraySpec`s
that describe the shapes, dtypes and elementwise lower and upper bounds
for the array(s) returned by `self.step`.
"""
raise
NotImplementedError
()
@
abc
.
abstractmethod
def
get_observation
(
self
,
physics
):
"""Returns an observation from the environment.
Args:
physics: Instance of `Physics`.
"""
@
abc
.
abstractmethod
def
get_reward
(
self
,
physics
):
"""Returns a reward from the environment.
Args:
physics: Instance of `Physics`.
"""
def
get_termination
(
self
,
physics
):
"""If the episode should end, returns a final discount, otherwise None."""
def
observation_spec
(
self
,
physics
):
"""Optional method that returns the observation spec.
If not implemented, the Environment infers the spec from the observation.
Args:
physics: Instance of `Physics`.
Returns:
A dict mapping observation name to `ArraySpec` containing observation
shape and dtype.
"""
raise
NotImplementedError
()
def
flatten_observation
(
observation
,
output_key
=
FLAT_OBSERVATION_KEY
):
"""Flattens multiple observation arrays into a single numpy array.
Args:
observation: A mutable mapping from observation names to numpy arrays.
output_key: The key for the flattened observation array in the output.
Returns:
A mutable mapping of the same type as `observation`. This will contain a
single key-value pair consisting of `output_key` and the flattened
and concatenated observation array.
Raises:
ValueError: If `observation` is not a `collections.MutableMapping`.
"""
if
not
isinstance
(
observation
,
collections
.
MutableMapping
):
raise
ValueError
(
'Can only flatten dict-like observations.'
)
if
isinstance
(
observation
,
collections
.
OrderedDict
):
keys
=
six
.
iterkeys
(
observation
)
else
:
# Keep a consistent ordering for other mappings.
keys
=
sorted
(
six
.
iterkeys
(
observation
))
observation_arrays
=
[
observation
[
key
].
ravel
()
for
key
in
keys
]
return
type
(
observation
)([(
output_key
,
np
.
concatenate
(
observation_arrays
))])
Back
|
FazBrowse Home
|
New Git URL