-
Notifications
You must be signed in to change notification settings - Fork 430
/
ShadowHandOpenAI_FF.yaml
191 lines (178 loc) · 7.49 KB
/
ShadowHandOpenAI_FF.yaml
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
# used to create the object
name: ShadowHand
physics_engine: ${..physics_engine}
# if given, will override the device setting in gym.
env:
numEnvs: ${resolve_default:16384,${...num_envs}}
envSpacing: 0.75
episodeLength: 160 # Not used, but would be 8 sec if resetTime is not set
resetTime: 8 # Max time till reset, in seconds, if a goal wasn't achieved. Will overwrite the episodeLength if is > 0.
enableDebugVis: False
aggregateMode: 1
clipObservations: 5.0
clipActions: 1.0
stiffnessScale: 1.0
forceLimitScale: 1.0
useRelativeControl: False
dofSpeedScale: 20.0
actionsMovingAverage: 0.3
controlFrequencyInv: 3 # 20 Hz
startPositionNoise: 0.01
startRotationNoise: 0.0
resetPositionNoise: 0.01
resetRotationNoise: 0.0
resetDofPosRandomInterval: 0.2
resetDofVelRandomInterval: 0.0
# Random forces applied to the object
forceScale: 1.0
forceProbRange: [0.001, 0.1]
forceDecay: 0.99
forceDecayInterval: 0.08
distRewardScale: -10.0
rotRewardScale: 1.0
rotEps: 0.1
actionPenaltyScale: -0.0002
reachGoalBonus: 250
fallDistance: 0.24
fallPenalty: -50.0
objectType: "block" # can be block, egg or pen
observationType: "openai" # can be "openai", "full_no_vel", "full","full_state"
asymmetric_observations: True
successTolerance: 0.4
printNumSuccesses: False
maxConsecutiveSuccesses: 50
averFactor: 0.1 # running mean factor for consecutive successes calculation
asset:
assetRoot: "../assets"
assetFileName: "mjcf/open_ai_assets/hand/shadow_hand.xml"
assetFileNameBlock: "urdf/objects/cube_multicolor.urdf"
assetFileNameEgg: "mjcf/open_ai_assets/hand/egg.xml"
assetFileNamePen: "mjcf/open_ai_assets/hand/pen.xml"
# set to True if you use camera sensors in the environment
enableCameraSensors: False
task:
randomize: True
randomization_params:
frequency: 720 # Define how many simulation steps between generating new randomizations
observations:
range: [0, .002] # range for the white noise
range_correlated: [0, .001 ] # range for correlated noise, refreshed with freq `frequency`
operation: "additive"
distribution: "gaussian"
# schedule: "linear" # "constant" is to turn on noise after `schedule_steps` num steps
# schedule_steps: 40000
actions:
range: [0., .05]
range_correlated: [0, .015] # range for correlated noise, refreshed with freq `frequency`
operation: "additive"
distribution: "gaussian"
# schedule: "linear" # "linear" will linearly interpolate between no rand and max rand
# schedule_steps: 40000
sim_params:
gravity:
range: [0, 0.4]
operation: "additive"
distribution: "gaussian"
# schedule: "linear" # "linear" will linearly interpolate between no rand and max rand
# schedule_steps: 40000
actor_params:
hand:
color: True
tendon_properties:
damping:
range: [0.3, 3.0]
operation: "scaling"
distribution: "loguniform"
# schedule: "linear" # "linear" will scale the current random sample by `min(current num steps, schedule_steps) / schedule_steps`
# schedule_steps: 30000
stiffness:
range: [0.75, 1.5]
operation: "scaling"
distribution: "loguniform"
# schedule: "linear" # "linear" will scale the current random sample by `min(current num steps, schedule_steps) / schedule_steps`
# schedule_steps: 30000
dof_properties:
damping:
range: [0.3, 3.0]
operation: "scaling"
distribution: "loguniform"
# schedule: "linear" # "linear" will scale the current random sample by `min(current num steps, schedule_steps) / schedule_steps`
# schedule_steps: 30000
stiffness:
range: [0.75, 1.5]
operation: "scaling"
distribution: "loguniform"
# schedule: "linear" # "linear" will scale the current random sample by `min(current num steps, schedule_steps) / schedule_steps`
# schedule_steps: 30000
lower:
range: [0, 0.01]
operation: "additive"
distribution: "gaussian"
# schedule: "linear" # "linear" will scale the current random sample by `min(current num steps, schedule_steps) / schedule_steps`
# schedule_steps: 30000
upper:
range: [0, 0.01]
operation: "additive"
distribution: "gaussian"
# schedule: "linear" # "linear" will scale the current random sample by `min(current num steps, schedule_steps) / schedule_steps`
# schedule_steps: 30000
rigid_body_properties:
mass:
range: [0.5, 1.5]
operation: "scaling"
distribution: "uniform"
setup_only: True # Property will only be randomized once before simulation is started. See Domain Randomization Documentation for more info.
# schedule: "linear" # "linear" will scale the current random sample by `min(current num steps, schedule_steps) / schedule_steps`
# schedule_steps: 30000
rigid_shape_properties:
friction:
num_buckets: 250
range: [0.7, 1.3]
operation: "scaling"
distribution: "uniform"
# schedule: "linear" # "linear" will scale the current random sample by `min(current num steps, schedule_steps) / schedule_steps`
# schedule_steps: 30000
object:
scale:
range: [0.95, 1.05]
operation: "scaling"
distribution: "uniform"
setup_only: True # Property will only be randomized once before simulation is started. See Domain Randomization Documentation for more info.
# schedule: "linear" # "linear" will scale the current random sample by ``min(current num steps, schedule_steps) / schedule_steps`
# schedule_steps: 30000
rigid_body_properties:
mass:
range: [0.5, 1.5]
operation: "scaling"
distribution: "uniform"
setup_only: True # Property will only be randomized once before simulation is started. See Domain Randomization Documentation for more info.
# schedule: "linear" # "linear" will scale the current random sample by ``min(current num steps, schedule_steps) / schedule_steps`
# schedule_steps: 30000
rigid_shape_properties:
friction:
num_buckets: 250
range: [0.7, 1.3]
operation: "scaling"
distribution: "uniform"
# schedule: "linear" # "linear" will scale the current random sample by `min(current num steps, schedule_steps) / schedule_steps`
# schedule_steps: 30000
sim:
dt: 0.01667 # 1/60
substeps: 2
up_axis: "z"
use_gpu_pipeline: ${eq:${...pipeline},"gpu"}
gravity: [0.0, 0.0, -9.81]
physx:
num_threads: ${....num_threads}
solver_type: ${....solver_type}
use_gpu: ${contains:"cuda",${....sim_device}} # set to False to run on CPU
num_position_iterations: 8
num_velocity_iterations: 0
max_gpu_contact_pairs: 8388608 # 8*1024*1024
num_subscenes: ${....num_subscenes}
contact_offset: 0.002
rest_offset: 0.0
bounce_threshold_velocity: 0.2
max_depenetration_velocity: 1000.0
default_buffer_size_multiplier: 5.0
contact_collection: 0 # 0: CC_NEVER (don't collect contact info), 1: CC_LAST_SUBSTEP (collect only contacts on last substep), 2: CC_ALL_SUBSTEPS (broken - do not use!)