BIue-Jay commited on
Commit
b04edfb
·
verified ·
1 Parent(s): 6e3119c

Add files using upload-large-folder tool

Browse files
Files changed (48) hide show
  1. README.md +24 -0
  2. a18-pour-water/epoch-010/assets/h2_a18/norm_stats.json +636 -0
  3. a18-pour-water/epoch-010/metadata.pt +3 -0
  4. a18-pour-water/epoch-010/model.safetensors +3 -0
  5. a18-pour-water/epoch-010/protocol.json +56 -0
  6. a18-pour-water/epoch-010/training_plan.json +23 -0
  7. a18-pour-water/epoch-020/assets/h2_a18/norm_stats.json +636 -0
  8. a18-pour-water/epoch-020/metadata.pt +3 -0
  9. a18-pour-water/epoch-020/model.safetensors +3 -0
  10. a18-pour-water/epoch-020/protocol.json +56 -0
  11. a18-pour-water/epoch-020/training_plan.json +23 -0
  12. a3-fridge-open/epoch-010/assets/g1_a3/norm_stats.json +504 -0
  13. a3-fridge-open/epoch-010/metadata.pt +3 -0
  14. a3-fridge-open/epoch-010/model.safetensors +3 -0
  15. a3-fridge-open/epoch-010/protocol.json +56 -0
  16. a3-fridge-open/epoch-010/training_plan.json +23 -0
  17. a3-fridge-open/epoch-020/assets/g1_a3/norm_stats.json +504 -0
  18. a3-fridge-open/epoch-020/metadata.pt +3 -0
  19. a3-fridge-open/epoch-020/model.safetensors +3 -0
  20. a3-fridge-open/epoch-020/protocol.json +56 -0
  21. a3-fridge-open/epoch-020/training_plan.json +23 -0
  22. a7-cabinet-close/epoch-010/assets/g1_a7/norm_stats.json +504 -0
  23. a7-cabinet-close/epoch-010/metadata.pt +3 -0
  24. a7-cabinet-close/epoch-010/model.safetensors +3 -0
  25. a7-cabinet-close/epoch-010/protocol.json +56 -0
  26. a7-cabinet-close/epoch-010/training_plan.json +23 -0
  27. a7-cabinet-close/epoch-020/assets/g1_a7/norm_stats.json +504 -0
  28. a7-cabinet-close/epoch-020/metadata.pt +3 -0
  29. a7-cabinet-close/epoch-020/model.safetensors +3 -0
  30. a7-cabinet-close/epoch-020/protocol.json +56 -0
  31. a7-cabinet-close/epoch-020/training_plan.json +23 -0
  32. b6-pedal-bin/epoch-010/assets/g1_b6/norm_stats.json +504 -0
  33. b6-pedal-bin/epoch-010/metadata.pt +3 -0
  34. b6-pedal-bin/epoch-010/model.safetensors +3 -0
  35. b6-pedal-bin/epoch-010/protocol.json +56 -0
  36. b6-pedal-bin/epoch-010/training_plan.json +23 -0
  37. b6-pedal-bin/epoch-020/assets/g1_b6/norm_stats.json +504 -0
  38. b6-pedal-bin/epoch-020/metadata.pt +3 -0
  39. b6-pedal-bin/epoch-020/model.safetensors +3 -0
  40. b6-pedal-bin/epoch-020/protocol.json +56 -0
  41. b6-pedal-bin/epoch-020/training_plan.json +23 -0
  42. code/hb_pi05/hb_aug_v2.py +163 -0
  43. code/hb_pi05/hb_g1_config.py +144 -0
  44. code/hb_pi05/hb_g1_data.py +264 -0
  45. code/hb_pi05/hb_init_weights.py +173 -0
  46. code/hb_pi05/hb_prepare.py +278 -0
  47. code/hb_pi05/hb_train.py +153 -0
  48. code/hb_pi05/verify_cpu.py +222 -0
README.md ADDED
@@ -0,0 +1,24 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ tags:
3
+ - robotics
4
+ - humanoid
5
+ - imitation-learning
6
+ ---
7
+ # pi0.5 — Humanoid Benchmark
8
+
9
+ pi0.5 (openpi) checkpoints fine-tuned on Humanoid Benchmark tasks (2026-11-10 training round).
10
+
11
+ | Folder | Task | Robot |
12
+ |---|---|---|
13
+ | `b6-pedal-bin/` | Pick up a chip barrel, step on the pedal bin, drop it in | Unitree G1 |
14
+ | `a3-fridge-open/` | Open the fridge door | Unitree G1 |
15
+ | `a7-cabinet-close/` | Close the cabinet door | Unitree G1 |
16
+ | `a18-pour-water/` | Pour water from a pitcher into a glass | H2 |
17
+
18
+ Each task folder has `epoch-010/` and `epoch-020/` (100 demonstrations per task, 20 epochs, saved at epoch 10 and 20).
19
+ Every epoch folder is self-contained: weights, normalization statistics, training config and `protocol.json`
20
+ (dimensions, input/output layout, step counts, file list). Optimizer / resume state is not included.
21
+ `code/` holds the adapter code used for training and inference.
22
+
23
+ Training used the Humanoid Benchmark online data augmentation spec v2 (photometric, geometric, sensor and lighting
24
+ augmentation plus proprioception noise/delay/drop), with the model's own image augmentation turned off.
a18-pour-water/epoch-010/assets/h2_a18/norm_stats.json ADDED
@@ -0,0 +1,636 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "norm_stats": {
3
+ "state": {
4
+ "mean": [
5
+ -0.02828288823366165,
6
+ 0.03546403720974922,
7
+ 0.03124321810901165,
8
+ 0.16529245674610138,
9
+ 0.033904630690813065,
10
+ -0.1248306930065155,
11
+ -0.020945074036717415,
12
+ -0.03086675889790058,
13
+ 0.013680738396942616,
14
+ 0.20139853656291962,
15
+ -0.0015518380096182227,
16
+ -0.1498299092054367,
17
+ 0.00765480799600482,
18
+ -0.01444979291409254,
19
+ 0.030299680307507515,
20
+ 0.1457539200782776,
21
+ 0.08196060359477997,
22
+ -0.04197724163532257,
23
+ 1.0582942962646484,
24
+ 0.004078589379787445,
25
+ 0.10490317642688751,
26
+ 0.11859612911939621,
27
+ -0.19652463495731354,
28
+ -0.048440903425216675,
29
+ -0.06604290008544922,
30
+ 0.31633737683296204,
31
+ 0.8490549325942993,
32
+ 0.16632796823978424,
33
+ -0.20058786869049072,
34
+ 0.5188571214675903,
35
+ -0.0062464503571391106,
36
+ 0.0012028953060507774,
37
+ 0.0012025771429762244,
38
+ 0.000013984431461722124,
39
+ 0.0012024634052067995,
40
+ 0.008900434710085392,
41
+ -0.0008743949001654983,
42
+ -0.0008745864615775645,
43
+ 0.001202479936182499,
44
+ -0.0008735220180824399,
45
+ -0.09596802294254303,
46
+ 0.00023934645287226886,
47
+ 0.0002393399045104161,
48
+ -0.0008724569343030453,
49
+ 0.0002393364702584222,
50
+ -0.010859426110982895,
51
+ 0.00004498684938880615,
52
+ 0.00004498622729443014,
53
+ 0.00023933025659061968,
54
+ 0.00004498567068367265,
55
+ 0.0009174197912216187,
56
+ 0.000044984764826949686,
57
+ 0.000024329165171366185,
58
+ 0.1904839128255844,
59
+ 0.37486860156059265,
60
+ 0.020827073603868484,
61
+ 0.3586229681968689,
62
+ 0.3983735144138336,
63
+ -0.025473948568105698,
64
+ -0.07296764850616455,
65
+ 0.2999779284000397,
66
+ -0.004135874565690756,
67
+ -0.17300280928611755,
68
+ 0.9253543019294739,
69
+ 0.8886311054229736,
70
+ 0.08327575027942657,
71
+ 0.8554991483688354,
72
+ 0.25384166836738586,
73
+ 0.677445650100708,
74
+ 0.644685685634613,
75
+ 0.9370030164718628,
76
+ 0.15102075040340424,
77
+ -0.003838402684777975,
78
+ 0.5436624884605408,
79
+ 0.4556223452091217,
80
+ -0.023036951199173927,
81
+ -0.010215787217020988,
82
+ -0.9998420476913452
83
+ ],
84
+ "std": [
85
+ 0.03512008860707283,
86
+ 0.01740800216794014,
87
+ 0.03788851574063301,
88
+ 0.03280680626630783,
89
+ 0.022302838042378426,
90
+ 0.01713656447827816,
91
+ 0.03212040290236473,
92
+ 0.016956482082605362,
93
+ 0.05338956415653229,
94
+ 0.027900811284780502,
95
+ 0.012169546447694302,
96
+ 0.015364138409495354,
97
+ 0.01877790130674839,
98
+ 0.012579463422298431,
99
+ 0.021866735070943832,
100
+ 0.03327784687280655,
101
+ 0.017236528918147087,
102
+ 0.11609876155853271,
103
+ 0.05033324658870697,
104
+ 0.05683313310146332,
105
+ 0.10001008212566376,
106
+ 0.06167083978652954,
107
+ 0.35286715626716614,
108
+ 0.09832895547151566,
109
+ 0.23262782394886017,
110
+ 0.4843495786190033,
111
+ 0.6070270538330078,
112
+ 0.17739585041999817,
113
+ 0.32976317405700684,
114
+ 0.2399081438779831,
115
+ 0.10201188921928406,
116
+ 0.0003750584146473557,
117
+ 0.00037507538218051195,
118
+ 0.000020995448721805587,
119
+ 0.0003750614414457232,
120
+ 0.024341240525245667,
121
+ 0.00038038159254938364,
122
+ 0.0003799603146035224,
123
+ 0.00037506569060496986,
124
+ 0.00038039451465010643,
125
+ 0.01200117077678442,
126
+ 0.00007667305180802941,
127
+ 0.00007667495083296672,
128
+ 0.00038073238101787865,
129
+ 0.00007667316822335124,
130
+ 0.013414953835308552,
131
+ 0.000014869372535031289,
132
+ 0.000014868522157485131,
133
+ 0.0000766750454204157,
134
+ 0.000014868902326270472,
135
+ 0.00865352526307106,
136
+ 0.000014869230653857812,
137
+ 0.00003821943755610846,
138
+ 0.1619565486907959,
139
+ 0.3236274719238281,
140
+ 0.014654714614152908,
141
+ 0.31027916073799133,
142
+ 0.3515927493572235,
143
+ 0.022043032571673393,
144
+ 0.05582857504487038,
145
+ 0.25976499915122986,
146
+ 0.004044965375214815,
147
+ 0.07659266144037247,
148
+ 0.7724840641021729,
149
+ 0.7353335618972778,
150
+ 0.06149245798587799,
151
+ 0.7260345220565796,
152
+ 0.2404094636440277,
153
+ 0.5688204765319824,
154
+ 0.5405519604682922,
155
+ 0.7903794050216675,
156
+ 0.1275968998670578,
157
+ 0.00503132538869977,
158
+ 0.45670056343078613,
159
+ 0.43200525641441345,
160
+ 0.01495918445289135,
161
+ 0.009642813354730606,
162
+ 0.0
163
+ ],
164
+ "q01": [
165
+ -0.09696375873552374,
166
+ 0.0007942055537837334,
167
+ -0.051792578461062305,
168
+ 0.13033591651223178,
169
+ -0.01713594465283505,
170
+ -0.15955360559398968,
171
+ -0.0964198788970706,
172
+ -0.06363997625562103,
173
+ -0.1000344696582508,
174
+ 0.1482597910888919,
175
+ -0.027435287981779383,
176
+ -0.18897574493706598,
177
+ -0.044199422308726685,
178
+ -0.05024038314818473,
179
+ -0.01132426475313017,
180
+ 0.08324008762353496,
181
+ 0.038280332696323874,
182
+ -0.3242681729192237,
183
+ 0.9701593956163327,
184
+ -0.13880584287678877,
185
+ -0.09421699412132398,
186
+ 0.020241137548000453,
187
+ -0.7021140373221906,
188
+ -0.27443564700923767,
189
+ -0.5346578447649877,
190
+ -0.4259114304386858,
191
+ -0.08661681087593327,
192
+ -0.12122693202118925,
193
+ -0.9081178965514068,
194
+ 0.005415167589460248,
195
+ -0.2458772536665035,
196
+ 0.00006636888889494642,
197
+ 0.0000663710596573749,
198
+ -1.0299070379400978e-6,
199
+ 0.00006638123974712031,
200
+ -0.051145545630711886,
201
+ -0.0013889520672031631,
202
+ -0.0013888838999733162,
203
+ 0.00006639188308495093,
204
+ -0.0013883197191158069,
205
+ -0.1278356132528748,
206
+ 0.000012710355428494887,
207
+ 0.000012642206731161422,
208
+ -0.0013877263354788194,
209
+ 0.000012643952976322337,
210
+ -0.053515475887209486,
211
+ 2.275902296722304e-6,
212
+ 2.276176049376893e-6,
213
+ 0.00001264573177261127,
214
+ 2.2764722897171513e-6,
215
+ -0.021655506657829896,
216
+ 2.276778616162694e-6,
217
+ -0.0000463191339635946,
218
+ -0.0008273905561459105,
219
+ -0.008519364052306896,
220
+ 0.000015183388484773017,
221
+ -0.007527138107404751,
222
+ 0.009861442021763134,
223
+ -0.06757076855718339,
224
+ -0.13470218816765142,
225
+ -0.005825813357401111,
226
+ -0.010244786777927912,
227
+ -0.256434232508372,
228
+ -0.00025278038957425066,
229
+ -0.000052326070924881094,
230
+ -0.0005493609337659858,
231
+ -0.0002497415102234446,
232
+ -0.0034235131458670207,
233
+ -0.00019695640924990575,
234
+ -0.00011751209532239902,
235
+ -0.0002919652158541938,
236
+ 1.0082914067513594e-8,
237
+ -0.012793561069488308,
238
+ -0.00018825285256620066,
239
+ -0.000057134988586219314,
240
+ -0.05098403841318886,
241
+ -0.03447218551372179,
242
+ -0.9999983127043873
243
+ ],
244
+ "q99": [
245
+ 0.022343595302178443,
246
+ 0.08494462532764574,
247
+ 0.12511245712753394,
248
+ 0.2416719058894359,
249
+ 0.0677620012453274,
250
+ -0.08898747787709982,
251
+ 0.02318780476657832,
252
+ -0.0015199510267689398,
253
+ 0.14604715039697608,
254
+ 0.2637284779196589,
255
+ 0.02539288635049395,
256
+ -0.11248975168235031,
257
+ 0.045069882762597095,
258
+ 0.010824615722864028,
259
+ 0.06618172762760902,
260
+ 0.212699701484278,
261
+ 0.1207836883960253,
262
+ 0.17330300987310265,
263
+ 1.1781945639273799,
264
+ 0.15023529210570086,
265
+ 0.2845899072977251,
266
+ 0.24516796460341958,
267
+ 0.5740916040221649,
268
+ 0.2010916694726258,
269
+ 0.4289175018925724,
270
+ 1.0899582403284118,
271
+ 1.6852502702105407,
272
+ 0.43534713681491477,
273
+ 0.6240539385241481,
274
+ 0.6550094590556405,
275
+ 0.11099284940619253,
276
+ 0.0016734303027044354,
277
+ 0.0016734576745192527,
278
+ 0.00007065863147983712,
279
+ 0.0016737097039632794,
280
+ 0.04393054750393248,
281
+ -0.00003521824024034785,
282
+ -0.000036440479353125766,
283
+ 0.001673639154479647,
284
+ -0.00003455918041791865,
285
+ -0.07552439234949913,
286
+ 0.00033051224737817186,
287
+ 0.0003304989688443708,
288
+ -0.00003296538223554121,
289
+ 0.00033056630809084984,
290
+ 0.002957078129294119,
291
+ 0.00006171814166910019,
292
+ 0.00006171964270592179,
293
+ 0.00033056845467495713,
294
+ 0.00006172230211941015,
295
+ 0.02593200647046453,
296
+ 0.00006172569891158286,
297
+ 0.00011897972866163018,
298
+ 0.3787801509389085,
299
+ 0.7209945422769666,
300
+ 0.0407652068077642,
301
+ 0.6618924614843803,
302
+ 0.7517948510379678,
303
+ 0.005806987208462441,
304
+ 0.002885183540840164,
305
+ 0.570767374535971,
306
+ 0.0014458511145520186,
307
+ -0.06102761673970994,
308
+ 1.7102710429866648,
309
+ 1.6174167523095186,
310
+ 0.1397007980669106,
311
+ 1.554299900071352,
312
+ 0.5039048781097059,
313
+ 1.239142416988845,
314
+ 1.1760776755933253,
315
+ 1.7112652279961076,
316
+ 0.2737724787321517,
317
+ 0.008886616448485709,
318
+ 0.988431640873346,
319
+ 0.8869660620081917,
320
+ 0.0035224365899956395,
321
+ 0.016511748692902894,
322
+ -0.9981961741900639
323
+ ]
324
+ },
325
+ "actions": {
326
+ "mean": [
327
+ -0.039711859077215195,
328
+ 0.042044274508953094,
329
+ 0.031410254538059235,
330
+ 0.1572873443365097,
331
+ 0.05045900121331215,
332
+ -0.1497725248336792,
333
+ -0.006279041990637779,
334
+ -0.02705513872206211,
335
+ 0.014146559871733189,
336
+ 0.21941550076007843,
337
+ -0.025561144575476646,
338
+ -0.030189640820026398,
339
+ 0.009197989478707314,
340
+ -0.049546219408512115,
341
+ 0.03526940196752548,
342
+ 0.14168091118335724,
343
+ 0.11557798087596893,
344
+ -0.04095250740647316,
345
+ 0.9854463934898376,
346
+ 0.008960100822150707,
347
+ 0.08984862267971039,
348
+ 0.12244174629449844,
349
+ -0.19101494550704956,
350
+ -0.050739310681819916,
351
+ -0.06203479319810867,
352
+ 0.30884090065956116,
353
+ 0.8466822504997253,
354
+ 0.16446106135845184,
355
+ -0.19862252473831177,
356
+ 0.4859173893928528,
357
+ -0.0057493275962769985,
358
+ 0.0,
359
+ 0.0,
360
+ 0.0,
361
+ 0.0,
362
+ 0.0,
363
+ 0.0,
364
+ 0.0,
365
+ 0.0,
366
+ 0.0,
367
+ 0.0,
368
+ 0.0,
369
+ 0.0,
370
+ 0.0,
371
+ 0.0,
372
+ 0.0,
373
+ 0.0,
374
+ 0.0,
375
+ 0.0,
376
+ 0.0,
377
+ 0.0,
378
+ 0.0,
379
+ 0.0,
380
+ 0.19399771094322205,
381
+ 0.4265739321708679,
382
+ 0.023019764572381973,
383
+ 0.3584686815738678,
384
+ 0.40135377645492554,
385
+ 0.045260727405548096,
386
+ -0.176839679479599,
387
+ 0.31074193120002747,
388
+ -0.009844515472650528,
389
+ -0.11006658524274826,
390
+ 0.9708828926086426,
391
+ 0.974152684211731,
392
+ 0.07610955834388733,
393
+ 0.8584804534912109,
394
+ 0.2781544029712677,
395
+ 0.6949679851531982,
396
+ 0.6698334813117981,
397
+ 0.9649102091789246,
398
+ 0.15337009727954865,
399
+ -0.0038274202961474657,
400
+ 0.5546799898147583,
401
+ 0.4666762948036194
402
+ ],
403
+ "std": [
404
+ 0.039720017462968826,
405
+ 0.03597322106361389,
406
+ 0.038342930376529694,
407
+ 0.04752495139837265,
408
+ 0.02657449059188366,
409
+ 0.05687854439020157,
410
+ 0.04023086652159691,
411
+ 0.03455903008580208,
412
+ 0.05493085831403732,
413
+ 0.042149096727371216,
414
+ 0.012635765597224236,
415
+ 0.07036220282316208,
416
+ 0.015654711052775383,
417
+ 0.03608144074678421,
418
+ 0.03046419285237789,
419
+ 0.05074617266654968,
420
+ 0.020575977861881256,
421
+ 0.12018261849880219,
422
+ 0.1278754025697708,
423
+ 0.0572555735707283,
424
+ 0.10964718461036682,
425
+ 0.0703325942158699,
426
+ 0.3559415340423584,
427
+ 0.09842006117105484,
428
+ 0.23309923708438873,
429
+ 0.47085040807724,
430
+ 0.6077632308006287,
431
+ 0.18328437209129333,
432
+ 0.33357715606689453,
433
+ 0.22189438343048096,
434
+ 0.10337754338979721,
435
+ 0.0,
436
+ 0.0,
437
+ 0.0,
438
+ 0.0,
439
+ 0.0,
440
+ 0.0,
441
+ 0.0,
442
+ 0.0,
443
+ 0.0,
444
+ 0.0,
445
+ 0.0,
446
+ 0.0,
447
+ 0.0,
448
+ 0.0,
449
+ 0.0,
450
+ 0.0,
451
+ 0.0,
452
+ 0.0,
453
+ 0.0,
454
+ 0.0,
455
+ 0.0,
456
+ 0.0,
457
+ 0.1725616753101349,
458
+ 0.3615170419216156,
459
+ 0.018322760239243507,
460
+ 0.31214597821235657,
461
+ 0.3793746829032898,
462
+ 0.03442889451980591,
463
+ 0.1367262452840805,
464
+ 0.2696494162082672,
465
+ 0.00759609742090106,
466
+ 0.08290299028158188,
467
+ 0.8084851503372192,
468
+ 0.8061898946762085,
469
+ 0.05788164213299751,
470
+ 0.72284334897995,
471
+ 0.2660578489303589,
472
+ 0.5892078876495361,
473
+ 0.5761005282402039,
474
+ 0.811720609664917,
475
+ 0.12750492990016937,
476
+ 0.0029384188819676638,
477
+ 0.46889087557792664,
478
+ 0.4423946440219879
479
+ ],
480
+ "q01": [
481
+ -0.13136589226552758,
482
+ -0.024246096761391306,
483
+ -0.05189719808149282,
484
+ -1e-10,
485
+ -0.0026160656495360507,
486
+ -0.23931071566014658,
487
+ -0.10000512030107096,
488
+ -0.12678847973570523,
489
+ -0.10167653675904859,
490
+ -1e-10,
491
+ -0.04783909519164232,
492
+ -0.15011731969174025,
493
+ -0.03566053932528732,
494
+ -0.14333088852474807,
495
+ -0.015240181828669704,
496
+ -1e-10,
497
+ -1e-10,
498
+ -0.33686717257179044,
499
+ -1e-10,
500
+ -0.1350961140365755,
501
+ -0.11334149588543854,
502
+ -1e-10,
503
+ -0.7022344961392326,
504
+ -0.2776577313950343,
505
+ -0.5394677002780325,
506
+ -0.4033692333062952,
507
+ -0.09314109218487832,
508
+ -0.14045375546192315,
509
+ -0.9111085386817417,
510
+ -1e-10,
511
+ -0.24797243346869297,
512
+ 0.0,
513
+ 0.0,
514
+ 0.0,
515
+ 0.0,
516
+ 0.0,
517
+ 0.0,
518
+ 0.0,
519
+ 0.0,
520
+ 0.0,
521
+ 0.0,
522
+ 0.0,
523
+ 0.0,
524
+ 0.0,
525
+ 0.0,
526
+ 0.0,
527
+ 0.0,
528
+ 0.0,
529
+ 0.0,
530
+ 0.0,
531
+ 0.0,
532
+ 0.0,
533
+ 0.0,
534
+ -1e-10,
535
+ -1e-10,
536
+ -1e-10,
537
+ -1e-10,
538
+ -1e-10,
539
+ -1e-10,
540
+ -0.2870999873684479,
541
+ -1e-10,
542
+ -0.01590000102983246,
543
+ -0.17630000422464143,
544
+ -1e-10,
545
+ -1e-10,
546
+ -1e-10,
547
+ -1e-10,
548
+ -1e-10,
549
+ -1e-10,
550
+ -1e-10,
551
+ -1e-10,
552
+ -1e-10,
553
+ -0.006200000278068876,
554
+ -1e-10,
555
+ -1e-10
556
+ ],
557
+ "q99": [
558
+ 0.02158951045498614,
559
+ 0.16638232857872537,
560
+ 0.12718728261393447,
561
+ 0.24990865929830552,
562
+ 0.080158638796064,
563
+ 0.03880121835006106,
564
+ 0.051546728917431256,
565
+ 0.03648561011608348,
566
+ 0.14874143446309396,
567
+ 0.3012807006313445,
568
+ 8.760044392333577e-6,
569
+ 0.1050926848649123,
570
+ 0.04378821893175901,
571
+ 0.00007304405802485703,
572
+ 0.12353252486804131,
573
+ 0.2499613958553109,
574
+ 0.1548596104397836,
575
+ 0.18439460120944595,
576
+ 1.1400343631722312,
577
+ 0.15592786586500085,
578
+ 0.304366214206148,
579
+ 0.2555923827020851,
580
+ 0.5288230844682025,
581
+ 0.19660598214024083,
582
+ 0.429336508913561,
583
+ 1.0832967488560827,
584
+ 1.67866495521356,
585
+ 0.43620634968740046,
586
+ 0.624262357965645,
587
+ 0.6080994198696249,
588
+ 0.11173681515406503,
589
+ 0.0,
590
+ 0.0,
591
+ 0.0,
592
+ 0.0,
593
+ 0.0,
594
+ 0.0,
595
+ 0.0,
596
+ 0.0,
597
+ 0.0,
598
+ 0.0,
599
+ 0.0,
600
+ 0.0,
601
+ 0.0,
602
+ 0.0,
603
+ 0.0,
604
+ 0.0,
605
+ 0.0,
606
+ 0.0,
607
+ 0.0,
608
+ 0.0,
609
+ 0.0,
610
+ 0.0,
611
+ 0.36262744560700955,
612
+ 0.7772445368812712,
613
+ 0.04019195886639062,
614
+ 0.6608678134487301,
615
+ 0.7698459810302883,
616
+ 0.07268546139908137,
617
+ -0.00005741989749369347,
618
+ 0.573185344104591,
619
+ -3.179900225965676e-6,
620
+ -0.000035259900864920146,
621
+ 1.7449509946392208,
622
+ 1.7449509946392208,
623
+ 0.12297540349674525,
624
+ 1.5553888046072155,
625
+ 0.5408918086813123,
626
+ 1.2590481336162715,
627
+ 1.2128573366211086,
628
+ 1.7449509946392208,
629
+ 0.27464504608375134,
630
+ -1.2399000756139686e-6,
631
+ 1.0020995641277461,
632
+ 0.8994201090858609
633
+ ]
634
+ }
635
+ }
636
+ }
a18-pour-water/epoch-010/metadata.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:9a933f678931269f9db579981fe4fcda37c3359f64cc0ea22986a6f43e867029
3
+ size 3379
a18-pour-water/epoch-010/model.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:33c7d4c57d689a906fc51f9169605557d645960d975b6a16000171e44e8f6f55
3
+ size 7473452088
a18-pour-water/epoch-010/protocol.json ADDED
@@ -0,0 +1,56 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "model": "pi0.5 (openpi, PyTorch weights)",
3
+ "task": "A18 pour water (H2)",
4
+ "task_dir": "a18-pour-water",
5
+ "openpi_config": "pi05_hb_h2_a18",
6
+ "layout": "h2",
7
+ "asset_id": "h2_a18",
8
+ "epoch_completed": 10,
9
+ "global_step": 237,
10
+ "total_epochs": 20.0,
11
+ "total_steps": 474,
12
+ "global_batch": 3072,
13
+ "per_gpu_batch": 96,
14
+ "world_size": 32,
15
+ "peak_lr": 5e-05,
16
+ "warmup_steps": 24,
17
+ "max_token_len": 384,
18
+ "init_weights": "lerobot/pi05_base converted to PyTorch, action/state projections resized to 76-D ($P/policies-20261109/models/openpi/pi05_base_g1a76_pytorch)",
19
+ "augmentation": "Humanoid Benchmark online augmentation spec v2 (2026-11-09), on from epoch 1; model-native augmentation off",
20
+ "aug_version": "hb-aug-v2-20261109",
21
+ "model_action_dim": 76,
22
+ "action_horizon": 16,
23
+ "episodes": 100,
24
+ "frames": 72752,
25
+ "fps": 30.0,
26
+ "control_hz": 30.0,
27
+ "prompt": "Grasp the pitcher's handle with the right hand, lift it, pour water into the glass tumbler, then put the pitcher back where it stood and let go.",
28
+ "raw_state_dim": 75,
29
+ "state_layout": "observation.state75 (legs12|waist3|arms14|head2|hands44) | projected_gravity3",
30
+ "action_layout": "action.body_target31 | action.hands44 (+1 zero pad dim -> 76), rows t..t+15",
31
+ "cameras": {
32
+ "base_0_rgb": "observation.images.head",
33
+ "left_wrist_0_rgb": "observation.images.chest",
34
+ "right_wrist_0_rgb": "zeros, image_mask False"
35
+ },
36
+ "raw_image_hw": [
37
+ 480,
38
+ 640
39
+ ],
40
+ "content_box": [
41
+ 28,
42
+ 196,
43
+ 0,
44
+ 224
45
+ ],
46
+ "training_data": "$P/a18-h2-dataset-audit-8e956a10/dataset",
47
+ "normalization": "assets/h2_a18/norm_stats.json (openpi quantile normalization)",
48
+ "inference": "Register configs by importing code/hb_pi05/hb_g1_config.py (needs hb_g1_data.py on the path), then openpi policy_config.create_trained_policy(get_config(openpi_config), <this folder>). Clients send observation/<camera> raw RGB, observation/state and prompt; see HBInputs.",
49
+ "resume_state_included": false,
50
+ "files": {
51
+ "assets/h2_a18/norm_stats.json": 16565,
52
+ "metadata.pt": 3379,
53
+ "model.safetensors": 7473452088,
54
+ "training_plan.json": 746
55
+ }
56
+ }
a18-pour-water/epoch-010/training_plan.json ADDED
@@ -0,0 +1,23 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "task": "a18",
3
+ "layout": "h2",
4
+ "config": "pi05_hb_h2_a18",
5
+ "exp_name": "run",
6
+ "frames": 72752,
7
+ "world_size": 32,
8
+ "per_gpu_batch": 96,
9
+ "global_batch": 3072,
10
+ "epochs": 20.0,
11
+ "num_train_steps": 474,
12
+ "save_steps": [
13
+ 237,
14
+ 474
15
+ ],
16
+ "warmup_steps": 24,
17
+ "peak_lr": 5e-05,
18
+ "weights": "/mnt/nas/cross/Team/RobotAgent/Personal/baijiaqi/humanoid-benchmark-work/preflight-20261104T1648Z/policies-20261109/models/openpi/pi05_base_g1a76_pytorch",
19
+ "checkpoint_dir": "/mnt/nas/cross/Team/RobotAgent/Personal/baijiaqi/humanoid-benchmark-work/preflight-20261104T1648Z/policies-20261109/pi05-runs/chain-hb1110/a18/run/ckpt/pi05_hb_h2_a18/run",
20
+ "augment": true,
21
+ "aug_version": "hb-aug-v2-20261109",
22
+ "max_token_len": 384
23
+ }
a18-pour-water/epoch-020/assets/h2_a18/norm_stats.json ADDED
@@ -0,0 +1,636 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "norm_stats": {
3
+ "state": {
4
+ "mean": [
5
+ -0.02828288823366165,
6
+ 0.03546403720974922,
7
+ 0.03124321810901165,
8
+ 0.16529245674610138,
9
+ 0.033904630690813065,
10
+ -0.1248306930065155,
11
+ -0.020945074036717415,
12
+ -0.03086675889790058,
13
+ 0.013680738396942616,
14
+ 0.20139853656291962,
15
+ -0.0015518380096182227,
16
+ -0.1498299092054367,
17
+ 0.00765480799600482,
18
+ -0.01444979291409254,
19
+ 0.030299680307507515,
20
+ 0.1457539200782776,
21
+ 0.08196060359477997,
22
+ -0.04197724163532257,
23
+ 1.0582942962646484,
24
+ 0.004078589379787445,
25
+ 0.10490317642688751,
26
+ 0.11859612911939621,
27
+ -0.19652463495731354,
28
+ -0.048440903425216675,
29
+ -0.06604290008544922,
30
+ 0.31633737683296204,
31
+ 0.8490549325942993,
32
+ 0.16632796823978424,
33
+ -0.20058786869049072,
34
+ 0.5188571214675903,
35
+ -0.0062464503571391106,
36
+ 0.0012028953060507774,
37
+ 0.0012025771429762244,
38
+ 0.000013984431461722124,
39
+ 0.0012024634052067995,
40
+ 0.008900434710085392,
41
+ -0.0008743949001654983,
42
+ -0.0008745864615775645,
43
+ 0.001202479936182499,
44
+ -0.0008735220180824399,
45
+ -0.09596802294254303,
46
+ 0.00023934645287226886,
47
+ 0.0002393399045104161,
48
+ -0.0008724569343030453,
49
+ 0.0002393364702584222,
50
+ -0.010859426110982895,
51
+ 0.00004498684938880615,
52
+ 0.00004498622729443014,
53
+ 0.00023933025659061968,
54
+ 0.00004498567068367265,
55
+ 0.0009174197912216187,
56
+ 0.000044984764826949686,
57
+ 0.000024329165171366185,
58
+ 0.1904839128255844,
59
+ 0.37486860156059265,
60
+ 0.020827073603868484,
61
+ 0.3586229681968689,
62
+ 0.3983735144138336,
63
+ -0.025473948568105698,
64
+ -0.07296764850616455,
65
+ 0.2999779284000397,
66
+ -0.004135874565690756,
67
+ -0.17300280928611755,
68
+ 0.9253543019294739,
69
+ 0.8886311054229736,
70
+ 0.08327575027942657,
71
+ 0.8554991483688354,
72
+ 0.25384166836738586,
73
+ 0.677445650100708,
74
+ 0.644685685634613,
75
+ 0.9370030164718628,
76
+ 0.15102075040340424,
77
+ -0.003838402684777975,
78
+ 0.5436624884605408,
79
+ 0.4556223452091217,
80
+ -0.023036951199173927,
81
+ -0.010215787217020988,
82
+ -0.9998420476913452
83
+ ],
84
+ "std": [
85
+ 0.03512008860707283,
86
+ 0.01740800216794014,
87
+ 0.03788851574063301,
88
+ 0.03280680626630783,
89
+ 0.022302838042378426,
90
+ 0.01713656447827816,
91
+ 0.03212040290236473,
92
+ 0.016956482082605362,
93
+ 0.05338956415653229,
94
+ 0.027900811284780502,
95
+ 0.012169546447694302,
96
+ 0.015364138409495354,
97
+ 0.01877790130674839,
98
+ 0.012579463422298431,
99
+ 0.021866735070943832,
100
+ 0.03327784687280655,
101
+ 0.017236528918147087,
102
+ 0.11609876155853271,
103
+ 0.05033324658870697,
104
+ 0.05683313310146332,
105
+ 0.10001008212566376,
106
+ 0.06167083978652954,
107
+ 0.35286715626716614,
108
+ 0.09832895547151566,
109
+ 0.23262782394886017,
110
+ 0.4843495786190033,
111
+ 0.6070270538330078,
112
+ 0.17739585041999817,
113
+ 0.32976317405700684,
114
+ 0.2399081438779831,
115
+ 0.10201188921928406,
116
+ 0.0003750584146473557,
117
+ 0.00037507538218051195,
118
+ 0.000020995448721805587,
119
+ 0.0003750614414457232,
120
+ 0.024341240525245667,
121
+ 0.00038038159254938364,
122
+ 0.0003799603146035224,
123
+ 0.00037506569060496986,
124
+ 0.00038039451465010643,
125
+ 0.01200117077678442,
126
+ 0.00007667305180802941,
127
+ 0.00007667495083296672,
128
+ 0.00038073238101787865,
129
+ 0.00007667316822335124,
130
+ 0.013414953835308552,
131
+ 0.000014869372535031289,
132
+ 0.000014868522157485131,
133
+ 0.0000766750454204157,
134
+ 0.000014868902326270472,
135
+ 0.00865352526307106,
136
+ 0.000014869230653857812,
137
+ 0.00003821943755610846,
138
+ 0.1619565486907959,
139
+ 0.3236274719238281,
140
+ 0.014654714614152908,
141
+ 0.31027916073799133,
142
+ 0.3515927493572235,
143
+ 0.022043032571673393,
144
+ 0.05582857504487038,
145
+ 0.25976499915122986,
146
+ 0.004044965375214815,
147
+ 0.07659266144037247,
148
+ 0.7724840641021729,
149
+ 0.7353335618972778,
150
+ 0.06149245798587799,
151
+ 0.7260345220565796,
152
+ 0.2404094636440277,
153
+ 0.5688204765319824,
154
+ 0.5405519604682922,
155
+ 0.7903794050216675,
156
+ 0.1275968998670578,
157
+ 0.00503132538869977,
158
+ 0.45670056343078613,
159
+ 0.43200525641441345,
160
+ 0.01495918445289135,
161
+ 0.009642813354730606,
162
+ 0.0
163
+ ],
164
+ "q01": [
165
+ -0.09696375873552374,
166
+ 0.0007942055537837334,
167
+ -0.051792578461062305,
168
+ 0.13033591651223178,
169
+ -0.01713594465283505,
170
+ -0.15955360559398968,
171
+ -0.0964198788970706,
172
+ -0.06363997625562103,
173
+ -0.1000344696582508,
174
+ 0.1482597910888919,
175
+ -0.027435287981779383,
176
+ -0.18897574493706598,
177
+ -0.044199422308726685,
178
+ -0.05024038314818473,
179
+ -0.01132426475313017,
180
+ 0.08324008762353496,
181
+ 0.038280332696323874,
182
+ -0.3242681729192237,
183
+ 0.9701593956163327,
184
+ -0.13880584287678877,
185
+ -0.09421699412132398,
186
+ 0.020241137548000453,
187
+ -0.7021140373221906,
188
+ -0.27443564700923767,
189
+ -0.5346578447649877,
190
+ -0.4259114304386858,
191
+ -0.08661681087593327,
192
+ -0.12122693202118925,
193
+ -0.9081178965514068,
194
+ 0.005415167589460248,
195
+ -0.2458772536665035,
196
+ 0.00006636888889494642,
197
+ 0.0000663710596573749,
198
+ -1.0299070379400978e-6,
199
+ 0.00006638123974712031,
200
+ -0.051145545630711886,
201
+ -0.0013889520672031631,
202
+ -0.0013888838999733162,
203
+ 0.00006639188308495093,
204
+ -0.0013883197191158069,
205
+ -0.1278356132528748,
206
+ 0.000012710355428494887,
207
+ 0.000012642206731161422,
208
+ -0.0013877263354788194,
209
+ 0.000012643952976322337,
210
+ -0.053515475887209486,
211
+ 2.275902296722304e-6,
212
+ 2.276176049376893e-6,
213
+ 0.00001264573177261127,
214
+ 2.2764722897171513e-6,
215
+ -0.021655506657829896,
216
+ 2.276778616162694e-6,
217
+ -0.0000463191339635946,
218
+ -0.0008273905561459105,
219
+ -0.008519364052306896,
220
+ 0.000015183388484773017,
221
+ -0.007527138107404751,
222
+ 0.009861442021763134,
223
+ -0.06757076855718339,
224
+ -0.13470218816765142,
225
+ -0.005825813357401111,
226
+ -0.010244786777927912,
227
+ -0.256434232508372,
228
+ -0.00025278038957425066,
229
+ -0.000052326070924881094,
230
+ -0.0005493609337659858,
231
+ -0.0002497415102234446,
232
+ -0.0034235131458670207,
233
+ -0.00019695640924990575,
234
+ -0.00011751209532239902,
235
+ -0.0002919652158541938,
236
+ 1.0082914067513594e-8,
237
+ -0.012793561069488308,
238
+ -0.00018825285256620066,
239
+ -0.000057134988586219314,
240
+ -0.05098403841318886,
241
+ -0.03447218551372179,
242
+ -0.9999983127043873
243
+ ],
244
+ "q99": [
245
+ 0.022343595302178443,
246
+ 0.08494462532764574,
247
+ 0.12511245712753394,
248
+ 0.2416719058894359,
249
+ 0.0677620012453274,
250
+ -0.08898747787709982,
251
+ 0.02318780476657832,
252
+ -0.0015199510267689398,
253
+ 0.14604715039697608,
254
+ 0.2637284779196589,
255
+ 0.02539288635049395,
256
+ -0.11248975168235031,
257
+ 0.045069882762597095,
258
+ 0.010824615722864028,
259
+ 0.06618172762760902,
260
+ 0.212699701484278,
261
+ 0.1207836883960253,
262
+ 0.17330300987310265,
263
+ 1.1781945639273799,
264
+ 0.15023529210570086,
265
+ 0.2845899072977251,
266
+ 0.24516796460341958,
267
+ 0.5740916040221649,
268
+ 0.2010916694726258,
269
+ 0.4289175018925724,
270
+ 1.0899582403284118,
271
+ 1.6852502702105407,
272
+ 0.43534713681491477,
273
+ 0.6240539385241481,
274
+ 0.6550094590556405,
275
+ 0.11099284940619253,
276
+ 0.0016734303027044354,
277
+ 0.0016734576745192527,
278
+ 0.00007065863147983712,
279
+ 0.0016737097039632794,
280
+ 0.04393054750393248,
281
+ -0.00003521824024034785,
282
+ -0.000036440479353125766,
283
+ 0.001673639154479647,
284
+ -0.00003455918041791865,
285
+ -0.07552439234949913,
286
+ 0.00033051224737817186,
287
+ 0.0003304989688443708,
288
+ -0.00003296538223554121,
289
+ 0.00033056630809084984,
290
+ 0.002957078129294119,
291
+ 0.00006171814166910019,
292
+ 0.00006171964270592179,
293
+ 0.00033056845467495713,
294
+ 0.00006172230211941015,
295
+ 0.02593200647046453,
296
+ 0.00006172569891158286,
297
+ 0.00011897972866163018,
298
+ 0.3787801509389085,
299
+ 0.7209945422769666,
300
+ 0.0407652068077642,
301
+ 0.6618924614843803,
302
+ 0.7517948510379678,
303
+ 0.005806987208462441,
304
+ 0.002885183540840164,
305
+ 0.570767374535971,
306
+ 0.0014458511145520186,
307
+ -0.06102761673970994,
308
+ 1.7102710429866648,
309
+ 1.6174167523095186,
310
+ 0.1397007980669106,
311
+ 1.554299900071352,
312
+ 0.5039048781097059,
313
+ 1.239142416988845,
314
+ 1.1760776755933253,
315
+ 1.7112652279961076,
316
+ 0.2737724787321517,
317
+ 0.008886616448485709,
318
+ 0.988431640873346,
319
+ 0.8869660620081917,
320
+ 0.0035224365899956395,
321
+ 0.016511748692902894,
322
+ -0.9981961741900639
323
+ ]
324
+ },
325
+ "actions": {
326
+ "mean": [
327
+ -0.039711859077215195,
328
+ 0.042044274508953094,
329
+ 0.031410254538059235,
330
+ 0.1572873443365097,
331
+ 0.05045900121331215,
332
+ -0.1497725248336792,
333
+ -0.006279041990637779,
334
+ -0.02705513872206211,
335
+ 0.014146559871733189,
336
+ 0.21941550076007843,
337
+ -0.025561144575476646,
338
+ -0.030189640820026398,
339
+ 0.009197989478707314,
340
+ -0.049546219408512115,
341
+ 0.03526940196752548,
342
+ 0.14168091118335724,
343
+ 0.11557798087596893,
344
+ -0.04095250740647316,
345
+ 0.9854463934898376,
346
+ 0.008960100822150707,
347
+ 0.08984862267971039,
348
+ 0.12244174629449844,
349
+ -0.19101494550704956,
350
+ -0.050739310681819916,
351
+ -0.06203479319810867,
352
+ 0.30884090065956116,
353
+ 0.8466822504997253,
354
+ 0.16446106135845184,
355
+ -0.19862252473831177,
356
+ 0.4859173893928528,
357
+ -0.0057493275962769985,
358
+ 0.0,
359
+ 0.0,
360
+ 0.0,
361
+ 0.0,
362
+ 0.0,
363
+ 0.0,
364
+ 0.0,
365
+ 0.0,
366
+ 0.0,
367
+ 0.0,
368
+ 0.0,
369
+ 0.0,
370
+ 0.0,
371
+ 0.0,
372
+ 0.0,
373
+ 0.0,
374
+ 0.0,
375
+ 0.0,
376
+ 0.0,
377
+ 0.0,
378
+ 0.0,
379
+ 0.0,
380
+ 0.19399771094322205,
381
+ 0.4265739321708679,
382
+ 0.023019764572381973,
383
+ 0.3584686815738678,
384
+ 0.40135377645492554,
385
+ 0.045260727405548096,
386
+ -0.176839679479599,
387
+ 0.31074193120002747,
388
+ -0.009844515472650528,
389
+ -0.11006658524274826,
390
+ 0.9708828926086426,
391
+ 0.974152684211731,
392
+ 0.07610955834388733,
393
+ 0.8584804534912109,
394
+ 0.2781544029712677,
395
+ 0.6949679851531982,
396
+ 0.6698334813117981,
397
+ 0.9649102091789246,
398
+ 0.15337009727954865,
399
+ -0.0038274202961474657,
400
+ 0.5546799898147583,
401
+ 0.4666762948036194
402
+ ],
403
+ "std": [
404
+ 0.039720017462968826,
405
+ 0.03597322106361389,
406
+ 0.038342930376529694,
407
+ 0.04752495139837265,
408
+ 0.02657449059188366,
409
+ 0.05687854439020157,
410
+ 0.04023086652159691,
411
+ 0.03455903008580208,
412
+ 0.05493085831403732,
413
+ 0.042149096727371216,
414
+ 0.012635765597224236,
415
+ 0.07036220282316208,
416
+ 0.015654711052775383,
417
+ 0.03608144074678421,
418
+ 0.03046419285237789,
419
+ 0.05074617266654968,
420
+ 0.020575977861881256,
421
+ 0.12018261849880219,
422
+ 0.1278754025697708,
423
+ 0.0572555735707283,
424
+ 0.10964718461036682,
425
+ 0.0703325942158699,
426
+ 0.3559415340423584,
427
+ 0.09842006117105484,
428
+ 0.23309923708438873,
429
+ 0.47085040807724,
430
+ 0.6077632308006287,
431
+ 0.18328437209129333,
432
+ 0.33357715606689453,
433
+ 0.22189438343048096,
434
+ 0.10337754338979721,
435
+ 0.0,
436
+ 0.0,
437
+ 0.0,
438
+ 0.0,
439
+ 0.0,
440
+ 0.0,
441
+ 0.0,
442
+ 0.0,
443
+ 0.0,
444
+ 0.0,
445
+ 0.0,
446
+ 0.0,
447
+ 0.0,
448
+ 0.0,
449
+ 0.0,
450
+ 0.0,
451
+ 0.0,
452
+ 0.0,
453
+ 0.0,
454
+ 0.0,
455
+ 0.0,
456
+ 0.0,
457
+ 0.1725616753101349,
458
+ 0.3615170419216156,
459
+ 0.018322760239243507,
460
+ 0.31214597821235657,
461
+ 0.3793746829032898,
462
+ 0.03442889451980591,
463
+ 0.1367262452840805,
464
+ 0.2696494162082672,
465
+ 0.00759609742090106,
466
+ 0.08290299028158188,
467
+ 0.8084851503372192,
468
+ 0.8061898946762085,
469
+ 0.05788164213299751,
470
+ 0.72284334897995,
471
+ 0.2660578489303589,
472
+ 0.5892078876495361,
473
+ 0.5761005282402039,
474
+ 0.811720609664917,
475
+ 0.12750492990016937,
476
+ 0.0029384188819676638,
477
+ 0.46889087557792664,
478
+ 0.4423946440219879
479
+ ],
480
+ "q01": [
481
+ -0.13136589226552758,
482
+ -0.024246096761391306,
483
+ -0.05189719808149282,
484
+ -1e-10,
485
+ -0.0026160656495360507,
486
+ -0.23931071566014658,
487
+ -0.10000512030107096,
488
+ -0.12678847973570523,
489
+ -0.10167653675904859,
490
+ -1e-10,
491
+ -0.04783909519164232,
492
+ -0.15011731969174025,
493
+ -0.03566053932528732,
494
+ -0.14333088852474807,
495
+ -0.015240181828669704,
496
+ -1e-10,
497
+ -1e-10,
498
+ -0.33686717257179044,
499
+ -1e-10,
500
+ -0.1350961140365755,
501
+ -0.11334149588543854,
502
+ -1e-10,
503
+ -0.7022344961392326,
504
+ -0.2776577313950343,
505
+ -0.5394677002780325,
506
+ -0.4033692333062952,
507
+ -0.09314109218487832,
508
+ -0.14045375546192315,
509
+ -0.9111085386817417,
510
+ -1e-10,
511
+ -0.24797243346869297,
512
+ 0.0,
513
+ 0.0,
514
+ 0.0,
515
+ 0.0,
516
+ 0.0,
517
+ 0.0,
518
+ 0.0,
519
+ 0.0,
520
+ 0.0,
521
+ 0.0,
522
+ 0.0,
523
+ 0.0,
524
+ 0.0,
525
+ 0.0,
526
+ 0.0,
527
+ 0.0,
528
+ 0.0,
529
+ 0.0,
530
+ 0.0,
531
+ 0.0,
532
+ 0.0,
533
+ 0.0,
534
+ -1e-10,
535
+ -1e-10,
536
+ -1e-10,
537
+ -1e-10,
538
+ -1e-10,
539
+ -1e-10,
540
+ -0.2870999873684479,
541
+ -1e-10,
542
+ -0.01590000102983246,
543
+ -0.17630000422464143,
544
+ -1e-10,
545
+ -1e-10,
546
+ -1e-10,
547
+ -1e-10,
548
+ -1e-10,
549
+ -1e-10,
550
+ -1e-10,
551
+ -1e-10,
552
+ -1e-10,
553
+ -0.006200000278068876,
554
+ -1e-10,
555
+ -1e-10
556
+ ],
557
+ "q99": [
558
+ 0.02158951045498614,
559
+ 0.16638232857872537,
560
+ 0.12718728261393447,
561
+ 0.24990865929830552,
562
+ 0.080158638796064,
563
+ 0.03880121835006106,
564
+ 0.051546728917431256,
565
+ 0.03648561011608348,
566
+ 0.14874143446309396,
567
+ 0.3012807006313445,
568
+ 8.760044392333577e-6,
569
+ 0.1050926848649123,
570
+ 0.04378821893175901,
571
+ 0.00007304405802485703,
572
+ 0.12353252486804131,
573
+ 0.2499613958553109,
574
+ 0.1548596104397836,
575
+ 0.18439460120944595,
576
+ 1.1400343631722312,
577
+ 0.15592786586500085,
578
+ 0.304366214206148,
579
+ 0.2555923827020851,
580
+ 0.5288230844682025,
581
+ 0.19660598214024083,
582
+ 0.429336508913561,
583
+ 1.0832967488560827,
584
+ 1.67866495521356,
585
+ 0.43620634968740046,
586
+ 0.624262357965645,
587
+ 0.6080994198696249,
588
+ 0.11173681515406503,
589
+ 0.0,
590
+ 0.0,
591
+ 0.0,
592
+ 0.0,
593
+ 0.0,
594
+ 0.0,
595
+ 0.0,
596
+ 0.0,
597
+ 0.0,
598
+ 0.0,
599
+ 0.0,
600
+ 0.0,
601
+ 0.0,
602
+ 0.0,
603
+ 0.0,
604
+ 0.0,
605
+ 0.0,
606
+ 0.0,
607
+ 0.0,
608
+ 0.0,
609
+ 0.0,
610
+ 0.0,
611
+ 0.36262744560700955,
612
+ 0.7772445368812712,
613
+ 0.04019195886639062,
614
+ 0.6608678134487301,
615
+ 0.7698459810302883,
616
+ 0.07268546139908137,
617
+ -0.00005741989749369347,
618
+ 0.573185344104591,
619
+ -3.179900225965676e-6,
620
+ -0.000035259900864920146,
621
+ 1.7449509946392208,
622
+ 1.7449509946392208,
623
+ 0.12297540349674525,
624
+ 1.5553888046072155,
625
+ 0.5408918086813123,
626
+ 1.2590481336162715,
627
+ 1.2128573366211086,
628
+ 1.7449509946392208,
629
+ 0.27464504608375134,
630
+ -1.2399000756139686e-6,
631
+ 1.0020995641277461,
632
+ 0.8994201090858609
633
+ ]
634
+ }
635
+ }
636
+ }
a18-pour-water/epoch-020/metadata.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:d003aaaa8863efdb79f8c5c09b37b4e5bb19a309f4598f6dbf3d51aeb4199ee6
3
+ size 3379
a18-pour-water/epoch-020/model.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:adc0b3883d08cfa7c7d7ccacda14eadb2f99eb4f323bd68ef8cd4182c55bde9c
3
+ size 7473452088
a18-pour-water/epoch-020/protocol.json ADDED
@@ -0,0 +1,56 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "model": "pi0.5 (openpi, PyTorch weights)",
3
+ "task": "A18 pour water (H2)",
4
+ "task_dir": "a18-pour-water",
5
+ "openpi_config": "pi05_hb_h2_a18",
6
+ "layout": "h2",
7
+ "asset_id": "h2_a18",
8
+ "epoch_completed": 20,
9
+ "global_step": 474,
10
+ "total_epochs": 20.0,
11
+ "total_steps": 474,
12
+ "global_batch": 3072,
13
+ "per_gpu_batch": 96,
14
+ "world_size": 32,
15
+ "peak_lr": 5e-05,
16
+ "warmup_steps": 24,
17
+ "max_token_len": 384,
18
+ "init_weights": "lerobot/pi05_base converted to PyTorch, action/state projections resized to 76-D ($P/policies-20261109/models/openpi/pi05_base_g1a76_pytorch)",
19
+ "augmentation": "Humanoid Benchmark online augmentation spec v2 (2026-11-09), on from epoch 1; model-native augmentation off",
20
+ "aug_version": "hb-aug-v2-20261109",
21
+ "model_action_dim": 76,
22
+ "action_horizon": 16,
23
+ "episodes": 100,
24
+ "frames": 72752,
25
+ "fps": 30.0,
26
+ "control_hz": 30.0,
27
+ "prompt": "Grasp the pitcher's handle with the right hand, lift it, pour water into the glass tumbler, then put the pitcher back where it stood and let go.",
28
+ "raw_state_dim": 75,
29
+ "state_layout": "observation.state75 (legs12|waist3|arms14|head2|hands44) | projected_gravity3",
30
+ "action_layout": "action.body_target31 | action.hands44 (+1 zero pad dim -> 76), rows t..t+15",
31
+ "cameras": {
32
+ "base_0_rgb": "observation.images.head",
33
+ "left_wrist_0_rgb": "observation.images.chest",
34
+ "right_wrist_0_rgb": "zeros, image_mask False"
35
+ },
36
+ "raw_image_hw": [
37
+ 480,
38
+ 640
39
+ ],
40
+ "content_box": [
41
+ 28,
42
+ 196,
43
+ 0,
44
+ 224
45
+ ],
46
+ "training_data": "$P/a18-h2-dataset-audit-8e956a10/dataset",
47
+ "normalization": "assets/h2_a18/norm_stats.json (openpi quantile normalization)",
48
+ "inference": "Register configs by importing code/hb_pi05/hb_g1_config.py (needs hb_g1_data.py on the path), then openpi policy_config.create_trained_policy(get_config(openpi_config), <this folder>). Clients send observation/<camera> raw RGB, observation/state and prompt; see HBInputs.",
49
+ "resume_state_included": false,
50
+ "files": {
51
+ "assets/h2_a18/norm_stats.json": 16565,
52
+ "metadata.pt": 3379,
53
+ "model.safetensors": 7473452088,
54
+ "training_plan.json": 746
55
+ }
56
+ }
a18-pour-water/epoch-020/training_plan.json ADDED
@@ -0,0 +1,23 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "task": "a18",
3
+ "layout": "h2",
4
+ "config": "pi05_hb_h2_a18",
5
+ "exp_name": "run",
6
+ "frames": 72752,
7
+ "world_size": 32,
8
+ "per_gpu_batch": 96,
9
+ "global_batch": 3072,
10
+ "epochs": 20.0,
11
+ "num_train_steps": 474,
12
+ "save_steps": [
13
+ 237,
14
+ 474
15
+ ],
16
+ "warmup_steps": 24,
17
+ "peak_lr": 5e-05,
18
+ "weights": "/mnt/nas/cross/Team/RobotAgent/Personal/baijiaqi/humanoid-benchmark-work/preflight-20261104T1648Z/policies-20261109/models/openpi/pi05_base_g1a76_pytorch",
19
+ "checkpoint_dir": "/mnt/nas/cross/Team/RobotAgent/Personal/baijiaqi/humanoid-benchmark-work/preflight-20261104T1648Z/policies-20261109/pi05-runs/chain-hb1110/a18/run/ckpt/pi05_hb_h2_a18/run",
20
+ "augment": true,
21
+ "aug_version": "hb-aug-v2-20261109",
22
+ "max_token_len": 384
23
+ }
a3-fridge-open/epoch-010/assets/g1_a3/norm_stats.json ADDED
@@ -0,0 +1,504 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "norm_stats": {
3
+ "state": {
4
+ "mean": [
5
+ -0.05641397461295128,
6
+ 0.07833150029182434,
7
+ 0.11269253492355347,
8
+ 0.3092529773712158,
9
+ -0.18253354728221893,
10
+ -0.021320076659321785,
11
+ -0.10698527842760086,
12
+ -0.05834974721074104,
13
+ -0.20540674030780792,
14
+ 0.3034929633140564,
15
+ -0.1302507370710373,
16
+ -0.013441003859043121,
17
+ -0.030619896948337555,
18
+ -0.03488051891326904,
19
+ 0.020995844155550003,
20
+ 0.19224053621292114,
21
+ 0.14336878061294556,
22
+ 0.006716981064528227,
23
+ -0.1879279464483261,
24
+ -0.0025916018057614565,
25
+ 0.005836201831698418,
26
+ -0.010614745318889618,
27
+ 0.15154524147510529,
28
+ -0.17589232325553894,
29
+ -0.2622005045413971,
30
+ -0.47192373871803284,
31
+ 0.2191326767206192,
32
+ 0.20637701451778412,
33
+ -0.009809371083974838,
34
+ 0.999921977519989,
35
+ 0.9999970197677612,
36
+ 0.9999962449073792,
37
+ 0.9999969005584717,
38
+ 0.9999812245368958,
39
+ 0.004240452777594328,
40
+ 0.8682456612586975,
41
+ 0.869608461856842,
42
+ 0.873170793056488,
43
+ 0.881648600101471,
44
+ 0.8788400888442993,
45
+ 0.004410883411765099,
46
+ -0.03297373279929161,
47
+ 0.009942770935595036,
48
+ -0.9982253909111023
49
+ ],
50
+ "std": [
51
+ 0.15939587354660034,
52
+ 0.06632916629314423,
53
+ 0.07504713535308838,
54
+ 0.17704583704471588,
55
+ 0.1108374074101448,
56
+ 0.04978697746992111,
57
+ 0.1295444816350937,
58
+ 0.062313150614500046,
59
+ 0.08294235169887543,
60
+ 0.1889764368534088,
61
+ 0.1166033148765564,
62
+ 0.05051586776971817,
63
+ 0.06022709235548973,
64
+ 0.05384501442313194,
65
+ 0.02067396230995655,
66
+ 0.02285316400229931,
67
+ 0.012044434435665607,
68
+ 0.013775781728327274,
69
+ 0.021458350121974945,
70
+ 0.009183689951896667,
71
+ 0.011165750212967396,
72
+ 0.014735972508788109,
73
+ 0.3950332999229431,
74
+ 0.13744504749774933,
75
+ 0.37059733271598816,
76
+ 0.38896965980529785,
77
+ 0.19889448583126068,
78
+ 0.24846522510051727,
79
+ 0.08681652694940567,
80
+ 0.0,
81
+ 0.0,
82
+ 0.0,
83
+ 0.0,
84
+ 0.0,
85
+ 0.030091701075434685,
86
+ 0.21434949338436127,
87
+ 0.2120189517736435,
88
+ 0.2065865397453308,
89
+ 0.1933792531490326,
90
+ 0.1923707276582718,
91
+ 0.030065977945923805,
92
+ 0.03751828148961067,
93
+ 0.031833503395318985,
94
+ 0.0
95
+ ],
96
+ "q01": [
97
+ -0.46506815473136753,
98
+ -0.04889196949116753,
99
+ -0.03771492845147742,
100
+ 0.12403084383573584,
101
+ -0.4327747708072943,
102
+ -0.1480070360368902,
103
+ -0.4722705717922792,
104
+ -0.21336711200995134,
105
+ -0.4257794928347497,
106
+ 0.14030948715447483,
107
+ -0.4798982841342425,
108
+ -0.12194732715222922,
109
+ -0.2540194292733686,
110
+ -0.2156617373707264,
111
+ -0.03905836607585417,
112
+ 0.13080863007519333,
113
+ 0.11221892317597773,
114
+ -0.030211272256092374,
115
+ -0.2578536881135155,
116
+ -0.033230624382752746,
117
+ -0.01635900642697462,
118
+ -0.05323737701858385,
119
+ -1.0001329558172423,
120
+ -0.5889036026173377,
121
+ -1.3324901583751518,
122
+ -1.0482400876077553,
123
+ -0.043778369650742166,
124
+ -0.43015055529414703,
125
+ -0.2122822610443819,
126
+ 0.9979524703889018,
127
+ 0.9998749794303501,
128
+ 0.9998566181770135,
129
+ 0.9998688959205332,
130
+ 0.9992656699177477,
131
+ -1e-10,
132
+ 0.49980709718610594,
133
+ 0.49981258509911924,
134
+ 0.4997684336261139,
135
+ 0.509475866386762,
136
+ 0.5038498614157283,
137
+ -1e-10,
138
+ -0.1147158351111296,
139
+ -0.07648704438251047,
140
+ -0.9999822119043099
141
+ ],
142
+ "q99": [
143
+ 0.20402788861789334,
144
+ 0.2784175859561666,
145
+ 0.3222591326691111,
146
+ 1.0027285105278778,
147
+ 0.14594734426464373,
148
+ 0.10093794696942421,
149
+ 0.13905996096478002,
150
+ 0.06319464641553324,
151
+ -0.03222518681984321,
152
+ 1.0404026509018791,
153
+ 0.11550482268441065,
154
+ 0.116426306980423,
155
+ 0.07085631988559993,
156
+ 0.0778610076478406,
157
+ 0.07582880456800241,
158
+ 0.25341256991924554,
159
+ 0.17726504346664093,
160
+ 0.042624652664472365,
161
+ -0.14470140677618784,
162
+ 0.019137383483876676,
163
+ 0.03877572672478539,
164
+ 0.005319607630962106,
165
+ 1.2571805590330305,
166
+ 0.13922517023041814,
167
+ 0.1538355433300771,
168
+ 0.8272105335782394,
169
+ 0.6667848346517331,
170
+ 0.6415972574619173,
171
+ 0.20692930209156313,
172
+ 0.9999940131710202,
173
+ 0.9999995318577916,
174
+ 0.9999999581263691,
175
+ 0.9999999589369923,
176
+ 0.9999998499916226,
177
+ 0.21753646457975004,
178
+ 0.9998999014602333,
179
+ 0.9998998424277932,
180
+ 0.999899813525501,
181
+ 0.9999008442209394,
182
+ 0.9999001911805303,
183
+ 0.21746011237681775,
184
+ 0.07922326138295144,
185
+ 0.10487203598122224,
186
+ -0.9906932630281675
187
+ ]
188
+ },
189
+ "actions": {
190
+ "mean": [
191
+ 0.15783676505088806,
192
+ 0.16300027072429657,
193
+ -0.04454663768410683,
194
+ -0.019376136362552643,
195
+ 0.03434408828616142,
196
+ 0.04013373702764511,
197
+ -0.0696130320429802,
198
+ 0.16177308559417725,
199
+ 0.09804673492908478,
200
+ -0.21799813210964203,
201
+ -0.1298724114894867,
202
+ -0.14870725572109222,
203
+ -0.009640868753194809,
204
+ 0.09577104449272156,
205
+ 0.17926457524299622,
206
+ -0.0506000779569149,
207
+ 0.16960610449314117,
208
+ 0.13840295374393463,
209
+ -0.0558854341506958,
210
+ -0.15770410001277924,
211
+ -0.1273193508386612,
212
+ -0.07514706999063492,
213
+ -0.04349217191338539,
214
+ 0.0012820761185139418,
215
+ 0.08928761631250381,
216
+ -0.0955679714679718,
217
+ -0.12635554373264313,
218
+ 0.01848260685801506,
219
+ 0.009467734955251217,
220
+ 0.12164337188005447,
221
+ 0.12644965946674347,
222
+ -0.0803895965218544,
223
+ -0.014793431386351585,
224
+ 0.19501863420009613,
225
+ 0.04290221258997917,
226
+ 0.03360012546181679,
227
+ -0.01547778770327568,
228
+ 0.06523936241865158,
229
+ 0.08329988270998001,
230
+ -0.06897857785224915,
231
+ -0.03849208727478981,
232
+ -0.057777289301157,
233
+ 0.019498951733112335,
234
+ -0.1662408858537674,
235
+ 0.06778231263160706,
236
+ -0.13049370050430298,
237
+ -0.02015572413802147,
238
+ -0.014168542809784412,
239
+ -0.16451606154441833,
240
+ -0.060739967972040176,
241
+ 0.09654980152845383,
242
+ 0.16325144469738007,
243
+ -0.02566051296889782,
244
+ -0.21247075498104095,
245
+ -0.009581679478287697,
246
+ 0.05965486913919449,
247
+ -0.11509575694799423,
248
+ 0.19388903677463531,
249
+ -0.003796325996518135,
250
+ -0.16428831219673157,
251
+ -0.08977831900119781,
252
+ 0.12237153202295303,
253
+ 0.20675969123840332,
254
+ -0.04955295845866203,
255
+ 0.9999598860740662,
256
+ 0.9999998211860657,
257
+ 0.9999997019767761,
258
+ 0.9999997019767761,
259
+ 0.9999992847442627,
260
+ 0.0007821584003977478,
261
+ 0.8682477474212646,
262
+ 0.8695738911628723,
263
+ 0.8731648921966553,
264
+ 0.8815546035766602,
265
+ 0.8789401650428772,
266
+ 0.0009576878510415554
267
+ ],
268
+ "std": [
269
+ 0.08065593242645264,
270
+ 0.08959350734949112,
271
+ 0.11977304518222809,
272
+ 0.10043507069349289,
273
+ 0.14305159449577332,
274
+ 0.08850719034671783,
275
+ 0.14513453841209412,
276
+ 0.12789000570774078,
277
+ 0.06797633320093155,
278
+ 0.050453633069992065,
279
+ 0.0903969332575798,
280
+ 0.06595497578382492,
281
+ 0.11297961324453354,
282
+ 0.07852301746606827,
283
+ 0.08324233442544937,
284
+ 0.08588460087776184,
285
+ 0.0624857097864151,
286
+ 0.09392557293176651,
287
+ 0.11839237064123154,
288
+ 0.0951315388083458,
289
+ 0.08514631539583206,
290
+ 0.09915973991155624,
291
+ 0.06709878891706467,
292
+ 0.0878397673368454,
293
+ 0.0614795908331871,
294
+ 0.08896732330322266,
295
+ 0.06850738823413849,
296
+ 0.08969742804765701,
297
+ 0.11463499069213867,
298
+ 0.08277275413274765,
299
+ 0.10993892699480057,
300
+ 0.07832051813602448,
301
+ 0.0719744935631752,
302
+ 0.07010309398174286,
303
+ 0.04459095746278763,
304
+ 0.07442177087068558,
305
+ 0.10430904477834702,
306
+ 0.1396985650062561,
307
+ 0.08694476634263992,
308
+ 0.07248101383447647,
309
+ 0.08530312776565552,
310
+ 0.1339278668165207,
311
+ 0.082322858273983,
312
+ 0.06349117308855057,
313
+ 0.11075259000062943,
314
+ 0.08747898787260056,
315
+ 0.10215355455875397,
316
+ 0.13771311938762665,
317
+ 0.10990352183580399,
318
+ 0.11261215806007385,
319
+ 0.08981925994157791,
320
+ 0.06594730913639069,
321
+ 0.16168172657489777,
322
+ 0.09319587051868439,
323
+ 0.06820440292358398,
324
+ 0.12118405848741531,
325
+ 0.11162112653255463,
326
+ 0.045856211334466934,
327
+ 0.11358722299337387,
328
+ 0.08551610261201859,
329
+ 0.07454151660203934,
330
+ 0.0910487249493599,
331
+ 0.06291131675243378,
332
+ 0.11520238220691681,
333
+ 0.0,
334
+ 0.0,
335
+ 0.0,
336
+ 0.0,
337
+ 0.0,
338
+ 0.01162454579025507,
339
+ 0.21439313888549805,
340
+ 0.21226760745048523,
341
+ 0.20670393109321594,
342
+ 0.1937810778617859,
343
+ 0.1922834813594818,
344
+ 0.011621065437793732
345
+ ],
346
+ "q01": [
347
+ -7.499999232774002e-11,
348
+ -0.06250000008,
349
+ -0.25003750006923997,
350
+ -0.25001250006364,
351
+ -0.25007500005716,
352
+ -0.18750000003999998,
353
+ -0.37507500005716,
354
+ -0.12507500008184003,
355
+ -0.06252500006668,
356
+ -0.31252500006668,
357
+ -0.312500000075,
358
+ -0.3125000001,
359
+ -0.25001250006364,
360
+ -0.06252500007144002,
361
+ -0.06250000003999998,
362
+ -0.25001250005555997,
363
+ -0.000012500055559999668,
364
+ -0.062500000075,
365
+ -0.25001250006364,
366
+ -0.37501250005555997,
367
+ -0.31251250005555997,
368
+ -0.37505000007144,
369
+ -0.25001250005555997,
370
+ -0.18751250005556,
371
+ -0.06252500006668,
372
+ -0.3125625000778,
373
+ -0.25000000005,
374
+ -0.25001250006364,
375
+ -0.25000000005,
376
+ -0.06256250007780001,
377
+ -0.06250000005,
378
+ -0.2500625000778,
379
+ -0.18750000005,
380
+ -0.000025000071439996996,
381
+ -0.06250000002,
382
+ -0.12500000005,
383
+ -0.25010000008336003,
384
+ -0.25005000007144,
385
+ -0.1250625000778,
386
+ -0.2500000001,
387
+ -0.25000000008,
388
+ -0.31253750006923997,
389
+ -0.18751250005556,
390
+ -0.31252500007144,
391
+ -0.18757500008184003,
392
+ -0.3125625000778,
393
+ -0.25000000006,
394
+ -0.31250000006,
395
+ -0.4375000001,
396
+ -0.37505000007144,
397
+ -0.12500000006,
398
+ -1e-10,
399
+ -0.43760000008752,
400
+ -0.37500000006,
401
+ -0.12505000004288003,
402
+ -0.25010000008336003,
403
+ -0.37507500008184,
404
+ 0.12497499993332001,
405
+ -0.25010000003335997,
406
+ -0.31251250005555997,
407
+ -0.187500000025,
408
+ -0.12500000008,
409
+ 0.062474999928559996,
410
+ -0.37510000008464,
411
+ 0.9979524703889018,
412
+ 0.9998749794303501,
413
+ 0.9998566181770135,
414
+ 0.9998688959205332,
415
+ 0.9995942571558815,
416
+ 0.0,
417
+ 0.49980709718610594,
418
+ 0.49981258509911924,
419
+ 0.4997684336261139,
420
+ 0.509475866386762,
421
+ 0.5038498614157283,
422
+ 0.0
423
+ ],
424
+ "q99": [
425
+ 0.37490000007496,
426
+ 0.37487500005996005,
427
+ 0.31237500006920005,
428
+ 0.24993750008179993,
429
+ 0.37485000008568004,
430
+ 0.2498750000999601,
431
+ 0.24985000008568004,
432
+ 0.43743750008179993,
433
+ 0.24992500009996005,
434
+ -0.06254999993335997,
435
+ 0.062400000074959994,
436
+ -0.00003749995716001431,
437
+ 0.24993750008179993,
438
+ 0.24993750007139998,
439
+ 0.3123750000799601,
440
+ 0.12495000007776003,
441
+ 0.31240000005552004,
442
+ 0.31240000007496,
443
+ 0.24993750008179993,
444
+ -0.00004999992223997385,
445
+ 0.062450000077760026,
446
+ 0.12492500004284002,
447
+ 0.12495000007776003,
448
+ 0.18745000007776003,
449
+ 0.18745000006664003,
450
+ 0.12495000007776003,
451
+ 0.062400000074959994,
452
+ 0.18737500006359997,
453
+ 0.24995000008332002,
454
+ 0.31240000005552004,
455
+ 0.37490000006664004,
456
+ 0.12490000005552004,
457
+ 0.12490000007496,
458
+ 0.3124375000714,
459
+ 0.12493750009996002,
460
+ 0.18740000007496,
461
+ 0.31240000006664004,
462
+ 0.43735000008568004,
463
+ 0.31245000007776,
464
+ 0.12493750007139998,
465
+ 0.18737500005996005,
466
+ 0.31243750008460003,
467
+ 0.18745000007776003,
468
+ -0.06253749995716001,
469
+ 0.37493750008179993,
470
+ 0.12495000007776003,
471
+ 0.1873750000799601,
472
+ 0.31243750007332005,
473
+ 0.06237500005996005,
474
+ 0.18740000005712,
475
+ 0.3123750000799601,
476
+ 0.24993750005996,
477
+ 0.37480000007495995,
478
+ 0.06237500007996011,
479
+ 0.12493750007139998,
480
+ 0.31240000006664004,
481
+ 0.18743750008179993,
482
+ 0.31247500003332,
483
+ 0.18745000008332002,
484
+ 0.062450000077760026,
485
+ 0.062400000074959994,
486
+ 0.31237500005996005,
487
+ 0.31246250004284,
488
+ 0.24987500006920005,
489
+ 0.9999940131710202,
490
+ 0.9999995318577916,
491
+ 0.9999999581263691,
492
+ 0.9999999589369923,
493
+ 0.9999998499916226,
494
+ 0.30305999517440796,
495
+ 0.9998999014602333,
496
+ 0.9998998424277932,
497
+ 0.999899813525501,
498
+ 0.9999008442209394,
499
+ 0.9999001911805303,
500
+ 0.3030380606651306
501
+ ]
502
+ }
503
+ }
504
+ }
a3-fridge-open/epoch-010/metadata.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:654f0b149b64b7c5784cbd4b474d653386cc079183251c0de638406555ad5c1c
3
+ size 3379
a3-fridge-open/epoch-010/model.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:80967b5b948cc0f1d2fc9d254b49ba1513478d7ab3d4a40e7733d20c308d0638
3
+ size 7473452088
a3-fridge-open/epoch-010/protocol.json ADDED
@@ -0,0 +1,56 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "model": "pi0.5 (openpi, PyTorch weights)",
3
+ "task": "A3 open fridge door (G1)",
4
+ "task_dir": "a3-fridge-open",
5
+ "openpi_config": "pi05_hb_g1_a3",
6
+ "layout": "g1",
7
+ "asset_id": "g1_a3",
8
+ "epoch_completed": 10,
9
+ "global_step": 92,
10
+ "total_epochs": 20.0,
11
+ "total_steps": 184,
12
+ "global_batch": 3072,
13
+ "per_gpu_batch": 96,
14
+ "world_size": 32,
15
+ "peak_lr": 5e-05,
16
+ "warmup_steps": 9,
17
+ "max_token_len": 256,
18
+ "init_weights": "lerobot/pi05_base converted to PyTorch, action/state projections resized to 76-D ($P/policies-20261109/models/openpi/pi05_base_g1a76_pytorch)",
19
+ "augmentation": "Humanoid Benchmark online augmentation spec v2 (2026-11-09), on from epoch 1; model-native augmentation off",
20
+ "aug_version": "hb-aug-v2-20261109",
21
+ "model_action_dim": 76,
22
+ "action_horizon": 16,
23
+ "episodes": 100,
24
+ "frames": 28180,
25
+ "fps": 30.0,
26
+ "control_hz": 30.0,
27
+ "prompt": "Walk to the refrigerator, pull its upper door open with your right hand, release the handle and pull your hand back.",
28
+ "raw_state_dim": 41,
29
+ "state_layout": "body29(observation.state by name) | inspire_left6 | inspire_right6 | projected_gravity3",
30
+ "action_layout": "action.motion_token64 | inspire_left6 | inspire_right6 (same-row observed hands), rows t..t+15",
31
+ "cameras": {
32
+ "base_0_rgb": "observation.images.ego_view",
33
+ "left_wrist_0_rgb": "observation.images.chest",
34
+ "right_wrist_0_rgb": "zeros, image_mask False"
35
+ },
36
+ "raw_image_hw": [
37
+ 480,
38
+ 640
39
+ ],
40
+ "content_box": [
41
+ 28,
42
+ 196,
43
+ 0,
44
+ 224
45
+ ],
46
+ "training_data": "$P/a3-a7-20261109/a3-fridge-open-expert-v1/780fcf1e4bd7db947a8c2d8f7b92e61f24f71b45",
47
+ "normalization": "assets/g1_a3/norm_stats.json (openpi quantile normalization)",
48
+ "inference": "Register configs by importing code/hb_pi05/hb_g1_config.py (needs hb_g1_data.py on the path), then openpi policy_config.create_trained_policy(get_config(openpi_config), <this folder>). Clients send observation/<camera> raw RGB, observation/state and prompt; see HBInputs.",
49
+ "resume_state_included": false,
50
+ "files": {
51
+ "assets/g1_a3/norm_stats.json": 13795,
52
+ "metadata.pt": 3379,
53
+ "model.safetensors": 7473452088,
54
+ "training_plan.json": 740
55
+ }
56
+ }
a3-fridge-open/epoch-010/training_plan.json ADDED
@@ -0,0 +1,23 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "task": "a3",
3
+ "layout": "g1",
4
+ "config": "pi05_hb_g1_a3",
5
+ "exp_name": "run",
6
+ "frames": 28180,
7
+ "world_size": 32,
8
+ "per_gpu_batch": 96,
9
+ "global_batch": 3072,
10
+ "epochs": 20.0,
11
+ "num_train_steps": 184,
12
+ "save_steps": [
13
+ 92,
14
+ 184
15
+ ],
16
+ "warmup_steps": 9,
17
+ "peak_lr": 5e-05,
18
+ "weights": "/mnt/nas/cross/Team/RobotAgent/Personal/baijiaqi/humanoid-benchmark-work/preflight-20261104T1648Z/policies-20261109/models/openpi/pi05_base_g1a76_pytorch",
19
+ "checkpoint_dir": "/mnt/nas/cross/Team/RobotAgent/Personal/baijiaqi/humanoid-benchmark-work/preflight-20261104T1648Z/policies-20261109/pi05-runs/chain-hb1110/a3/run/ckpt/pi05_hb_g1_a3/run",
20
+ "augment": true,
21
+ "aug_version": "hb-aug-v2-20261109",
22
+ "max_token_len": 256
23
+ }
a3-fridge-open/epoch-020/assets/g1_a3/norm_stats.json ADDED
@@ -0,0 +1,504 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "norm_stats": {
3
+ "state": {
4
+ "mean": [
5
+ -0.05641397461295128,
6
+ 0.07833150029182434,
7
+ 0.11269253492355347,
8
+ 0.3092529773712158,
9
+ -0.18253354728221893,
10
+ -0.021320076659321785,
11
+ -0.10698527842760086,
12
+ -0.05834974721074104,
13
+ -0.20540674030780792,
14
+ 0.3034929633140564,
15
+ -0.1302507370710373,
16
+ -0.013441003859043121,
17
+ -0.030619896948337555,
18
+ -0.03488051891326904,
19
+ 0.020995844155550003,
20
+ 0.19224053621292114,
21
+ 0.14336878061294556,
22
+ 0.006716981064528227,
23
+ -0.1879279464483261,
24
+ -0.0025916018057614565,
25
+ 0.005836201831698418,
26
+ -0.010614745318889618,
27
+ 0.15154524147510529,
28
+ -0.17589232325553894,
29
+ -0.2622005045413971,
30
+ -0.47192373871803284,
31
+ 0.2191326767206192,
32
+ 0.20637701451778412,
33
+ -0.009809371083974838,
34
+ 0.999921977519989,
35
+ 0.9999970197677612,
36
+ 0.9999962449073792,
37
+ 0.9999969005584717,
38
+ 0.9999812245368958,
39
+ 0.004240452777594328,
40
+ 0.8682456612586975,
41
+ 0.869608461856842,
42
+ 0.873170793056488,
43
+ 0.881648600101471,
44
+ 0.8788400888442993,
45
+ 0.004410883411765099,
46
+ -0.03297373279929161,
47
+ 0.009942770935595036,
48
+ -0.9982253909111023
49
+ ],
50
+ "std": [
51
+ 0.15939587354660034,
52
+ 0.06632916629314423,
53
+ 0.07504713535308838,
54
+ 0.17704583704471588,
55
+ 0.1108374074101448,
56
+ 0.04978697746992111,
57
+ 0.1295444816350937,
58
+ 0.062313150614500046,
59
+ 0.08294235169887543,
60
+ 0.1889764368534088,
61
+ 0.1166033148765564,
62
+ 0.05051586776971817,
63
+ 0.06022709235548973,
64
+ 0.05384501442313194,
65
+ 0.02067396230995655,
66
+ 0.02285316400229931,
67
+ 0.012044434435665607,
68
+ 0.013775781728327274,
69
+ 0.021458350121974945,
70
+ 0.009183689951896667,
71
+ 0.011165750212967396,
72
+ 0.014735972508788109,
73
+ 0.3950332999229431,
74
+ 0.13744504749774933,
75
+ 0.37059733271598816,
76
+ 0.38896965980529785,
77
+ 0.19889448583126068,
78
+ 0.24846522510051727,
79
+ 0.08681652694940567,
80
+ 0.0,
81
+ 0.0,
82
+ 0.0,
83
+ 0.0,
84
+ 0.0,
85
+ 0.030091701075434685,
86
+ 0.21434949338436127,
87
+ 0.2120189517736435,
88
+ 0.2065865397453308,
89
+ 0.1933792531490326,
90
+ 0.1923707276582718,
91
+ 0.030065977945923805,
92
+ 0.03751828148961067,
93
+ 0.031833503395318985,
94
+ 0.0
95
+ ],
96
+ "q01": [
97
+ -0.46506815473136753,
98
+ -0.04889196949116753,
99
+ -0.03771492845147742,
100
+ 0.12403084383573584,
101
+ -0.4327747708072943,
102
+ -0.1480070360368902,
103
+ -0.4722705717922792,
104
+ -0.21336711200995134,
105
+ -0.4257794928347497,
106
+ 0.14030948715447483,
107
+ -0.4798982841342425,
108
+ -0.12194732715222922,
109
+ -0.2540194292733686,
110
+ -0.2156617373707264,
111
+ -0.03905836607585417,
112
+ 0.13080863007519333,
113
+ 0.11221892317597773,
114
+ -0.030211272256092374,
115
+ -0.2578536881135155,
116
+ -0.033230624382752746,
117
+ -0.01635900642697462,
118
+ -0.05323737701858385,
119
+ -1.0001329558172423,
120
+ -0.5889036026173377,
121
+ -1.3324901583751518,
122
+ -1.0482400876077553,
123
+ -0.043778369650742166,
124
+ -0.43015055529414703,
125
+ -0.2122822610443819,
126
+ 0.9979524703889018,
127
+ 0.9998749794303501,
128
+ 0.9998566181770135,
129
+ 0.9998688959205332,
130
+ 0.9992656699177477,
131
+ -1e-10,
132
+ 0.49980709718610594,
133
+ 0.49981258509911924,
134
+ 0.4997684336261139,
135
+ 0.509475866386762,
136
+ 0.5038498614157283,
137
+ -1e-10,
138
+ -0.1147158351111296,
139
+ -0.07648704438251047,
140
+ -0.9999822119043099
141
+ ],
142
+ "q99": [
143
+ 0.20402788861789334,
144
+ 0.2784175859561666,
145
+ 0.3222591326691111,
146
+ 1.0027285105278778,
147
+ 0.14594734426464373,
148
+ 0.10093794696942421,
149
+ 0.13905996096478002,
150
+ 0.06319464641553324,
151
+ -0.03222518681984321,
152
+ 1.0404026509018791,
153
+ 0.11550482268441065,
154
+ 0.116426306980423,
155
+ 0.07085631988559993,
156
+ 0.0778610076478406,
157
+ 0.07582880456800241,
158
+ 0.25341256991924554,
159
+ 0.17726504346664093,
160
+ 0.042624652664472365,
161
+ -0.14470140677618784,
162
+ 0.019137383483876676,
163
+ 0.03877572672478539,
164
+ 0.005319607630962106,
165
+ 1.2571805590330305,
166
+ 0.13922517023041814,
167
+ 0.1538355433300771,
168
+ 0.8272105335782394,
169
+ 0.6667848346517331,
170
+ 0.6415972574619173,
171
+ 0.20692930209156313,
172
+ 0.9999940131710202,
173
+ 0.9999995318577916,
174
+ 0.9999999581263691,
175
+ 0.9999999589369923,
176
+ 0.9999998499916226,
177
+ 0.21753646457975004,
178
+ 0.9998999014602333,
179
+ 0.9998998424277932,
180
+ 0.999899813525501,
181
+ 0.9999008442209394,
182
+ 0.9999001911805303,
183
+ 0.21746011237681775,
184
+ 0.07922326138295144,
185
+ 0.10487203598122224,
186
+ -0.9906932630281675
187
+ ]
188
+ },
189
+ "actions": {
190
+ "mean": [
191
+ 0.15783676505088806,
192
+ 0.16300027072429657,
193
+ -0.04454663768410683,
194
+ -0.019376136362552643,
195
+ 0.03434408828616142,
196
+ 0.04013373702764511,
197
+ -0.0696130320429802,
198
+ 0.16177308559417725,
199
+ 0.09804673492908478,
200
+ -0.21799813210964203,
201
+ -0.1298724114894867,
202
+ -0.14870725572109222,
203
+ -0.009640868753194809,
204
+ 0.09577104449272156,
205
+ 0.17926457524299622,
206
+ -0.0506000779569149,
207
+ 0.16960610449314117,
208
+ 0.13840295374393463,
209
+ -0.0558854341506958,
210
+ -0.15770410001277924,
211
+ -0.1273193508386612,
212
+ -0.07514706999063492,
213
+ -0.04349217191338539,
214
+ 0.0012820761185139418,
215
+ 0.08928761631250381,
216
+ -0.0955679714679718,
217
+ -0.12635554373264313,
218
+ 0.01848260685801506,
219
+ 0.009467734955251217,
220
+ 0.12164337188005447,
221
+ 0.12644965946674347,
222
+ -0.0803895965218544,
223
+ -0.014793431386351585,
224
+ 0.19501863420009613,
225
+ 0.04290221258997917,
226
+ 0.03360012546181679,
227
+ -0.01547778770327568,
228
+ 0.06523936241865158,
229
+ 0.08329988270998001,
230
+ -0.06897857785224915,
231
+ -0.03849208727478981,
232
+ -0.057777289301157,
233
+ 0.019498951733112335,
234
+ -0.1662408858537674,
235
+ 0.06778231263160706,
236
+ -0.13049370050430298,
237
+ -0.02015572413802147,
238
+ -0.014168542809784412,
239
+ -0.16451606154441833,
240
+ -0.060739967972040176,
241
+ 0.09654980152845383,
242
+ 0.16325144469738007,
243
+ -0.02566051296889782,
244
+ -0.21247075498104095,
245
+ -0.009581679478287697,
246
+ 0.05965486913919449,
247
+ -0.11509575694799423,
248
+ 0.19388903677463531,
249
+ -0.003796325996518135,
250
+ -0.16428831219673157,
251
+ -0.08977831900119781,
252
+ 0.12237153202295303,
253
+ 0.20675969123840332,
254
+ -0.04955295845866203,
255
+ 0.9999598860740662,
256
+ 0.9999998211860657,
257
+ 0.9999997019767761,
258
+ 0.9999997019767761,
259
+ 0.9999992847442627,
260
+ 0.0007821584003977478,
261
+ 0.8682477474212646,
262
+ 0.8695738911628723,
263
+ 0.8731648921966553,
264
+ 0.8815546035766602,
265
+ 0.8789401650428772,
266
+ 0.0009576878510415554
267
+ ],
268
+ "std": [
269
+ 0.08065593242645264,
270
+ 0.08959350734949112,
271
+ 0.11977304518222809,
272
+ 0.10043507069349289,
273
+ 0.14305159449577332,
274
+ 0.08850719034671783,
275
+ 0.14513453841209412,
276
+ 0.12789000570774078,
277
+ 0.06797633320093155,
278
+ 0.050453633069992065,
279
+ 0.0903969332575798,
280
+ 0.06595497578382492,
281
+ 0.11297961324453354,
282
+ 0.07852301746606827,
283
+ 0.08324233442544937,
284
+ 0.08588460087776184,
285
+ 0.0624857097864151,
286
+ 0.09392557293176651,
287
+ 0.11839237064123154,
288
+ 0.0951315388083458,
289
+ 0.08514631539583206,
290
+ 0.09915973991155624,
291
+ 0.06709878891706467,
292
+ 0.0878397673368454,
293
+ 0.0614795908331871,
294
+ 0.08896732330322266,
295
+ 0.06850738823413849,
296
+ 0.08969742804765701,
297
+ 0.11463499069213867,
298
+ 0.08277275413274765,
299
+ 0.10993892699480057,
300
+ 0.07832051813602448,
301
+ 0.0719744935631752,
302
+ 0.07010309398174286,
303
+ 0.04459095746278763,
304
+ 0.07442177087068558,
305
+ 0.10430904477834702,
306
+ 0.1396985650062561,
307
+ 0.08694476634263992,
308
+ 0.07248101383447647,
309
+ 0.08530312776565552,
310
+ 0.1339278668165207,
311
+ 0.082322858273983,
312
+ 0.06349117308855057,
313
+ 0.11075259000062943,
314
+ 0.08747898787260056,
315
+ 0.10215355455875397,
316
+ 0.13771311938762665,
317
+ 0.10990352183580399,
318
+ 0.11261215806007385,
319
+ 0.08981925994157791,
320
+ 0.06594730913639069,
321
+ 0.16168172657489777,
322
+ 0.09319587051868439,
323
+ 0.06820440292358398,
324
+ 0.12118405848741531,
325
+ 0.11162112653255463,
326
+ 0.045856211334466934,
327
+ 0.11358722299337387,
328
+ 0.08551610261201859,
329
+ 0.07454151660203934,
330
+ 0.0910487249493599,
331
+ 0.06291131675243378,
332
+ 0.11520238220691681,
333
+ 0.0,
334
+ 0.0,
335
+ 0.0,
336
+ 0.0,
337
+ 0.0,
338
+ 0.01162454579025507,
339
+ 0.21439313888549805,
340
+ 0.21226760745048523,
341
+ 0.20670393109321594,
342
+ 0.1937810778617859,
343
+ 0.1922834813594818,
344
+ 0.011621065437793732
345
+ ],
346
+ "q01": [
347
+ -7.499999232774002e-11,
348
+ -0.06250000008,
349
+ -0.25003750006923997,
350
+ -0.25001250006364,
351
+ -0.25007500005716,
352
+ -0.18750000003999998,
353
+ -0.37507500005716,
354
+ -0.12507500008184003,
355
+ -0.06252500006668,
356
+ -0.31252500006668,
357
+ -0.312500000075,
358
+ -0.3125000001,
359
+ -0.25001250006364,
360
+ -0.06252500007144002,
361
+ -0.06250000003999998,
362
+ -0.25001250005555997,
363
+ -0.000012500055559999668,
364
+ -0.062500000075,
365
+ -0.25001250006364,
366
+ -0.37501250005555997,
367
+ -0.31251250005555997,
368
+ -0.37505000007144,
369
+ -0.25001250005555997,
370
+ -0.18751250005556,
371
+ -0.06252500006668,
372
+ -0.3125625000778,
373
+ -0.25000000005,
374
+ -0.25001250006364,
375
+ -0.25000000005,
376
+ -0.06256250007780001,
377
+ -0.06250000005,
378
+ -0.2500625000778,
379
+ -0.18750000005,
380
+ -0.000025000071439996996,
381
+ -0.06250000002,
382
+ -0.12500000005,
383
+ -0.25010000008336003,
384
+ -0.25005000007144,
385
+ -0.1250625000778,
386
+ -0.2500000001,
387
+ -0.25000000008,
388
+ -0.31253750006923997,
389
+ -0.18751250005556,
390
+ -0.31252500007144,
391
+ -0.18757500008184003,
392
+ -0.3125625000778,
393
+ -0.25000000006,
394
+ -0.31250000006,
395
+ -0.4375000001,
396
+ -0.37505000007144,
397
+ -0.12500000006,
398
+ -1e-10,
399
+ -0.43760000008752,
400
+ -0.37500000006,
401
+ -0.12505000004288003,
402
+ -0.25010000008336003,
403
+ -0.37507500008184,
404
+ 0.12497499993332001,
405
+ -0.25010000003335997,
406
+ -0.31251250005555997,
407
+ -0.187500000025,
408
+ -0.12500000008,
409
+ 0.062474999928559996,
410
+ -0.37510000008464,
411
+ 0.9979524703889018,
412
+ 0.9998749794303501,
413
+ 0.9998566181770135,
414
+ 0.9998688959205332,
415
+ 0.9995942571558815,
416
+ 0.0,
417
+ 0.49980709718610594,
418
+ 0.49981258509911924,
419
+ 0.4997684336261139,
420
+ 0.509475866386762,
421
+ 0.5038498614157283,
422
+ 0.0
423
+ ],
424
+ "q99": [
425
+ 0.37490000007496,
426
+ 0.37487500005996005,
427
+ 0.31237500006920005,
428
+ 0.24993750008179993,
429
+ 0.37485000008568004,
430
+ 0.2498750000999601,
431
+ 0.24985000008568004,
432
+ 0.43743750008179993,
433
+ 0.24992500009996005,
434
+ -0.06254999993335997,
435
+ 0.062400000074959994,
436
+ -0.00003749995716001431,
437
+ 0.24993750008179993,
438
+ 0.24993750007139998,
439
+ 0.3123750000799601,
440
+ 0.12495000007776003,
441
+ 0.31240000005552004,
442
+ 0.31240000007496,
443
+ 0.24993750008179993,
444
+ -0.00004999992223997385,
445
+ 0.062450000077760026,
446
+ 0.12492500004284002,
447
+ 0.12495000007776003,
448
+ 0.18745000007776003,
449
+ 0.18745000006664003,
450
+ 0.12495000007776003,
451
+ 0.062400000074959994,
452
+ 0.18737500006359997,
453
+ 0.24995000008332002,
454
+ 0.31240000005552004,
455
+ 0.37490000006664004,
456
+ 0.12490000005552004,
457
+ 0.12490000007496,
458
+ 0.3124375000714,
459
+ 0.12493750009996002,
460
+ 0.18740000007496,
461
+ 0.31240000006664004,
462
+ 0.43735000008568004,
463
+ 0.31245000007776,
464
+ 0.12493750007139998,
465
+ 0.18737500005996005,
466
+ 0.31243750008460003,
467
+ 0.18745000007776003,
468
+ -0.06253749995716001,
469
+ 0.37493750008179993,
470
+ 0.12495000007776003,
471
+ 0.1873750000799601,
472
+ 0.31243750007332005,
473
+ 0.06237500005996005,
474
+ 0.18740000005712,
475
+ 0.3123750000799601,
476
+ 0.24993750005996,
477
+ 0.37480000007495995,
478
+ 0.06237500007996011,
479
+ 0.12493750007139998,
480
+ 0.31240000006664004,
481
+ 0.18743750008179993,
482
+ 0.31247500003332,
483
+ 0.18745000008332002,
484
+ 0.062450000077760026,
485
+ 0.062400000074959994,
486
+ 0.31237500005996005,
487
+ 0.31246250004284,
488
+ 0.24987500006920005,
489
+ 0.9999940131710202,
490
+ 0.9999995318577916,
491
+ 0.9999999581263691,
492
+ 0.9999999589369923,
493
+ 0.9999998499916226,
494
+ 0.30305999517440796,
495
+ 0.9998999014602333,
496
+ 0.9998998424277932,
497
+ 0.999899813525501,
498
+ 0.9999008442209394,
499
+ 0.9999001911805303,
500
+ 0.3030380606651306
501
+ ]
502
+ }
503
+ }
504
+ }
a3-fridge-open/epoch-020/metadata.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:197030de119da0e36a1edb9a7526e252787a6b35ae7f736c50708bf280bf4e47
3
+ size 3379
a3-fridge-open/epoch-020/model.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:cb08cb26cd7ddcc6a947ca4a109bbf66e0de515cf897b65707e4f4ccfdabe8f1
3
+ size 7473452088
a3-fridge-open/epoch-020/protocol.json ADDED
@@ -0,0 +1,56 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "model": "pi0.5 (openpi, PyTorch weights)",
3
+ "task": "A3 open fridge door (G1)",
4
+ "task_dir": "a3-fridge-open",
5
+ "openpi_config": "pi05_hb_g1_a3",
6
+ "layout": "g1",
7
+ "asset_id": "g1_a3",
8
+ "epoch_completed": 20,
9
+ "global_step": 184,
10
+ "total_epochs": 20.0,
11
+ "total_steps": 184,
12
+ "global_batch": 3072,
13
+ "per_gpu_batch": 96,
14
+ "world_size": 32,
15
+ "peak_lr": 5e-05,
16
+ "warmup_steps": 9,
17
+ "max_token_len": 256,
18
+ "init_weights": "lerobot/pi05_base converted to PyTorch, action/state projections resized to 76-D ($P/policies-20261109/models/openpi/pi05_base_g1a76_pytorch)",
19
+ "augmentation": "Humanoid Benchmark online augmentation spec v2 (2026-11-09), on from epoch 1; model-native augmentation off",
20
+ "aug_version": "hb-aug-v2-20261109",
21
+ "model_action_dim": 76,
22
+ "action_horizon": 16,
23
+ "episodes": 100,
24
+ "frames": 28180,
25
+ "fps": 30.0,
26
+ "control_hz": 30.0,
27
+ "prompt": "Walk to the refrigerator, pull its upper door open with your right hand, release the handle and pull your hand back.",
28
+ "raw_state_dim": 41,
29
+ "state_layout": "body29(observation.state by name) | inspire_left6 | inspire_right6 | projected_gravity3",
30
+ "action_layout": "action.motion_token64 | inspire_left6 | inspire_right6 (same-row observed hands), rows t..t+15",
31
+ "cameras": {
32
+ "base_0_rgb": "observation.images.ego_view",
33
+ "left_wrist_0_rgb": "observation.images.chest",
34
+ "right_wrist_0_rgb": "zeros, image_mask False"
35
+ },
36
+ "raw_image_hw": [
37
+ 480,
38
+ 640
39
+ ],
40
+ "content_box": [
41
+ 28,
42
+ 196,
43
+ 0,
44
+ 224
45
+ ],
46
+ "training_data": "$P/a3-a7-20261109/a3-fridge-open-expert-v1/780fcf1e4bd7db947a8c2d8f7b92e61f24f71b45",
47
+ "normalization": "assets/g1_a3/norm_stats.json (openpi quantile normalization)",
48
+ "inference": "Register configs by importing code/hb_pi05/hb_g1_config.py (needs hb_g1_data.py on the path), then openpi policy_config.create_trained_policy(get_config(openpi_config), <this folder>). Clients send observation/<camera> raw RGB, observation/state and prompt; see HBInputs.",
49
+ "resume_state_included": false,
50
+ "files": {
51
+ "assets/g1_a3/norm_stats.json": 13795,
52
+ "metadata.pt": 3379,
53
+ "model.safetensors": 7473452088,
54
+ "training_plan.json": 740
55
+ }
56
+ }
a3-fridge-open/epoch-020/training_plan.json ADDED
@@ -0,0 +1,23 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "task": "a3",
3
+ "layout": "g1",
4
+ "config": "pi05_hb_g1_a3",
5
+ "exp_name": "run",
6
+ "frames": 28180,
7
+ "world_size": 32,
8
+ "per_gpu_batch": 96,
9
+ "global_batch": 3072,
10
+ "epochs": 20.0,
11
+ "num_train_steps": 184,
12
+ "save_steps": [
13
+ 92,
14
+ 184
15
+ ],
16
+ "warmup_steps": 9,
17
+ "peak_lr": 5e-05,
18
+ "weights": "/mnt/nas/cross/Team/RobotAgent/Personal/baijiaqi/humanoid-benchmark-work/preflight-20261104T1648Z/policies-20261109/models/openpi/pi05_base_g1a76_pytorch",
19
+ "checkpoint_dir": "/mnt/nas/cross/Team/RobotAgent/Personal/baijiaqi/humanoid-benchmark-work/preflight-20261104T1648Z/policies-20261109/pi05-runs/chain-hb1110/a3/run/ckpt/pi05_hb_g1_a3/run",
20
+ "augment": true,
21
+ "aug_version": "hb-aug-v2-20261109",
22
+ "max_token_len": 256
23
+ }
a7-cabinet-close/epoch-010/assets/g1_a7/norm_stats.json ADDED
@@ -0,0 +1,504 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "norm_stats": {
3
+ "state": {
4
+ "mean": [
5
+ -0.35561004281044006,
6
+ 0.037816599011421204,
7
+ 0.09485039860010147,
8
+ 0.5194756388664246,
9
+ -0.1980385035276413,
10
+ -0.006605133879929781,
11
+ -0.3908240795135498,
12
+ -0.1153804287314415,
13
+ -0.2100031077861786,
14
+ 0.5305911898612976,
15
+ -0.18644081056118011,
16
+ -0.0369383879005909,
17
+ 0.015539770945906639,
18
+ 0.013902479782700539,
19
+ 0.016449792310595512,
20
+ 0.20338064432144165,
21
+ 0.1432969570159912,
22
+ 0.03276205062866211,
23
+ -0.21513384580612183,
24
+ 0.004481739364564419,
25
+ 0.007482144050300121,
26
+ -0.02025328017771244,
27
+ -0.20058324933052063,
28
+ -0.07724551111459732,
29
+ 0.036659061908721924,
30
+ 0.21988210082054138,
31
+ 0.014881269074976444,
32
+ 0.009545939043164253,
33
+ 0.0062908814288675785,
34
+ 0.9994741678237915,
35
+ 0.9999215006828308,
36
+ 0.9999319911003113,
37
+ 0.9999361038208008,
38
+ 0.9997094869613647,
39
+ 0.0039056746754795313,
40
+ 0.9999874234199524,
41
+ 0.9999957084655762,
42
+ 0.9999594688415527,
43
+ 0.9998860359191895,
44
+ 0.9993945956230164,
45
+ 0.004138351418077946,
46
+ 0.08264563977718353,
47
+ -0.01804521679878235,
48
+ -0.986268162727356
49
+ ],
50
+ "std": [
51
+ 0.28290310502052307,
52
+ 0.09233158826828003,
53
+ 0.19659872353076935,
54
+ 0.20531529188156128,
55
+ 0.08587288856506348,
56
+ 0.07257651537656784,
57
+ 0.2948111593723297,
58
+ 0.09278968721628189,
59
+ 0.14734666049480438,
60
+ 0.20262238383293152,
61
+ 0.09397441893815994,
62
+ 0.07671834528446198,
63
+ 0.05012736842036247,
64
+ 0.03560950979590416,
65
+ 0.024248890578746796,
66
+ 0.03378785029053688,
67
+ 0.025765331462025642,
68
+ 0.035841915756464005,
69
+ 0.03535226732492447,
70
+ 0.012149378657341003,
71
+ 0.02936926670372486,
72
+ 0.03796697035431862,
73
+ 0.4766698181629181,
74
+ 0.12800851464271545,
75
+ 0.08127840608358383,
76
+ 0.5201791524887085,
77
+ 0.03146743029356003,
78
+ 0.030610734596848488,
79
+ 0.08500799536705017,
80
+ 0.004740353673696518,
81
+ 0.0,
82
+ 0.0,
83
+ 0.0,
84
+ 0.0,
85
+ 0.02886965498328209,
86
+ 0.0,
87
+ 0.0,
88
+ 0.0,
89
+ 0.0,
90
+ 0.0,
91
+ 0.028847994282841682,
92
+ 0.1356819123029709,
93
+ 0.038989145308732986,
94
+ 0.014254551380872726
95
+ ],
96
+ "q01": [
97
+ -0.8435057642134977,
98
+ -0.1539409343454199,
99
+ -0.38289439146854154,
100
+ 0.19744580789334945,
101
+ -0.4023641065573069,
102
+ -0.15783057514232443,
103
+ -0.8902416186952684,
104
+ -0.3213813964588269,
105
+ -0.580422671914016,
106
+ 0.1688319038662181,
107
+ -0.3692423825023103,
108
+ -0.1722233697723961,
109
+ -0.05082127009701552,
110
+ -0.05095853811305899,
111
+ -0.04851158314737884,
112
+ 0.12016225313020878,
113
+ 0.07590308442343385,
114
+ -0.03233522117337442,
115
+ -0.33273061226802,
116
+ -0.023371822398616807,
117
+ -0.07994668760927265,
118
+ -0.10301587557162084,
119
+ -1.2144301321222808,
120
+ -0.3396789082267284,
121
+ -0.08274552846634385,
122
+ -0.40091806847836514,
123
+ -0.0650506398949991,
124
+ -0.11048650279830803,
125
+ -0.21627054291998973,
126
+ 0.9829885311015435,
127
+ 0.9513298273086548,
128
+ 0.9574251770973206,
129
+ 0.9504732489585876,
130
+ 0.851092517375946,
131
+ -1e-10,
132
+ 0.9825374484062195,
133
+ 0.9934524297714233,
134
+ 0.9735382199287415,
135
+ 0.9164777994155884,
136
+ 0.9852659983204475,
137
+ -1e-10,
138
+ -0.13615792221606549,
139
+ -0.12976165536695589,
140
+ -0.9999653876588571
141
+ ],
142
+ "q99": [
143
+ 0.19439826722640507,
144
+ 0.24245334795716866,
145
+ 0.45560911755377265,
146
+ 1.1228674726221726,
147
+ 0.008258910985718626,
148
+ 0.12484455139244438,
149
+ 0.1404618034872802,
150
+ 0.06307184297845925,
151
+ 0.06084788480354464,
152
+ 0.9531215072192949,
153
+ 0.07626591385718673,
154
+ 0.15503785397921882,
155
+ 0.1525609748567458,
156
+ 0.08131745231611116,
157
+ 0.08366167413077756,
158
+ 0.3211568147841109,
159
+ 0.20607530826763967,
160
+ 0.15977249451081388,
161
+ -0.1695193791277912,
162
+ 0.04427388410215958,
163
+ 0.06231649593288226,
164
+ 0.07538968521664166,
165
+ 0.39386444146296595,
166
+ 0.16625795727069403,
167
+ 0.2778081297818395,
168
+ 1.3490795560201072,
169
+ 0.095980245563843,
170
+ 0.07410926639083143,
171
+ 0.13639384343573485,
172
+ 0.999956823782745,
173
+ 1.0,
174
+ 1.0,
175
+ 1.0,
176
+ 1.0,
177
+ 0.16704642294477218,
178
+ 1.0,
179
+ 1.0,
180
+ 1.0,
181
+ 1.0,
182
+ 0.999987639360252,
183
+ 0.16709456220701158,
184
+ 0.3249750240081849,
185
+ 0.06491606969695807,
186
+ -0.9452246287075486
187
+ ]
188
+ },
189
+ "actions": {
190
+ "mean": [
191
+ 0.10757830739021301,
192
+ 0.0010036963503807783,
193
+ -0.1955849826335907,
194
+ 0.03575943410396576,
195
+ -0.02901560068130493,
196
+ 0.10509791970252991,
197
+ 0.15584516525268555,
198
+ -0.036249756813049316,
199
+ 0.00029133816133253276,
200
+ -0.19356472790241241,
201
+ -0.06754574179649353,
202
+ -0.07026422023773193,
203
+ -0.10356352478265762,
204
+ 0.1086534708738327,
205
+ 0.179977685213089,
206
+ -0.09078477323055267,
207
+ 0.04756608605384827,
208
+ -0.006199375726282597,
209
+ -0.1645488739013672,
210
+ 0.03273191303014755,
211
+ 0.013353785499930382,
212
+ -0.0864969938993454,
213
+ 0.033461980521678925,
214
+ 0.149399533867836,
215
+ 0.09672172367572784,
216
+ 0.03816988691687584,
217
+ -0.13994601368904114,
218
+ 0.0559534914791584,
219
+ 0.1010548546910286,
220
+ -0.025988206267356873,
221
+ 0.12739083170890808,
222
+ -0.08735954016447067,
223
+ -0.002861789194867015,
224
+ 0.1596086025238037,
225
+ -0.018322203308343887,
226
+ 0.007539633195847273,
227
+ 0.04284964129328728,
228
+ -0.09214604645967484,
229
+ 0.14863164722919464,
230
+ 0.016650989651679993,
231
+ 0.027098143473267555,
232
+ -0.08947993069887161,
233
+ 0.05657769367098808,
234
+ -0.11662724614143372,
235
+ 0.0791701003909111,
236
+ -0.10034364461898804,
237
+ -0.09565547853708267,
238
+ -0.2216619849205017,
239
+ 0.008683966472744942,
240
+ -0.15533827245235443,
241
+ 0.0966157391667366,
242
+ -0.003970836289227009,
243
+ 0.08637572079896927,
244
+ -0.118191197514534,
245
+ 0.1010720506310463,
246
+ 0.14126704633235931,
247
+ -0.13963939249515533,
248
+ 0.07252243906259537,
249
+ 0.0847059041261673,
250
+ -0.04034632444381714,
251
+ -0.047200094908475876,
252
+ -0.002455164212733507,
253
+ 0.15698553621768951,
254
+ -0.03268032148480415,
255
+ 0.9995062351226807,
256
+ 0.9999467134475708,
257
+ 0.9999535083770752,
258
+ 0.9999532103538513,
259
+ 0.9997615218162537,
260
+ 0.0007274940144270658,
261
+ 0.9999932050704956,
262
+ 0.999999463558197,
263
+ 0.9999749064445496,
264
+ 0.9999121427536011,
265
+ 0.9996171593666077,
266
+ 0.0009604249498806894
267
+ ],
268
+ "std": [
269
+ 0.06684934347867966,
270
+ 0.12237707525491714,
271
+ 0.12836524844169617,
272
+ 0.1443842351436615,
273
+ 0.12026809900999069,
274
+ 0.06686931103467941,
275
+ 0.08585774153470993,
276
+ 0.12531286478042603,
277
+ 0.07042475789785385,
278
+ 0.04380151629447937,
279
+ 0.0754399374127388,
280
+ 0.07911296933889389,
281
+ 0.10523597151041031,
282
+ 0.04984365031123161,
283
+ 0.10585196316242218,
284
+ 0.1417231559753418,
285
+ 0.08640342950820923,
286
+ 0.09490349888801575,
287
+ 0.10842842608690262,
288
+ 0.12022152543067932,
289
+ 0.11372580379247665,
290
+ 0.08418825268745422,
291
+ 0.09572859108448029,
292
+ 0.11790358275175095,
293
+ 0.04139557480812073,
294
+ 0.11743805557489395,
295
+ 0.04096134379506111,
296
+ 0.1019567996263504,
297
+ 0.08468007296323776,
298
+ 0.09862369298934937,
299
+ 0.08369130641222,
300
+ 0.07918284833431244,
301
+ 0.12212815880775452,
302
+ 0.07682932168245316,
303
+ 0.06639379262924194,
304
+ 0.06500166654586792,
305
+ 0.1091652363538742,
306
+ 0.1271735578775406,
307
+ 0.07222087681293488,
308
+ 0.11637697368860245,
309
+ 0.09275760501623154,
310
+ 0.06542796641588211,
311
+ 0.064761683344841,
312
+ 0.04859273508191109,
313
+ 0.10564269870519638,
314
+ 0.05467769876122475,
315
+ 0.15121139585971832,
316
+ 0.13371562957763672,
317
+ 0.0961233377456665,
318
+ 0.11129939556121826,
319
+ 0.09783373773097992,
320
+ 0.12376602739095688,
321
+ 0.17745254933834076,
322
+ 0.050024744123220444,
323
+ 0.07118535041809082,
324
+ 0.15862050652503967,
325
+ 0.10293488204479218,
326
+ 0.11946934461593628,
327
+ 0.06005885452032089,
328
+ 0.11131948232650757,
329
+ 0.10331486165523529,
330
+ 0.06897100061178207,
331
+ 0.07636360079050064,
332
+ 0.07862718403339386,
333
+ 0.0,
334
+ 0.0,
335
+ 0.0,
336
+ 0.0,
337
+ 0.0,
338
+ 0.011146476492285728,
339
+ 0.0,
340
+ 0.0,
341
+ 0.0,
342
+ 0.0,
343
+ 0.0,
344
+ 0.011168749071657658
345
+ ],
346
+ "q01": [
347
+ -5.999999108663445e-11,
348
+ -0.2500000001,
349
+ -0.43750000006,
350
+ -0.25001250006364,
351
+ -0.2500625000778,
352
+ -0.0625000001,
353
+ -0.0625000001,
354
+ -0.250000000075,
355
+ -0.1250000001,
356
+ -0.25000000005,
357
+ -0.25000000005,
358
+ -0.25002500007144,
359
+ -0.312500000075,
360
+ -0.06252500006668,
361
+ -0.00006250007779998873,
362
+ -0.37507500008184,
363
+ -0.12502500006668,
364
+ -0.18751250005556,
365
+ -0.3750625000778,
366
+ -0.187500000075,
367
+ -0.2500625000778,
368
+ -0.31251250005555997,
369
+ -0.125000000075,
370
+ -0.12500000008,
371
+ -1e-10,
372
+ -0.1875000001,
373
+ -0.25000000006,
374
+ -0.125000000075,
375
+ -0.1250625000778,
376
+ -0.18752500007144002,
377
+ -0.062500000075,
378
+ -0.25002500007144,
379
+ -0.2500625000778,
380
+ -0.000025000071439996996,
381
+ -0.12500000006,
382
+ -0.18752500007144002,
383
+ -0.12500000008,
384
+ -0.3750000001,
385
+ -1e-10,
386
+ -0.1875000001,
387
+ -0.187500000075,
388
+ -0.25002500007144,
389
+ -0.06252500006668,
390
+ -0.18752500006668,
391
+ -0.062500000075,
392
+ -0.25002500006668,
393
+ -0.37500000008,
394
+ -0.50000000008,
395
+ -0.1875000001,
396
+ -0.4375625000778,
397
+ -0.18750000008,
398
+ -0.1875000001,
399
+ -0.1875000001,
400
+ -0.25002500006668,
401
+ -0.06252500006668,
402
+ -0.12507500008184003,
403
+ -0.375000000075,
404
+ -0.1250000001,
405
+ -0.06252500006668,
406
+ -0.250000000075,
407
+ -0.187500000075,
408
+ -0.12502500006668,
409
+ -0.0625000001,
410
+ -0.18752500007144002,
411
+ 0.9829885311015435,
412
+ 0.9513298273086548,
413
+ 0.9574251770973206,
414
+ 0.9504732489585876,
415
+ 0.851092517375946,
416
+ 0.0,
417
+ 0.9825374484062195,
418
+ 0.9934524297714233,
419
+ 0.9735382199287415,
420
+ 0.9164777994155884,
421
+ 0.9852659983204475,
422
+ 0.0
423
+ ],
424
+ "q99": [
425
+ 0.24993750009996002,
426
+ 0.24990000009996,
427
+ -0.00012499992003989213,
428
+ 0.3123625000999599,
429
+ 0.18745000007776003,
430
+ 0.24995000006664003,
431
+ 0.31242500009996005,
432
+ 0.18740000009996,
433
+ 0.12493750005995999,
434
+ -0.12504999995004001,
435
+ 0.062400000074959994,
436
+ 0.062437500071399976,
437
+ 0.062400000074959994,
438
+ 0.18745000006664003,
439
+ 0.43745000007776,
440
+ 0.18743750008179993,
441
+ 0.18742500009996005,
442
+ 0.18745000007776003,
443
+ 0.062450000077760026,
444
+ 0.24990000009996,
445
+ 0.18745000007776003,
446
+ 0.062450000077760026,
447
+ 0.24990000007496,
448
+ 0.3748750000799601,
449
+ 0.18745000004995999,
450
+ 0.24991250009995997,
451
+ -0.06256249994004001,
452
+ 0.24990000007496,
453
+ 0.24990000005552004,
454
+ 0.18741250009995997,
455
+ 0.31240000007496,
456
+ 0.062437500071399976,
457
+ 0.24988750009996008,
458
+ 0.3124375000714,
459
+ 0.12493750009996002,
460
+ 0.12493750007139998,
461
+ 0.3748750000799601,
462
+ 0.12495000007776003,
463
+ 0.31246250004284,
464
+ 0.24991250009995997,
465
+ 0.18740000007496,
466
+ 0.062437500071399976,
467
+ 0.18745000006664003,
468
+ -0.00002499996667998383,
469
+ 0.31240000007496,
470
+ -0.00004999993335996766,
471
+ 0.12487500007996011,
472
+ -0.00012499992003989213,
473
+ 0.18740000004996,
474
+ 0.06238750009996008,
475
+ 0.3123750000799601,
476
+ 0.18743750007139998,
477
+ 0.43743750008179993,
478
+ -0.00004999993335996766,
479
+ 0.18745000006664003,
480
+ 0.43743750008179993,
481
+ -0.00009999992504000632,
482
+ 0.24993750007139998,
483
+ 0.18745000006664003,
484
+ 0.18740000009996,
485
+ 0.18740000007496,
486
+ 0.12495000006664003,
487
+ 0.3124375000714,
488
+ 0.12493750007139998,
489
+ 0.999956823782745,
490
+ 1.0,
491
+ 1.0,
492
+ 1.0,
493
+ 1.0,
494
+ 0.30305954813957214,
495
+ 1.0,
496
+ 1.0,
497
+ 1.0,
498
+ 1.0,
499
+ 0.999987639360252,
500
+ 0.3030369281768799
501
+ ]
502
+ }
503
+ }
504
+ }
a7-cabinet-close/epoch-010/metadata.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:733d9438896952f7cdc1d31e8fe4d7b78c4351244230bc8ceb4d0620af399399
3
+ size 3379
a7-cabinet-close/epoch-010/model.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:6869f535c390980890549d616376c9a2092717d95513d3cc617a8632f0de628a
3
+ size 7473452088
a7-cabinet-close/epoch-010/protocol.json ADDED
@@ -0,0 +1,56 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "model": "pi0.5 (openpi, PyTorch weights)",
3
+ "task": "A7 close cabinet door (G1)",
4
+ "task_dir": "a7-cabinet-close",
5
+ "openpi_config": "pi05_hb_g1_a7",
6
+ "layout": "g1",
7
+ "asset_id": "g1_a7",
8
+ "epoch_completed": 10,
9
+ "global_step": 100,
10
+ "total_epochs": 20.0,
11
+ "total_steps": 200,
12
+ "global_batch": 3072,
13
+ "per_gpu_batch": 96,
14
+ "world_size": 32,
15
+ "peak_lr": 5e-05,
16
+ "warmup_steps": 10,
17
+ "max_token_len": 256,
18
+ "init_weights": "lerobot/pi05_base converted to PyTorch, action/state projections resized to 76-D ($P/policies-20261109/models/openpi/pi05_base_g1a76_pytorch)",
19
+ "augmentation": "Humanoid Benchmark online augmentation spec v2 (2026-11-09), on from epoch 1; model-native augmentation off",
20
+ "aug_version": "hb-aug-v2-20261109",
21
+ "model_action_dim": 76,
22
+ "action_horizon": 16,
23
+ "episodes": 100,
24
+ "frames": 30664,
25
+ "fps": 30.0,
26
+ "control_hz": 30.0,
27
+ "prompt": "Walk to the open cabinet door under the sink, push it shut with your right hand and pull your hand back.",
28
+ "raw_state_dim": 41,
29
+ "state_layout": "body29(observation.state by name) | inspire_left6 | inspire_right6 | projected_gravity3",
30
+ "action_layout": "action.motion_token64 | inspire_left6 | inspire_right6 (same-row observed hands), rows t..t+15",
31
+ "cameras": {
32
+ "base_0_rgb": "observation.images.ego_view",
33
+ "left_wrist_0_rgb": "observation.images.chest",
34
+ "right_wrist_0_rgb": "zeros, image_mask False"
35
+ },
36
+ "raw_image_hw": [
37
+ 480,
38
+ 640
39
+ ],
40
+ "content_box": [
41
+ 28,
42
+ 196,
43
+ 0,
44
+ 224
45
+ ],
46
+ "training_data": "$P/a3-a7-20261109/a7-cabinet-close-expert-v1/1935a98e8f55d4d1d2d1615ddb97c2679713595c",
47
+ "normalization": "assets/g1_a7/norm_stats.json (openpi quantile normalization)",
48
+ "inference": "Register configs by importing code/hb_pi05/hb_g1_config.py (needs hb_g1_data.py on the path), then openpi policy_config.create_trained_policy(get_config(openpi_config), <this folder>). Clients send observation/<camera> raw RGB, observation/state and prompt; see HBInputs.",
49
+ "resume_state_included": false,
50
+ "files": {
51
+ "assets/g1_a7/norm_stats.json": 13316,
52
+ "metadata.pt": 3379,
53
+ "model.safetensors": 7473452088,
54
+ "training_plan.json": 742
55
+ }
56
+ }
a7-cabinet-close/epoch-010/training_plan.json ADDED
@@ -0,0 +1,23 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "task": "a7",
3
+ "layout": "g1",
4
+ "config": "pi05_hb_g1_a7",
5
+ "exp_name": "run",
6
+ "frames": 30664,
7
+ "world_size": 32,
8
+ "per_gpu_batch": 96,
9
+ "global_batch": 3072,
10
+ "epochs": 20.0,
11
+ "num_train_steps": 200,
12
+ "save_steps": [
13
+ 100,
14
+ 200
15
+ ],
16
+ "warmup_steps": 10,
17
+ "peak_lr": 5e-05,
18
+ "weights": "/mnt/nas/cross/Team/RobotAgent/Personal/baijiaqi/humanoid-benchmark-work/preflight-20261104T1648Z/policies-20261109/models/openpi/pi05_base_g1a76_pytorch",
19
+ "checkpoint_dir": "/mnt/nas/cross/Team/RobotAgent/Personal/baijiaqi/humanoid-benchmark-work/preflight-20261104T1648Z/policies-20261109/pi05-runs/chain-hb1110/a7/run/ckpt/pi05_hb_g1_a7/run",
20
+ "augment": true,
21
+ "aug_version": "hb-aug-v2-20261109",
22
+ "max_token_len": 256
23
+ }
a7-cabinet-close/epoch-020/assets/g1_a7/norm_stats.json ADDED
@@ -0,0 +1,504 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "norm_stats": {
3
+ "state": {
4
+ "mean": [
5
+ -0.35561004281044006,
6
+ 0.037816599011421204,
7
+ 0.09485039860010147,
8
+ 0.5194756388664246,
9
+ -0.1980385035276413,
10
+ -0.006605133879929781,
11
+ -0.3908240795135498,
12
+ -0.1153804287314415,
13
+ -0.2100031077861786,
14
+ 0.5305911898612976,
15
+ -0.18644081056118011,
16
+ -0.0369383879005909,
17
+ 0.015539770945906639,
18
+ 0.013902479782700539,
19
+ 0.016449792310595512,
20
+ 0.20338064432144165,
21
+ 0.1432969570159912,
22
+ 0.03276205062866211,
23
+ -0.21513384580612183,
24
+ 0.004481739364564419,
25
+ 0.007482144050300121,
26
+ -0.02025328017771244,
27
+ -0.20058324933052063,
28
+ -0.07724551111459732,
29
+ 0.036659061908721924,
30
+ 0.21988210082054138,
31
+ 0.014881269074976444,
32
+ 0.009545939043164253,
33
+ 0.0062908814288675785,
34
+ 0.9994741678237915,
35
+ 0.9999215006828308,
36
+ 0.9999319911003113,
37
+ 0.9999361038208008,
38
+ 0.9997094869613647,
39
+ 0.0039056746754795313,
40
+ 0.9999874234199524,
41
+ 0.9999957084655762,
42
+ 0.9999594688415527,
43
+ 0.9998860359191895,
44
+ 0.9993945956230164,
45
+ 0.004138351418077946,
46
+ 0.08264563977718353,
47
+ -0.01804521679878235,
48
+ -0.986268162727356
49
+ ],
50
+ "std": [
51
+ 0.28290310502052307,
52
+ 0.09233158826828003,
53
+ 0.19659872353076935,
54
+ 0.20531529188156128,
55
+ 0.08587288856506348,
56
+ 0.07257651537656784,
57
+ 0.2948111593723297,
58
+ 0.09278968721628189,
59
+ 0.14734666049480438,
60
+ 0.20262238383293152,
61
+ 0.09397441893815994,
62
+ 0.07671834528446198,
63
+ 0.05012736842036247,
64
+ 0.03560950979590416,
65
+ 0.024248890578746796,
66
+ 0.03378785029053688,
67
+ 0.025765331462025642,
68
+ 0.035841915756464005,
69
+ 0.03535226732492447,
70
+ 0.012149378657341003,
71
+ 0.02936926670372486,
72
+ 0.03796697035431862,
73
+ 0.4766698181629181,
74
+ 0.12800851464271545,
75
+ 0.08127840608358383,
76
+ 0.5201791524887085,
77
+ 0.03146743029356003,
78
+ 0.030610734596848488,
79
+ 0.08500799536705017,
80
+ 0.004740353673696518,
81
+ 0.0,
82
+ 0.0,
83
+ 0.0,
84
+ 0.0,
85
+ 0.02886965498328209,
86
+ 0.0,
87
+ 0.0,
88
+ 0.0,
89
+ 0.0,
90
+ 0.0,
91
+ 0.028847994282841682,
92
+ 0.1356819123029709,
93
+ 0.038989145308732986,
94
+ 0.014254551380872726
95
+ ],
96
+ "q01": [
97
+ -0.8435057642134977,
98
+ -0.1539409343454199,
99
+ -0.38289439146854154,
100
+ 0.19744580789334945,
101
+ -0.4023641065573069,
102
+ -0.15783057514232443,
103
+ -0.8902416186952684,
104
+ -0.3213813964588269,
105
+ -0.580422671914016,
106
+ 0.1688319038662181,
107
+ -0.3692423825023103,
108
+ -0.1722233697723961,
109
+ -0.05082127009701552,
110
+ -0.05095853811305899,
111
+ -0.04851158314737884,
112
+ 0.12016225313020878,
113
+ 0.07590308442343385,
114
+ -0.03233522117337442,
115
+ -0.33273061226802,
116
+ -0.023371822398616807,
117
+ -0.07994668760927265,
118
+ -0.10301587557162084,
119
+ -1.2144301321222808,
120
+ -0.3396789082267284,
121
+ -0.08274552846634385,
122
+ -0.40091806847836514,
123
+ -0.0650506398949991,
124
+ -0.11048650279830803,
125
+ -0.21627054291998973,
126
+ 0.9829885311015435,
127
+ 0.9513298273086548,
128
+ 0.9574251770973206,
129
+ 0.9504732489585876,
130
+ 0.851092517375946,
131
+ -1e-10,
132
+ 0.9825374484062195,
133
+ 0.9934524297714233,
134
+ 0.9735382199287415,
135
+ 0.9164777994155884,
136
+ 0.9852659983204475,
137
+ -1e-10,
138
+ -0.13615792221606549,
139
+ -0.12976165536695589,
140
+ -0.9999653876588571
141
+ ],
142
+ "q99": [
143
+ 0.19439826722640507,
144
+ 0.24245334795716866,
145
+ 0.45560911755377265,
146
+ 1.1228674726221726,
147
+ 0.008258910985718626,
148
+ 0.12484455139244438,
149
+ 0.1404618034872802,
150
+ 0.06307184297845925,
151
+ 0.06084788480354464,
152
+ 0.9531215072192949,
153
+ 0.07626591385718673,
154
+ 0.15503785397921882,
155
+ 0.1525609748567458,
156
+ 0.08131745231611116,
157
+ 0.08366167413077756,
158
+ 0.3211568147841109,
159
+ 0.20607530826763967,
160
+ 0.15977249451081388,
161
+ -0.1695193791277912,
162
+ 0.04427388410215958,
163
+ 0.06231649593288226,
164
+ 0.07538968521664166,
165
+ 0.39386444146296595,
166
+ 0.16625795727069403,
167
+ 0.2778081297818395,
168
+ 1.3490795560201072,
169
+ 0.095980245563843,
170
+ 0.07410926639083143,
171
+ 0.13639384343573485,
172
+ 0.999956823782745,
173
+ 1.0,
174
+ 1.0,
175
+ 1.0,
176
+ 1.0,
177
+ 0.16704642294477218,
178
+ 1.0,
179
+ 1.0,
180
+ 1.0,
181
+ 1.0,
182
+ 0.999987639360252,
183
+ 0.16709456220701158,
184
+ 0.3249750240081849,
185
+ 0.06491606969695807,
186
+ -0.9452246287075486
187
+ ]
188
+ },
189
+ "actions": {
190
+ "mean": [
191
+ 0.10757830739021301,
192
+ 0.0010036963503807783,
193
+ -0.1955849826335907,
194
+ 0.03575943410396576,
195
+ -0.02901560068130493,
196
+ 0.10509791970252991,
197
+ 0.15584516525268555,
198
+ -0.036249756813049316,
199
+ 0.00029133816133253276,
200
+ -0.19356472790241241,
201
+ -0.06754574179649353,
202
+ -0.07026422023773193,
203
+ -0.10356352478265762,
204
+ 0.1086534708738327,
205
+ 0.179977685213089,
206
+ -0.09078477323055267,
207
+ 0.04756608605384827,
208
+ -0.006199375726282597,
209
+ -0.1645488739013672,
210
+ 0.03273191303014755,
211
+ 0.013353785499930382,
212
+ -0.0864969938993454,
213
+ 0.033461980521678925,
214
+ 0.149399533867836,
215
+ 0.09672172367572784,
216
+ 0.03816988691687584,
217
+ -0.13994601368904114,
218
+ 0.0559534914791584,
219
+ 0.1010548546910286,
220
+ -0.025988206267356873,
221
+ 0.12739083170890808,
222
+ -0.08735954016447067,
223
+ -0.002861789194867015,
224
+ 0.1596086025238037,
225
+ -0.018322203308343887,
226
+ 0.007539633195847273,
227
+ 0.04284964129328728,
228
+ -0.09214604645967484,
229
+ 0.14863164722919464,
230
+ 0.016650989651679993,
231
+ 0.027098143473267555,
232
+ -0.08947993069887161,
233
+ 0.05657769367098808,
234
+ -0.11662724614143372,
235
+ 0.0791701003909111,
236
+ -0.10034364461898804,
237
+ -0.09565547853708267,
238
+ -0.2216619849205017,
239
+ 0.008683966472744942,
240
+ -0.15533827245235443,
241
+ 0.0966157391667366,
242
+ -0.003970836289227009,
243
+ 0.08637572079896927,
244
+ -0.118191197514534,
245
+ 0.1010720506310463,
246
+ 0.14126704633235931,
247
+ -0.13963939249515533,
248
+ 0.07252243906259537,
249
+ 0.0847059041261673,
250
+ -0.04034632444381714,
251
+ -0.047200094908475876,
252
+ -0.002455164212733507,
253
+ 0.15698553621768951,
254
+ -0.03268032148480415,
255
+ 0.9995062351226807,
256
+ 0.9999467134475708,
257
+ 0.9999535083770752,
258
+ 0.9999532103538513,
259
+ 0.9997615218162537,
260
+ 0.0007274940144270658,
261
+ 0.9999932050704956,
262
+ 0.999999463558197,
263
+ 0.9999749064445496,
264
+ 0.9999121427536011,
265
+ 0.9996171593666077,
266
+ 0.0009604249498806894
267
+ ],
268
+ "std": [
269
+ 0.06684934347867966,
270
+ 0.12237707525491714,
271
+ 0.12836524844169617,
272
+ 0.1443842351436615,
273
+ 0.12026809900999069,
274
+ 0.06686931103467941,
275
+ 0.08585774153470993,
276
+ 0.12531286478042603,
277
+ 0.07042475789785385,
278
+ 0.04380151629447937,
279
+ 0.0754399374127388,
280
+ 0.07911296933889389,
281
+ 0.10523597151041031,
282
+ 0.04984365031123161,
283
+ 0.10585196316242218,
284
+ 0.1417231559753418,
285
+ 0.08640342950820923,
286
+ 0.09490349888801575,
287
+ 0.10842842608690262,
288
+ 0.12022152543067932,
289
+ 0.11372580379247665,
290
+ 0.08418825268745422,
291
+ 0.09572859108448029,
292
+ 0.11790358275175095,
293
+ 0.04139557480812073,
294
+ 0.11743805557489395,
295
+ 0.04096134379506111,
296
+ 0.1019567996263504,
297
+ 0.08468007296323776,
298
+ 0.09862369298934937,
299
+ 0.08369130641222,
300
+ 0.07918284833431244,
301
+ 0.12212815880775452,
302
+ 0.07682932168245316,
303
+ 0.06639379262924194,
304
+ 0.06500166654586792,
305
+ 0.1091652363538742,
306
+ 0.1271735578775406,
307
+ 0.07222087681293488,
308
+ 0.11637697368860245,
309
+ 0.09275760501623154,
310
+ 0.06542796641588211,
311
+ 0.064761683344841,
312
+ 0.04859273508191109,
313
+ 0.10564269870519638,
314
+ 0.05467769876122475,
315
+ 0.15121139585971832,
316
+ 0.13371562957763672,
317
+ 0.0961233377456665,
318
+ 0.11129939556121826,
319
+ 0.09783373773097992,
320
+ 0.12376602739095688,
321
+ 0.17745254933834076,
322
+ 0.050024744123220444,
323
+ 0.07118535041809082,
324
+ 0.15862050652503967,
325
+ 0.10293488204479218,
326
+ 0.11946934461593628,
327
+ 0.06005885452032089,
328
+ 0.11131948232650757,
329
+ 0.10331486165523529,
330
+ 0.06897100061178207,
331
+ 0.07636360079050064,
332
+ 0.07862718403339386,
333
+ 0.0,
334
+ 0.0,
335
+ 0.0,
336
+ 0.0,
337
+ 0.0,
338
+ 0.011146476492285728,
339
+ 0.0,
340
+ 0.0,
341
+ 0.0,
342
+ 0.0,
343
+ 0.0,
344
+ 0.011168749071657658
345
+ ],
346
+ "q01": [
347
+ -5.999999108663445e-11,
348
+ -0.2500000001,
349
+ -0.43750000006,
350
+ -0.25001250006364,
351
+ -0.2500625000778,
352
+ -0.0625000001,
353
+ -0.0625000001,
354
+ -0.250000000075,
355
+ -0.1250000001,
356
+ -0.25000000005,
357
+ -0.25000000005,
358
+ -0.25002500007144,
359
+ -0.312500000075,
360
+ -0.06252500006668,
361
+ -0.00006250007779998873,
362
+ -0.37507500008184,
363
+ -0.12502500006668,
364
+ -0.18751250005556,
365
+ -0.3750625000778,
366
+ -0.187500000075,
367
+ -0.2500625000778,
368
+ -0.31251250005555997,
369
+ -0.125000000075,
370
+ -0.12500000008,
371
+ -1e-10,
372
+ -0.1875000001,
373
+ -0.25000000006,
374
+ -0.125000000075,
375
+ -0.1250625000778,
376
+ -0.18752500007144002,
377
+ -0.062500000075,
378
+ -0.25002500007144,
379
+ -0.2500625000778,
380
+ -0.000025000071439996996,
381
+ -0.12500000006,
382
+ -0.18752500007144002,
383
+ -0.12500000008,
384
+ -0.3750000001,
385
+ -1e-10,
386
+ -0.1875000001,
387
+ -0.187500000075,
388
+ -0.25002500007144,
389
+ -0.06252500006668,
390
+ -0.18752500006668,
391
+ -0.062500000075,
392
+ -0.25002500006668,
393
+ -0.37500000008,
394
+ -0.50000000008,
395
+ -0.1875000001,
396
+ -0.4375625000778,
397
+ -0.18750000008,
398
+ -0.1875000001,
399
+ -0.1875000001,
400
+ -0.25002500006668,
401
+ -0.06252500006668,
402
+ -0.12507500008184003,
403
+ -0.375000000075,
404
+ -0.1250000001,
405
+ -0.06252500006668,
406
+ -0.250000000075,
407
+ -0.187500000075,
408
+ -0.12502500006668,
409
+ -0.0625000001,
410
+ -0.18752500007144002,
411
+ 0.9829885311015435,
412
+ 0.9513298273086548,
413
+ 0.9574251770973206,
414
+ 0.9504732489585876,
415
+ 0.851092517375946,
416
+ 0.0,
417
+ 0.9825374484062195,
418
+ 0.9934524297714233,
419
+ 0.9735382199287415,
420
+ 0.9164777994155884,
421
+ 0.9852659983204475,
422
+ 0.0
423
+ ],
424
+ "q99": [
425
+ 0.24993750009996002,
426
+ 0.24990000009996,
427
+ -0.00012499992003989213,
428
+ 0.3123625000999599,
429
+ 0.18745000007776003,
430
+ 0.24995000006664003,
431
+ 0.31242500009996005,
432
+ 0.18740000009996,
433
+ 0.12493750005995999,
434
+ -0.12504999995004001,
435
+ 0.062400000074959994,
436
+ 0.062437500071399976,
437
+ 0.062400000074959994,
438
+ 0.18745000006664003,
439
+ 0.43745000007776,
440
+ 0.18743750008179993,
441
+ 0.18742500009996005,
442
+ 0.18745000007776003,
443
+ 0.062450000077760026,
444
+ 0.24990000009996,
445
+ 0.18745000007776003,
446
+ 0.062450000077760026,
447
+ 0.24990000007496,
448
+ 0.3748750000799601,
449
+ 0.18745000004995999,
450
+ 0.24991250009995997,
451
+ -0.06256249994004001,
452
+ 0.24990000007496,
453
+ 0.24990000005552004,
454
+ 0.18741250009995997,
455
+ 0.31240000007496,
456
+ 0.062437500071399976,
457
+ 0.24988750009996008,
458
+ 0.3124375000714,
459
+ 0.12493750009996002,
460
+ 0.12493750007139998,
461
+ 0.3748750000799601,
462
+ 0.12495000007776003,
463
+ 0.31246250004284,
464
+ 0.24991250009995997,
465
+ 0.18740000007496,
466
+ 0.062437500071399976,
467
+ 0.18745000006664003,
468
+ -0.00002499996667998383,
469
+ 0.31240000007496,
470
+ -0.00004999993335996766,
471
+ 0.12487500007996011,
472
+ -0.00012499992003989213,
473
+ 0.18740000004996,
474
+ 0.06238750009996008,
475
+ 0.3123750000799601,
476
+ 0.18743750007139998,
477
+ 0.43743750008179993,
478
+ -0.00004999993335996766,
479
+ 0.18745000006664003,
480
+ 0.43743750008179993,
481
+ -0.00009999992504000632,
482
+ 0.24993750007139998,
483
+ 0.18745000006664003,
484
+ 0.18740000009996,
485
+ 0.18740000007496,
486
+ 0.12495000006664003,
487
+ 0.3124375000714,
488
+ 0.12493750007139998,
489
+ 0.999956823782745,
490
+ 1.0,
491
+ 1.0,
492
+ 1.0,
493
+ 1.0,
494
+ 0.30305954813957214,
495
+ 1.0,
496
+ 1.0,
497
+ 1.0,
498
+ 1.0,
499
+ 0.999987639360252,
500
+ 0.3030369281768799
501
+ ]
502
+ }
503
+ }
504
+ }
a7-cabinet-close/epoch-020/metadata.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:f71e7eea6fcde02a58ee7e8fb0481660b2112c9e30dfd7ffe0c554620e7f23bc
3
+ size 3379
a7-cabinet-close/epoch-020/model.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:c187ff07adc7f82a3212d7cd402217efa0f759df7fe29a6988dc2188e722066a
3
+ size 7473452088
a7-cabinet-close/epoch-020/protocol.json ADDED
@@ -0,0 +1,56 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "model": "pi0.5 (openpi, PyTorch weights)",
3
+ "task": "A7 close cabinet door (G1)",
4
+ "task_dir": "a7-cabinet-close",
5
+ "openpi_config": "pi05_hb_g1_a7",
6
+ "layout": "g1",
7
+ "asset_id": "g1_a7",
8
+ "epoch_completed": 20,
9
+ "global_step": 200,
10
+ "total_epochs": 20.0,
11
+ "total_steps": 200,
12
+ "global_batch": 3072,
13
+ "per_gpu_batch": 96,
14
+ "world_size": 32,
15
+ "peak_lr": 5e-05,
16
+ "warmup_steps": 10,
17
+ "max_token_len": 256,
18
+ "init_weights": "lerobot/pi05_base converted to PyTorch, action/state projections resized to 76-D ($P/policies-20261109/models/openpi/pi05_base_g1a76_pytorch)",
19
+ "augmentation": "Humanoid Benchmark online augmentation spec v2 (2026-11-09), on from epoch 1; model-native augmentation off",
20
+ "aug_version": "hb-aug-v2-20261109",
21
+ "model_action_dim": 76,
22
+ "action_horizon": 16,
23
+ "episodes": 100,
24
+ "frames": 30664,
25
+ "fps": 30.0,
26
+ "control_hz": 30.0,
27
+ "prompt": "Walk to the open cabinet door under the sink, push it shut with your right hand and pull your hand back.",
28
+ "raw_state_dim": 41,
29
+ "state_layout": "body29(observation.state by name) | inspire_left6 | inspire_right6 | projected_gravity3",
30
+ "action_layout": "action.motion_token64 | inspire_left6 | inspire_right6 (same-row observed hands), rows t..t+15",
31
+ "cameras": {
32
+ "base_0_rgb": "observation.images.ego_view",
33
+ "left_wrist_0_rgb": "observation.images.chest",
34
+ "right_wrist_0_rgb": "zeros, image_mask False"
35
+ },
36
+ "raw_image_hw": [
37
+ 480,
38
+ 640
39
+ ],
40
+ "content_box": [
41
+ 28,
42
+ 196,
43
+ 0,
44
+ 224
45
+ ],
46
+ "training_data": "$P/a3-a7-20261109/a7-cabinet-close-expert-v1/1935a98e8f55d4d1d2d1615ddb97c2679713595c",
47
+ "normalization": "assets/g1_a7/norm_stats.json (openpi quantile normalization)",
48
+ "inference": "Register configs by importing code/hb_pi05/hb_g1_config.py (needs hb_g1_data.py on the path), then openpi policy_config.create_trained_policy(get_config(openpi_config), <this folder>). Clients send observation/<camera> raw RGB, observation/state and prompt; see HBInputs.",
49
+ "resume_state_included": false,
50
+ "files": {
51
+ "assets/g1_a7/norm_stats.json": 13316,
52
+ "metadata.pt": 3379,
53
+ "model.safetensors": 7473452088,
54
+ "training_plan.json": 742
55
+ }
56
+ }
a7-cabinet-close/epoch-020/training_plan.json ADDED
@@ -0,0 +1,23 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "task": "a7",
3
+ "layout": "g1",
4
+ "config": "pi05_hb_g1_a7",
5
+ "exp_name": "run",
6
+ "frames": 30664,
7
+ "world_size": 32,
8
+ "per_gpu_batch": 96,
9
+ "global_batch": 3072,
10
+ "epochs": 20.0,
11
+ "num_train_steps": 200,
12
+ "save_steps": [
13
+ 100,
14
+ 200
15
+ ],
16
+ "warmup_steps": 10,
17
+ "peak_lr": 5e-05,
18
+ "weights": "/mnt/nas/cross/Team/RobotAgent/Personal/baijiaqi/humanoid-benchmark-work/preflight-20261104T1648Z/policies-20261109/models/openpi/pi05_base_g1a76_pytorch",
19
+ "checkpoint_dir": "/mnt/nas/cross/Team/RobotAgent/Personal/baijiaqi/humanoid-benchmark-work/preflight-20261104T1648Z/policies-20261109/pi05-runs/chain-hb1110/a7/run/ckpt/pi05_hb_g1_a7/run",
20
+ "augment": true,
21
+ "aug_version": "hb-aug-v2-20261109",
22
+ "max_token_len": 256
23
+ }
b6-pedal-bin/epoch-010/assets/g1_b6/norm_stats.json ADDED
@@ -0,0 +1,504 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "norm_stats": {
3
+ "state": {
4
+ "mean": [
5
+ -0.09289415925741196,
6
+ 0.045334845781326294,
7
+ 0.048640452325344086,
8
+ 0.3030760884284973,
9
+ -0.11417802423238754,
10
+ -0.03437678515911102,
11
+ -0.06910363584756851,
12
+ -0.06624675542116165,
13
+ -0.28657519817352295,
14
+ 0.3067435622215271,
15
+ -0.15567664802074432,
16
+ -0.023483101278543472,
17
+ 0.014021046459674835,
18
+ 0.014938326552510262,
19
+ 0.14744997024536133,
20
+ 0.24639774858951569,
21
+ 0.19172275066375732,
22
+ -0.04901121184229851,
23
+ -0.3446519374847412,
24
+ -0.366971492767334,
25
+ -0.03786687180399895,
26
+ 0.17884966731071472,
27
+ 0.07702597975730896,
28
+ -0.3111043870449066,
29
+ 0.006915700156241655,
30
+ -0.10259399563074112,
31
+ 0.3281053602695465,
32
+ 0.05004274472594261,
33
+ -0.041460540145635605,
34
+ 0.9999999403953552,
35
+ 1.0,
36
+ 1.0,
37
+ 1.0,
38
+ 0.9999556541442871,
39
+ 4.9299426763127485e-8,
40
+ 0.8562398552894592,
41
+ 0.8557164669036865,
42
+ 0.862363874912262,
43
+ 0.870231568813324,
44
+ 0.9374372959136963,
45
+ 0.0013326209736987948,
46
+ -0.0698074996471405,
47
+ 0.004781579133123159,
48
+ -0.9963920712471008
49
+ ],
50
+ "std": [
51
+ 0.22050054371356964,
52
+ 0.07048683613538742,
53
+ 0.134409561753273,
54
+ 0.2732667624950409,
55
+ 0.13167059421539307,
56
+ 0.052785616368055344,
57
+ 0.1399744749069214,
58
+ 0.05429463088512421,
59
+ 0.1418997198343277,
60
+ 0.1645597666501999,
61
+ 0.11816693097352982,
62
+ 0.04961647093296051,
63
+ 0.04337943345308304,
64
+ 0.019502397626638412,
65
+ 0.02250182442367077,
66
+ 0.06032394990324974,
67
+ 0.07465467602014542,
68
+ 0.10689205676317215,
69
+ 0.048137277364730835,
70
+ 0.04392699897289276,
71
+ 0.06392653286457062,
72
+ 0.06639137864112854,
73
+ 0.30172207951545715,
74
+ 0.23119832575321198,
75
+ 0.20930008590221405,
76
+ 0.5789200663566589,
77
+ 0.2542855143547058,
78
+ 0.23488010466098785,
79
+ 0.218588188290596,
80
+ 0.0,
81
+ 0.0,
82
+ 0.0,
83
+ 0.0,
84
+ 0.0,
85
+ 6.862556801934261e-6,
86
+ 0.16237150132656097,
87
+ 0.16346286237239838,
88
+ 0.1562580168247223,
89
+ 0.14771395921707153,
90
+ 0.06923356652259827,
91
+ 0.001731409691274166,
92
+ 0.04638409987092018,
93
+ 0.023614883422851562,
94
+ 0.0
95
+ ],
96
+ "q01": [
97
+ -0.880411240274788,
98
+ -0.11660030736746661,
99
+ -0.2591128731101498,
100
+ 0.1310143638281691,
101
+ -0.3975315273160438,
102
+ -0.12719849527707686,
103
+ -0.43698455556718624,
104
+ -0.2454130813051354,
105
+ -0.5891453627907941,
106
+ 0.15688254223430617,
107
+ -0.45365863902816306,
108
+ -0.1532304523474964,
109
+ -0.05649498579946606,
110
+ -0.03512588135941741,
111
+ 0.10754643191276736,
112
+ 0.09098392945162612,
113
+ 0.06845034108700554,
114
+ -0.2926570351384775,
115
+ -0.4391649112390142,
116
+ -0.4684700581492143,
117
+ -0.16496463962956748,
118
+ 0.07493410048820395,
119
+ -0.676092325818013,
120
+ -1.0350512746883667,
121
+ -0.43236514993125463,
122
+ -0.8278534394291747,
123
+ -0.051698092269253815,
124
+ -0.5183220580090175,
125
+ -0.4080306218162441,
126
+ 0.999207079410553,
127
+ 0.9995272755622864,
128
+ 0.9994345307350159,
129
+ 0.9995204210281372,
130
+ 0.9992793597638138,
131
+ 0.0,
132
+ 0.6215853575778556,
133
+ 0.6115407229258378,
134
+ 0.6271049689326582,
135
+ 0.6401020124295089,
136
+ 0.8316517208587076,
137
+ -1e-10,
138
+ -0.18501513771914246,
139
+ -0.05655392329480376,
140
+ -0.999953441719553
141
+ ],
142
+ "q99": [
143
+ 0.1713271221507653,
144
+ 0.16606888934973485,
145
+ 0.255133149306362,
146
+ 1.4343218416003525,
147
+ 0.18631403401933788,
148
+ 0.10816316061141407,
149
+ 0.24138335087282747,
150
+ 0.03144279724399046,
151
+ -0.027339548691011184,
152
+ 0.9277804614462278,
153
+ 0.06613453889172527,
154
+ 0.09583480280191259,
155
+ 0.14404179024680713,
156
+ 0.05739128976607806,
157
+ 0.235095005062959,
158
+ 0.34817335710089825,
159
+ 0.3896389798800358,
160
+ 0.10347168962327696,
161
+ -0.20304301404597563,
162
+ -0.2800530286244159,
163
+ 0.07539925654783991,
164
+ 0.3544070315619453,
165
+ 0.5230730046380223,
166
+ 0.06049067594124291,
167
+ 0.520795445975837,
168
+ 1.1251266785638767,
169
+ 1.12142253197707,
170
+ 0.5207122732182838,
171
+ 0.5347526467551684,
172
+ 1.0,
173
+ 1.0,
174
+ 1.0,
175
+ 1.0,
176
+ 0.9999972909258038,
177
+ 0.0010093163000419736,
178
+ 0.9999195548103482,
179
+ 0.9999173667715222,
180
+ 0.9999223136351736,
181
+ 0.9999271019862325,
182
+ 0.9999653391407162,
183
+ 0.005770046510408187,
184
+ 0.0100407091547432,
185
+ 0.06605743221344167,
186
+ -0.9818596055855306
187
+ ]
188
+ },
189
+ "actions": {
190
+ "mean": [
191
+ 0.03432026505470276,
192
+ 0.012461589649319649,
193
+ -0.05914423614740372,
194
+ -0.0072383214719593525,
195
+ 0.05019465461373329,
196
+ 0.11229633539915085,
197
+ 0.06718273460865021,
198
+ -0.004092664457857609,
199
+ 0.0245615653693676,
200
+ -0.2105632871389389,
201
+ -0.14285634458065033,
202
+ -0.199773371219635,
203
+ -0.041197869926691055,
204
+ 0.06098661199212074,
205
+ 0.1831471472978592,
206
+ -0.06344112753868103,
207
+ 0.05129233002662659,
208
+ -0.062251925468444824,
209
+ -0.009263630956411362,
210
+ -0.0808362290263176,
211
+ -0.07540543377399445,
212
+ -0.15363413095474243,
213
+ -0.11247213184833527,
214
+ -0.10093164443969727,
215
+ 0.01776888780295849,
216
+ -0.03236458823084831,
217
+ -0.17563124001026154,
218
+ -0.03566235303878784,
219
+ 0.03622003644704819,
220
+ -0.005657575558871031,
221
+ -0.002754294080659747,
222
+ -0.172866091132164,
223
+ 0.04396877810359001,
224
+ 0.12563470005989075,
225
+ 0.027359044179320335,
226
+ 0.08746480196714401,
227
+ -0.07898023724555969,
228
+ -0.05846341699361801,
229
+ 0.12352164089679718,
230
+ -0.07225605845451355,
231
+ -0.07172008603811264,
232
+ -0.0689985379576683,
233
+ 0.09783204644918442,
234
+ -0.1205044537782669,
235
+ 0.04385637864470482,
236
+ -0.0408664271235466,
237
+ -0.0202995203435421,
238
+ -0.05485040321946144,
239
+ -0.2798519432544708,
240
+ -0.06503159552812576,
241
+ 0.019104620441794395,
242
+ 0.07799244672060013,
243
+ 0.024153348058462143,
244
+ -0.07540969550609589,
245
+ -0.002842223271727562,
246
+ 0.09865490347146988,
247
+ -0.052052684128284454,
248
+ 0.07384795695543289,
249
+ 0.13516001403331757,
250
+ -0.16005200147628784,
251
+ -0.05993583798408508,
252
+ 0.06282256543636322,
253
+ 0.17966081202030182,
254
+ -0.06581006199121475,
255
+ 1.0,
256
+ 1.0,
257
+ 1.0,
258
+ 1.0,
259
+ 0.9999973177909851,
260
+ 4.929941965770013e-8,
261
+ 0.8565258383750916,
262
+ 0.8557460904121399,
263
+ 0.8630558848381042,
264
+ 0.8705127835273743,
265
+ 0.9372830390930176,
266
+ 0.0013317195698618889
267
+ ],
268
+ "std": [
269
+ 0.10664359480142593,
270
+ 0.1436038762331009,
271
+ 0.12584248185157776,
272
+ 0.10123310983181,
273
+ 0.18608538806438446,
274
+ 0.12901633977890015,
275
+ 0.16723662614822388,
276
+ 0.08674019575119019,
277
+ 0.09300430119037628,
278
+ 0.07599232345819473,
279
+ 0.13231238722801208,
280
+ 0.10487758368253708,
281
+ 0.12685155868530273,
282
+ 0.13161525130271912,
283
+ 0.10486916452646255,
284
+ 0.1401214748620987,
285
+ 0.08177037537097931,
286
+ 0.14273640513420105,
287
+ 0.15247996151447296,
288
+ 0.09722961485385895,
289
+ 0.09942055493593216,
290
+ 0.12243745476007462,
291
+ 0.10003399103879929,
292
+ 0.10224147886037827,
293
+ 0.08081497251987457,
294
+ 0.11588867008686066,
295
+ 0.0635068491101265,
296
+ 0.12075082212686539,
297
+ 0.10402055084705353,
298
+ 0.12647128105163574,
299
+ 0.13117629289627075,
300
+ 0.12363912910223007,
301
+ 0.08622833341360092,
302
+ 0.09993847459554672,
303
+ 0.051116377115249634,
304
+ 0.09662438929080963,
305
+ 0.12317366898059845,
306
+ 0.15041758120059967,
307
+ 0.08345945179462433,
308
+ 0.10266817361116409,
309
+ 0.11730080097913742,
310
+ 0.13897928595542908,
311
+ 0.13960976898670197,
312
+ 0.07891123741865158,
313
+ 0.107028529047966,
314
+ 0.13991400599479675,
315
+ 0.113197460770607,
316
+ 0.15233290195465088,
317
+ 0.10571637749671936,
318
+ 0.13236956298351288,
319
+ 0.11429065465927124,
320
+ 0.08670087158679962,
321
+ 0.19508449733257294,
322
+ 0.13858312368392944,
323
+ 0.06682883948087692,
324
+ 0.16072209179401398,
325
+ 0.14802610874176025,
326
+ 0.05860956013202667,
327
+ 0.12714003026485443,
328
+ 0.0833347961306572,
329
+ 0.1227123811841011,
330
+ 0.1028541624546051,
331
+ 0.10684986412525177,
332
+ 0.13590067625045776,
333
+ 0.0,
334
+ 0.0,
335
+ 0.0,
336
+ 0.0,
337
+ 0.0,
338
+ 6.8625545281975064e-6,
339
+ 0.1610528975725174,
340
+ 0.16486616432666779,
341
+ 0.1496121883392334,
342
+ 0.15113823115825653,
343
+ 0.06507572531700134,
344
+ 0.0017317550955340266
345
+ ],
346
+ "q01": [
347
+ -0.31260000004288,
348
+ -0.31260000008336003,
349
+ -0.31265000003336,
350
+ -0.25000000005,
351
+ -0.37502500008824,
352
+ -0.18753750006924,
353
+ -0.25000000005,
354
+ -0.18751250005556,
355
+ -0.25010000003335997,
356
+ -0.4375625000778,
357
+ -0.37513750005388,
358
+ -0.43750000008,
359
+ -0.25005000006668,
360
+ -0.18757500008184003,
361
+ -0.06263750005388,
362
+ -0.43756250008668,
363
+ -0.18750000003999998,
364
+ -0.31250000008,
365
+ -0.37502500008572,
366
+ -0.31250000005,
367
+ -0.25001250006364,
368
+ -0.437500000075,
369
+ -0.31257500008184,
370
+ -0.37510000008464,
371
+ -0.18751250005556,
372
+ -0.25000000008,
373
+ -0.31252500006668,
374
+ -0.31251250006364,
375
+ -0.2500625000778,
376
+ -0.25001250006364,
377
+ -0.31260000008336003,
378
+ -0.43755000007144,
379
+ -0.25003750006923997,
380
+ -0.12501250006364,
381
+ -0.12502500006668,
382
+ -0.18750000008,
383
+ -0.31255000007144,
384
+ -0.43756250008668,
385
+ -0.06256250007780001,
386
+ -0.25001250006364,
387
+ -0.25003750006923997,
388
+ -0.31253750006923997,
389
+ -0.18756250004668,
390
+ -0.25002500007144,
391
+ -0.25007500008184,
392
+ -0.25010000008336003,
393
+ -0.25007500008184,
394
+ -0.37505000007648004,
395
+ -0.50007500008184,
396
+ -0.375000000075,
397
+ -0.37515000003336,
398
+ -0.25000000008,
399
+ -0.50002500008824,
400
+ -0.375000000075,
401
+ -0.125000000075,
402
+ -0.37505000007648004,
403
+ -0.31260000008336003,
404
+ -0.06252500007144002,
405
+ -0.12500000005,
406
+ -0.3750000001,
407
+ -0.25013750005388,
408
+ -0.18757500003848,
409
+ -0.06258750004548003,
410
+ -0.43762500007336,
411
+ 0.999207079410553,
412
+ 0.9995272755622864,
413
+ 0.9994345307350159,
414
+ 0.9995204210281372,
415
+ 0.9992983239831871,
416
+ 0.0,
417
+ 0.6215853575778556,
418
+ 0.6115407229258378,
419
+ 0.6271049689326582,
420
+ 0.6401020124295089,
421
+ 0.8316517208587076,
422
+ -1e-10
423
+ ],
424
+ "q99": [
425
+ 0.24985000008568004,
426
+ 0.31245000008332,
427
+ 0.24985000006664004,
428
+ 0.24995000008332002,
429
+ 0.5623125000882,
430
+ 0.37487500006920005,
431
+ 0.4374000000874799,
432
+ 0.18745000007776003,
433
+ 0.18745000008332002,
434
+ -0.06259999994447996,
435
+ 0.18743750008460003,
436
+ 0.06237500007996011,
437
+ 0.24990000006664004,
438
+ 0.37493750008179993,
439
+ 0.43737500006920005,
440
+ 0.24981250005996003,
441
+ 0.1873750000799601,
442
+ 0.1873750000799601,
443
+ 0.37485000008568004,
444
+ 0.18745000008332002,
445
+ 0.18737500006359997,
446
+ 0.062300000024959945,
447
+ 0.18737500006359997,
448
+ 0.18747500005384,
449
+ 0.18745000007776003,
450
+ 0.2498750000799601,
451
+ -0.06254999993335997,
452
+ 0.18743750008179993,
453
+ 0.18745000007776003,
454
+ 0.24993750008179993,
455
+ 0.24990000006664004,
456
+ 0.12490000005711999,
457
+ 0.24997500005384,
458
+ 0.37493750008179993,
459
+ 0.12495000006664003,
460
+ 0.3123750000799601,
461
+ 0.31237500007140007,
462
+ 0.24981250005996003,
463
+ 0.31240000005552004,
464
+ 0.18737500006359997,
465
+ 0.31237500006920005,
466
+ 0.31243750008460003,
467
+ 0.37493750007332005,
468
+ 0.062437500071399976,
469
+ 0.31243750008179993,
470
+ 0.37495000008332,
471
+ 0.24987500006359997,
472
+ 0.31238750005292004,
473
+ -0.0001249999364000276,
474
+ 0.18740000003748003,
475
+ 0.24990000007776003,
476
+ 0.2498750000799601,
477
+ 0.4373125000882,
478
+ 0.24980000004995995,
479
+ 0.18740000004996,
480
+ 0.43733750007644,
481
+ 0.37485000009995995,
482
+ 0.24993750007139998,
483
+ 0.37495000008332,
484
+ -0.00009999995004000839,
485
+ 0.24987500006920005,
486
+ 0.24987500006920005,
487
+ 0.37493750008179993,
488
+ 0.24993750007332005,
489
+ 1.0,
490
+ 1.0,
491
+ 1.0,
492
+ 1.0,
493
+ 0.9999972909258038,
494
+ 0.0010093163000419736,
495
+ 0.9999195548103482,
496
+ 0.9999173667715222,
497
+ 0.9999223136351736,
498
+ 0.9999271019862325,
499
+ 0.9999653391407162,
500
+ 0.005770046510408187
501
+ ]
502
+ }
503
+ }
504
+ }
b6-pedal-bin/epoch-010/metadata.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:15092b21e2dec45d328533320061c2c9e086b38ccf7ab983fda63ea5e3a05f79
3
+ size 3379
b6-pedal-bin/epoch-010/model.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:b9fe1074498184cce34782df73272a6e5d662623f136e133b30e3bc200260ba7
3
+ size 7473452088
b6-pedal-bin/epoch-010/protocol.json ADDED
@@ -0,0 +1,56 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "model": "pi0.5 (openpi, PyTorch weights)",
3
+ "task": "B6 pedal-bin trash disposal (G1)",
4
+ "task_dir": "b6-pedal-bin",
5
+ "openpi_config": "pi05_hb_g1_b6",
6
+ "layout": "g1",
7
+ "asset_id": "g1_b6",
8
+ "epoch_completed": 10,
9
+ "global_step": 213,
10
+ "total_epochs": 20.0,
11
+ "total_steps": 425,
12
+ "global_batch": 3072,
13
+ "per_gpu_batch": 96,
14
+ "world_size": 32,
15
+ "peak_lr": 5e-05,
16
+ "warmup_steps": 21,
17
+ "max_token_len": 256,
18
+ "init_weights": "lerobot/pi05_base converted to PyTorch, action/state projections resized to 76-D ($P/policies-20261109/models/openpi/pi05_base_g1a76_pytorch)",
19
+ "augmentation": "Humanoid Benchmark online augmentation spec v2 (2026-11-09), on from epoch 1; model-native augmentation off",
20
+ "aug_version": "hb-aug-v2-20261109",
21
+ "model_action_dim": 76,
22
+ "action_horizon": 16,
23
+ "episodes": 100,
24
+ "frames": 65127,
25
+ "fps": 30.0,
26
+ "control_hz": 50.0,
27
+ "prompt": "Pick up the chip barrel from the cart, walk to the pedal bin, step on its pedal and drop it in.",
28
+ "raw_state_dim": 41,
29
+ "state_layout": "body29(observation.state by name) | inspire_left6 | inspire_right6 | projected_gravity3",
30
+ "action_layout": "action.motion_token64 | inspire_left6 | inspire_right6 (same-row observed hands), rows t..t+15",
31
+ "cameras": {
32
+ "base_0_rgb": "observation.images.ego_view",
33
+ "left_wrist_0_rgb": "observation.images.chest",
34
+ "right_wrist_0_rgb": "zeros, image_mask False"
35
+ },
36
+ "raw_image_hw": [
37
+ 480,
38
+ 640
39
+ ],
40
+ "content_box": [
41
+ 28,
42
+ 196,
43
+ 0,
44
+ 224
45
+ ],
46
+ "training_data": "$P/b6-v5t7-ready-start-60ep-20261109/data/deliveries/g1/b6-v5t7-ready-start-20261109",
47
+ "normalization": "assets/g1_b6/norm_stats.json (openpi quantile normalization)",
48
+ "inference": "Register configs by importing code/hb_pi05/hb_g1_config.py (needs hb_g1_data.py on the path), then openpi policy_config.create_trained_policy(get_config(openpi_config), <this folder>). Clients send observation/<camera> raw RGB, observation/state and prompt; see HBInputs.",
49
+ "resume_state_included": false,
50
+ "files": {
51
+ "assets/g1_b6/norm_stats.json": 13548,
52
+ "metadata.pt": 3379,
53
+ "model.safetensors": 7473452088,
54
+ "training_plan.json": 742
55
+ }
56
+ }
b6-pedal-bin/epoch-010/training_plan.json ADDED
@@ -0,0 +1,23 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "task": "b6",
3
+ "layout": "g1",
4
+ "config": "pi05_hb_g1_b6",
5
+ "exp_name": "run",
6
+ "frames": 65127,
7
+ "world_size": 32,
8
+ "per_gpu_batch": 96,
9
+ "global_batch": 3072,
10
+ "epochs": 20.0,
11
+ "num_train_steps": 425,
12
+ "save_steps": [
13
+ 213,
14
+ 425
15
+ ],
16
+ "warmup_steps": 21,
17
+ "peak_lr": 5e-05,
18
+ "weights": "/mnt/nas/cross/Team/RobotAgent/Personal/baijiaqi/humanoid-benchmark-work/preflight-20261104T1648Z/policies-20261109/models/openpi/pi05_base_g1a76_pytorch",
19
+ "checkpoint_dir": "/mnt/nas/cross/Team/RobotAgent/Personal/baijiaqi/humanoid-benchmark-work/preflight-20261104T1648Z/policies-20261109/pi05-runs/chain-hb1110/b6/run/ckpt/pi05_hb_g1_b6/run",
20
+ "augment": true,
21
+ "aug_version": "hb-aug-v2-20261109",
22
+ "max_token_len": 256
23
+ }
b6-pedal-bin/epoch-020/assets/g1_b6/norm_stats.json ADDED
@@ -0,0 +1,504 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "norm_stats": {
3
+ "state": {
4
+ "mean": [
5
+ -0.09289415925741196,
6
+ 0.045334845781326294,
7
+ 0.048640452325344086,
8
+ 0.3030760884284973,
9
+ -0.11417802423238754,
10
+ -0.03437678515911102,
11
+ -0.06910363584756851,
12
+ -0.06624675542116165,
13
+ -0.28657519817352295,
14
+ 0.3067435622215271,
15
+ -0.15567664802074432,
16
+ -0.023483101278543472,
17
+ 0.014021046459674835,
18
+ 0.014938326552510262,
19
+ 0.14744997024536133,
20
+ 0.24639774858951569,
21
+ 0.19172275066375732,
22
+ -0.04901121184229851,
23
+ -0.3446519374847412,
24
+ -0.366971492767334,
25
+ -0.03786687180399895,
26
+ 0.17884966731071472,
27
+ 0.07702597975730896,
28
+ -0.3111043870449066,
29
+ 0.006915700156241655,
30
+ -0.10259399563074112,
31
+ 0.3281053602695465,
32
+ 0.05004274472594261,
33
+ -0.041460540145635605,
34
+ 0.9999999403953552,
35
+ 1.0,
36
+ 1.0,
37
+ 1.0,
38
+ 0.9999556541442871,
39
+ 4.9299426763127485e-8,
40
+ 0.8562398552894592,
41
+ 0.8557164669036865,
42
+ 0.862363874912262,
43
+ 0.870231568813324,
44
+ 0.9374372959136963,
45
+ 0.0013326209736987948,
46
+ -0.0698074996471405,
47
+ 0.004781579133123159,
48
+ -0.9963920712471008
49
+ ],
50
+ "std": [
51
+ 0.22050054371356964,
52
+ 0.07048683613538742,
53
+ 0.134409561753273,
54
+ 0.2732667624950409,
55
+ 0.13167059421539307,
56
+ 0.052785616368055344,
57
+ 0.1399744749069214,
58
+ 0.05429463088512421,
59
+ 0.1418997198343277,
60
+ 0.1645597666501999,
61
+ 0.11816693097352982,
62
+ 0.04961647093296051,
63
+ 0.04337943345308304,
64
+ 0.019502397626638412,
65
+ 0.02250182442367077,
66
+ 0.06032394990324974,
67
+ 0.07465467602014542,
68
+ 0.10689205676317215,
69
+ 0.048137277364730835,
70
+ 0.04392699897289276,
71
+ 0.06392653286457062,
72
+ 0.06639137864112854,
73
+ 0.30172207951545715,
74
+ 0.23119832575321198,
75
+ 0.20930008590221405,
76
+ 0.5789200663566589,
77
+ 0.2542855143547058,
78
+ 0.23488010466098785,
79
+ 0.218588188290596,
80
+ 0.0,
81
+ 0.0,
82
+ 0.0,
83
+ 0.0,
84
+ 0.0,
85
+ 6.862556801934261e-6,
86
+ 0.16237150132656097,
87
+ 0.16346286237239838,
88
+ 0.1562580168247223,
89
+ 0.14771395921707153,
90
+ 0.06923356652259827,
91
+ 0.001731409691274166,
92
+ 0.04638409987092018,
93
+ 0.023614883422851562,
94
+ 0.0
95
+ ],
96
+ "q01": [
97
+ -0.880411240274788,
98
+ -0.11660030736746661,
99
+ -0.2591128731101498,
100
+ 0.1310143638281691,
101
+ -0.3975315273160438,
102
+ -0.12719849527707686,
103
+ -0.43698455556718624,
104
+ -0.2454130813051354,
105
+ -0.5891453627907941,
106
+ 0.15688254223430617,
107
+ -0.45365863902816306,
108
+ -0.1532304523474964,
109
+ -0.05649498579946606,
110
+ -0.03512588135941741,
111
+ 0.10754643191276736,
112
+ 0.09098392945162612,
113
+ 0.06845034108700554,
114
+ -0.2926570351384775,
115
+ -0.4391649112390142,
116
+ -0.4684700581492143,
117
+ -0.16496463962956748,
118
+ 0.07493410048820395,
119
+ -0.676092325818013,
120
+ -1.0350512746883667,
121
+ -0.43236514993125463,
122
+ -0.8278534394291747,
123
+ -0.051698092269253815,
124
+ -0.5183220580090175,
125
+ -0.4080306218162441,
126
+ 0.999207079410553,
127
+ 0.9995272755622864,
128
+ 0.9994345307350159,
129
+ 0.9995204210281372,
130
+ 0.9992793597638138,
131
+ 0.0,
132
+ 0.6215853575778556,
133
+ 0.6115407229258378,
134
+ 0.6271049689326582,
135
+ 0.6401020124295089,
136
+ 0.8316517208587076,
137
+ -1e-10,
138
+ -0.18501513771914246,
139
+ -0.05655392329480376,
140
+ -0.999953441719553
141
+ ],
142
+ "q99": [
143
+ 0.1713271221507653,
144
+ 0.16606888934973485,
145
+ 0.255133149306362,
146
+ 1.4343218416003525,
147
+ 0.18631403401933788,
148
+ 0.10816316061141407,
149
+ 0.24138335087282747,
150
+ 0.03144279724399046,
151
+ -0.027339548691011184,
152
+ 0.9277804614462278,
153
+ 0.06613453889172527,
154
+ 0.09583480280191259,
155
+ 0.14404179024680713,
156
+ 0.05739128976607806,
157
+ 0.235095005062959,
158
+ 0.34817335710089825,
159
+ 0.3896389798800358,
160
+ 0.10347168962327696,
161
+ -0.20304301404597563,
162
+ -0.2800530286244159,
163
+ 0.07539925654783991,
164
+ 0.3544070315619453,
165
+ 0.5230730046380223,
166
+ 0.06049067594124291,
167
+ 0.520795445975837,
168
+ 1.1251266785638767,
169
+ 1.12142253197707,
170
+ 0.5207122732182838,
171
+ 0.5347526467551684,
172
+ 1.0,
173
+ 1.0,
174
+ 1.0,
175
+ 1.0,
176
+ 0.9999972909258038,
177
+ 0.0010093163000419736,
178
+ 0.9999195548103482,
179
+ 0.9999173667715222,
180
+ 0.9999223136351736,
181
+ 0.9999271019862325,
182
+ 0.9999653391407162,
183
+ 0.005770046510408187,
184
+ 0.0100407091547432,
185
+ 0.06605743221344167,
186
+ -0.9818596055855306
187
+ ]
188
+ },
189
+ "actions": {
190
+ "mean": [
191
+ 0.03432026505470276,
192
+ 0.012461589649319649,
193
+ -0.05914423614740372,
194
+ -0.0072383214719593525,
195
+ 0.05019465461373329,
196
+ 0.11229633539915085,
197
+ 0.06718273460865021,
198
+ -0.004092664457857609,
199
+ 0.0245615653693676,
200
+ -0.2105632871389389,
201
+ -0.14285634458065033,
202
+ -0.199773371219635,
203
+ -0.041197869926691055,
204
+ 0.06098661199212074,
205
+ 0.1831471472978592,
206
+ -0.06344112753868103,
207
+ 0.05129233002662659,
208
+ -0.062251925468444824,
209
+ -0.009263630956411362,
210
+ -0.0808362290263176,
211
+ -0.07540543377399445,
212
+ -0.15363413095474243,
213
+ -0.11247213184833527,
214
+ -0.10093164443969727,
215
+ 0.01776888780295849,
216
+ -0.03236458823084831,
217
+ -0.17563124001026154,
218
+ -0.03566235303878784,
219
+ 0.03622003644704819,
220
+ -0.005657575558871031,
221
+ -0.002754294080659747,
222
+ -0.172866091132164,
223
+ 0.04396877810359001,
224
+ 0.12563470005989075,
225
+ 0.027359044179320335,
226
+ 0.08746480196714401,
227
+ -0.07898023724555969,
228
+ -0.05846341699361801,
229
+ 0.12352164089679718,
230
+ -0.07225605845451355,
231
+ -0.07172008603811264,
232
+ -0.0689985379576683,
233
+ 0.09783204644918442,
234
+ -0.1205044537782669,
235
+ 0.04385637864470482,
236
+ -0.0408664271235466,
237
+ -0.0202995203435421,
238
+ -0.05485040321946144,
239
+ -0.2798519432544708,
240
+ -0.06503159552812576,
241
+ 0.019104620441794395,
242
+ 0.07799244672060013,
243
+ 0.024153348058462143,
244
+ -0.07540969550609589,
245
+ -0.002842223271727562,
246
+ 0.09865490347146988,
247
+ -0.052052684128284454,
248
+ 0.07384795695543289,
249
+ 0.13516001403331757,
250
+ -0.16005200147628784,
251
+ -0.05993583798408508,
252
+ 0.06282256543636322,
253
+ 0.17966081202030182,
254
+ -0.06581006199121475,
255
+ 1.0,
256
+ 1.0,
257
+ 1.0,
258
+ 1.0,
259
+ 0.9999973177909851,
260
+ 4.929941965770013e-8,
261
+ 0.8565258383750916,
262
+ 0.8557460904121399,
263
+ 0.8630558848381042,
264
+ 0.8705127835273743,
265
+ 0.9372830390930176,
266
+ 0.0013317195698618889
267
+ ],
268
+ "std": [
269
+ 0.10664359480142593,
270
+ 0.1436038762331009,
271
+ 0.12584248185157776,
272
+ 0.10123310983181,
273
+ 0.18608538806438446,
274
+ 0.12901633977890015,
275
+ 0.16723662614822388,
276
+ 0.08674019575119019,
277
+ 0.09300430119037628,
278
+ 0.07599232345819473,
279
+ 0.13231238722801208,
280
+ 0.10487758368253708,
281
+ 0.12685155868530273,
282
+ 0.13161525130271912,
283
+ 0.10486916452646255,
284
+ 0.1401214748620987,
285
+ 0.08177037537097931,
286
+ 0.14273640513420105,
287
+ 0.15247996151447296,
288
+ 0.09722961485385895,
289
+ 0.09942055493593216,
290
+ 0.12243745476007462,
291
+ 0.10003399103879929,
292
+ 0.10224147886037827,
293
+ 0.08081497251987457,
294
+ 0.11588867008686066,
295
+ 0.0635068491101265,
296
+ 0.12075082212686539,
297
+ 0.10402055084705353,
298
+ 0.12647128105163574,
299
+ 0.13117629289627075,
300
+ 0.12363912910223007,
301
+ 0.08622833341360092,
302
+ 0.09993847459554672,
303
+ 0.051116377115249634,
304
+ 0.09662438929080963,
305
+ 0.12317366898059845,
306
+ 0.15041758120059967,
307
+ 0.08345945179462433,
308
+ 0.10266817361116409,
309
+ 0.11730080097913742,
310
+ 0.13897928595542908,
311
+ 0.13960976898670197,
312
+ 0.07891123741865158,
313
+ 0.107028529047966,
314
+ 0.13991400599479675,
315
+ 0.113197460770607,
316
+ 0.15233290195465088,
317
+ 0.10571637749671936,
318
+ 0.13236956298351288,
319
+ 0.11429065465927124,
320
+ 0.08670087158679962,
321
+ 0.19508449733257294,
322
+ 0.13858312368392944,
323
+ 0.06682883948087692,
324
+ 0.16072209179401398,
325
+ 0.14802610874176025,
326
+ 0.05860956013202667,
327
+ 0.12714003026485443,
328
+ 0.0833347961306572,
329
+ 0.1227123811841011,
330
+ 0.1028541624546051,
331
+ 0.10684986412525177,
332
+ 0.13590067625045776,
333
+ 0.0,
334
+ 0.0,
335
+ 0.0,
336
+ 0.0,
337
+ 0.0,
338
+ 6.8625545281975064e-6,
339
+ 0.1610528975725174,
340
+ 0.16486616432666779,
341
+ 0.1496121883392334,
342
+ 0.15113823115825653,
343
+ 0.06507572531700134,
344
+ 0.0017317550955340266
345
+ ],
346
+ "q01": [
347
+ -0.31260000004288,
348
+ -0.31260000008336003,
349
+ -0.31265000003336,
350
+ -0.25000000005,
351
+ -0.37502500008824,
352
+ -0.18753750006924,
353
+ -0.25000000005,
354
+ -0.18751250005556,
355
+ -0.25010000003335997,
356
+ -0.4375625000778,
357
+ -0.37513750005388,
358
+ -0.43750000008,
359
+ -0.25005000006668,
360
+ -0.18757500008184003,
361
+ -0.06263750005388,
362
+ -0.43756250008668,
363
+ -0.18750000003999998,
364
+ -0.31250000008,
365
+ -0.37502500008572,
366
+ -0.31250000005,
367
+ -0.25001250006364,
368
+ -0.437500000075,
369
+ -0.31257500008184,
370
+ -0.37510000008464,
371
+ -0.18751250005556,
372
+ -0.25000000008,
373
+ -0.31252500006668,
374
+ -0.31251250006364,
375
+ -0.2500625000778,
376
+ -0.25001250006364,
377
+ -0.31260000008336003,
378
+ -0.43755000007144,
379
+ -0.25003750006923997,
380
+ -0.12501250006364,
381
+ -0.12502500006668,
382
+ -0.18750000008,
383
+ -0.31255000007144,
384
+ -0.43756250008668,
385
+ -0.06256250007780001,
386
+ -0.25001250006364,
387
+ -0.25003750006923997,
388
+ -0.31253750006923997,
389
+ -0.18756250004668,
390
+ -0.25002500007144,
391
+ -0.25007500008184,
392
+ -0.25010000008336003,
393
+ -0.25007500008184,
394
+ -0.37505000007648004,
395
+ -0.50007500008184,
396
+ -0.375000000075,
397
+ -0.37515000003336,
398
+ -0.25000000008,
399
+ -0.50002500008824,
400
+ -0.375000000075,
401
+ -0.125000000075,
402
+ -0.37505000007648004,
403
+ -0.31260000008336003,
404
+ -0.06252500007144002,
405
+ -0.12500000005,
406
+ -0.3750000001,
407
+ -0.25013750005388,
408
+ -0.18757500003848,
409
+ -0.06258750004548003,
410
+ -0.43762500007336,
411
+ 0.999207079410553,
412
+ 0.9995272755622864,
413
+ 0.9994345307350159,
414
+ 0.9995204210281372,
415
+ 0.9992983239831871,
416
+ 0.0,
417
+ 0.6215853575778556,
418
+ 0.6115407229258378,
419
+ 0.6271049689326582,
420
+ 0.6401020124295089,
421
+ 0.8316517208587076,
422
+ -1e-10
423
+ ],
424
+ "q99": [
425
+ 0.24985000008568004,
426
+ 0.31245000008332,
427
+ 0.24985000006664004,
428
+ 0.24995000008332002,
429
+ 0.5623125000882,
430
+ 0.37487500006920005,
431
+ 0.4374000000874799,
432
+ 0.18745000007776003,
433
+ 0.18745000008332002,
434
+ -0.06259999994447996,
435
+ 0.18743750008460003,
436
+ 0.06237500007996011,
437
+ 0.24990000006664004,
438
+ 0.37493750008179993,
439
+ 0.43737500006920005,
440
+ 0.24981250005996003,
441
+ 0.1873750000799601,
442
+ 0.1873750000799601,
443
+ 0.37485000008568004,
444
+ 0.18745000008332002,
445
+ 0.18737500006359997,
446
+ 0.062300000024959945,
447
+ 0.18737500006359997,
448
+ 0.18747500005384,
449
+ 0.18745000007776003,
450
+ 0.2498750000799601,
451
+ -0.06254999993335997,
452
+ 0.18743750008179993,
453
+ 0.18745000007776003,
454
+ 0.24993750008179993,
455
+ 0.24990000006664004,
456
+ 0.12490000005711999,
457
+ 0.24997500005384,
458
+ 0.37493750008179993,
459
+ 0.12495000006664003,
460
+ 0.3123750000799601,
461
+ 0.31237500007140007,
462
+ 0.24981250005996003,
463
+ 0.31240000005552004,
464
+ 0.18737500006359997,
465
+ 0.31237500006920005,
466
+ 0.31243750008460003,
467
+ 0.37493750007332005,
468
+ 0.062437500071399976,
469
+ 0.31243750008179993,
470
+ 0.37495000008332,
471
+ 0.24987500006359997,
472
+ 0.31238750005292004,
473
+ -0.0001249999364000276,
474
+ 0.18740000003748003,
475
+ 0.24990000007776003,
476
+ 0.2498750000799601,
477
+ 0.4373125000882,
478
+ 0.24980000004995995,
479
+ 0.18740000004996,
480
+ 0.43733750007644,
481
+ 0.37485000009995995,
482
+ 0.24993750007139998,
483
+ 0.37495000008332,
484
+ -0.00009999995004000839,
485
+ 0.24987500006920005,
486
+ 0.24987500006920005,
487
+ 0.37493750008179993,
488
+ 0.24993750007332005,
489
+ 1.0,
490
+ 1.0,
491
+ 1.0,
492
+ 1.0,
493
+ 0.9999972909258038,
494
+ 0.0010093163000419736,
495
+ 0.9999195548103482,
496
+ 0.9999173667715222,
497
+ 0.9999223136351736,
498
+ 0.9999271019862325,
499
+ 0.9999653391407162,
500
+ 0.005770046510408187
501
+ ]
502
+ }
503
+ }
504
+ }
b6-pedal-bin/epoch-020/metadata.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:383ea61cbf4e945e92628b46b64a8525602ae9be5d4331d000beb84cc39bd6e0
3
+ size 3379
b6-pedal-bin/epoch-020/model.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:fc8132cec11350d4cd5c3e37219c38f1be3c35999d59da93b9a2998c22d585cb
3
+ size 7473452088
b6-pedal-bin/epoch-020/protocol.json ADDED
@@ -0,0 +1,56 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "model": "pi0.5 (openpi, PyTorch weights)",
3
+ "task": "B6 pedal-bin trash disposal (G1)",
4
+ "task_dir": "b6-pedal-bin",
5
+ "openpi_config": "pi05_hb_g1_b6",
6
+ "layout": "g1",
7
+ "asset_id": "g1_b6",
8
+ "epoch_completed": 20,
9
+ "global_step": 425,
10
+ "total_epochs": 20.0,
11
+ "total_steps": 425,
12
+ "global_batch": 3072,
13
+ "per_gpu_batch": 96,
14
+ "world_size": 32,
15
+ "peak_lr": 5e-05,
16
+ "warmup_steps": 21,
17
+ "max_token_len": 256,
18
+ "init_weights": "lerobot/pi05_base converted to PyTorch, action/state projections resized to 76-D ($P/policies-20261109/models/openpi/pi05_base_g1a76_pytorch)",
19
+ "augmentation": "Humanoid Benchmark online augmentation spec v2 (2026-11-09), on from epoch 1; model-native augmentation off",
20
+ "aug_version": "hb-aug-v2-20261109",
21
+ "model_action_dim": 76,
22
+ "action_horizon": 16,
23
+ "episodes": 100,
24
+ "frames": 65127,
25
+ "fps": 30.0,
26
+ "control_hz": 50.0,
27
+ "prompt": "Pick up the chip barrel from the cart, walk to the pedal bin, step on its pedal and drop it in.",
28
+ "raw_state_dim": 41,
29
+ "state_layout": "body29(observation.state by name) | inspire_left6 | inspire_right6 | projected_gravity3",
30
+ "action_layout": "action.motion_token64 | inspire_left6 | inspire_right6 (same-row observed hands), rows t..t+15",
31
+ "cameras": {
32
+ "base_0_rgb": "observation.images.ego_view",
33
+ "left_wrist_0_rgb": "observation.images.chest",
34
+ "right_wrist_0_rgb": "zeros, image_mask False"
35
+ },
36
+ "raw_image_hw": [
37
+ 480,
38
+ 640
39
+ ],
40
+ "content_box": [
41
+ 28,
42
+ 196,
43
+ 0,
44
+ 224
45
+ ],
46
+ "training_data": "$P/b6-v5t7-ready-start-60ep-20261109/data/deliveries/g1/b6-v5t7-ready-start-20261109",
47
+ "normalization": "assets/g1_b6/norm_stats.json (openpi quantile normalization)",
48
+ "inference": "Register configs by importing code/hb_pi05/hb_g1_config.py (needs hb_g1_data.py on the path), then openpi policy_config.create_trained_policy(get_config(openpi_config), <this folder>). Clients send observation/<camera> raw RGB, observation/state and prompt; see HBInputs.",
49
+ "resume_state_included": false,
50
+ "files": {
51
+ "assets/g1_b6/norm_stats.json": 13548,
52
+ "metadata.pt": 3379,
53
+ "model.safetensors": 7473452088,
54
+ "training_plan.json": 742
55
+ }
56
+ }
b6-pedal-bin/epoch-020/training_plan.json ADDED
@@ -0,0 +1,23 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "task": "b6",
3
+ "layout": "g1",
4
+ "config": "pi05_hb_g1_b6",
5
+ "exp_name": "run",
6
+ "frames": 65127,
7
+ "world_size": 32,
8
+ "per_gpu_batch": 96,
9
+ "global_batch": 3072,
10
+ "epochs": 20.0,
11
+ "num_train_steps": 425,
12
+ "save_steps": [
13
+ 213,
14
+ 425
15
+ ],
16
+ "warmup_steps": 21,
17
+ "peak_lr": 5e-05,
18
+ "weights": "/mnt/nas/cross/Team/RobotAgent/Personal/baijiaqi/humanoid-benchmark-work/preflight-20261104T1648Z/policies-20261109/models/openpi/pi05_base_g1a76_pytorch",
19
+ "checkpoint_dir": "/mnt/nas/cross/Team/RobotAgent/Personal/baijiaqi/humanoid-benchmark-work/preflight-20261104T1648Z/policies-20261109/pi05-runs/chain-hb1110/b6/run/ckpt/pi05_hb_g1_b6/run",
20
+ "augment": true,
21
+ "aug_version": "hb-aug-v2-20261109",
22
+ "max_token_len": 256
23
+ }
code/hb_pi05/hb_aug_v2.py ADDED
@@ -0,0 +1,163 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """Humanoid Benchmark data augmentation v2 (project CLAUDE.md), model-agnostic torch implementation.
2
+
3
+ augment(x, rng, obs_mask=None) takes x: float [T, V, 3, H, W] in [0, 1] (T frames, V cameras) and returns the
4
+ augmented tensor. Each call is one training sample: one of the 15 effects is chosen uniformly and its magnitude
5
+ is scaled by 1.5 around identity; the rest stay at x1. Colour is shared across cameras and time; geometry, light
6
+ patches and white balance are drawn per camera and fixed over time; sensor/optics effects (noise, gaussian blur,
7
+ JPEG, motion blur, sharpen, defocus) apply only to observation frames (obs_mask True).
8
+ Pixel-size parameters are defined at 224 px and scaled by H / 224.
9
+ perturb_joints(q, rng) applies the proprio spec: per-joint bias +-2 deg plus N(0, 0.25 deg) clipped at 3 sigma.
10
+ """
11
+ import io
12
+ import math
13
+
14
+ import numpy as np
15
+ import torch
16
+ import torch.nn.functional as F
17
+ from PIL import Image
18
+ from torchvision.transforms import InterpolationMode
19
+ from torchvision.transforms import functional as TF
20
+
21
+ EFFECTS = ("exposure", "contrast", "saturation", "hue", "gamma", "temperature", "rotation_crop", "noise",
22
+ "gaussian_blur", "jpeg", "motion_blur", "sharpen", "defocus", "light_shadow", "wb_channel")
23
+ CENTER = {"exposure": 0.0, "contrast": 1.0, "saturation": 1.0, "hue": 0.0, "gamma": 1.0, "temperature": 0.0}
24
+ RANGE = {"exposure": (-0.22, 0.22), "contrast": (0.88, 1.12), "saturation": (0.76, 1.24), "hue": (-0.015, 0.015),
25
+ "gamma": (0.89, 1.11), "temperature": (-0.066, 0.066)}
26
+ VERSION = "hb-aug-v2-20261109"
27
+
28
+
29
+ def sample(rng, views, height):
30
+ boost = EFFECTS[int(rng.integers(len(EFFECTS)))]
31
+
32
+ def k(name):
33
+ return 1.5 if boost == name else 1.0
34
+
35
+ px = height / 224.0
36
+ p = {"boost": boost}
37
+ for name, (a, b) in RANGE.items():
38
+ c = CENTER[name]
39
+ p[name] = float(rng.uniform(c + k(name) * (a - c), c + k(name) * (b - c)))
40
+ p["geometry"] = []
41
+ for _ in range(views):
42
+ angle = float(rng.uniform(-1.5 * k("rotation_crop"), 1.5 * k("rotation_crop")))
43
+ p["geometry"].append((angle, 1.0 + abs(angle) * 0.015 + float(rng.uniform(0, 0.02 * k("rotation_crop")))))
44
+ p["noise"] = 4.0 / 255.0 * k("noise") if rng.random() < 0.5 else 0.0
45
+ p["gaussian_blur"] = float(rng.uniform(0.25, 0.65)) * k("gaussian_blur") if rng.random() < 0.15 else 0.0
46
+ if rng.random() < 0.35:
47
+ lo = int(round(100 - k("jpeg") * 30))
48
+ hi = int(round(100 - k("jpeg") * 5))
49
+ p["jpeg"] = int(rng.integers(lo, hi + 1))
50
+ else:
51
+ p["jpeg"] = None
52
+ p["motion_blur"] = ((float(rng.uniform(3, 7)) * k("motion_blur") * px, float(rng.uniform(0, 180)))
53
+ if rng.random() < 0.15 else None)
54
+ p["sharpen"] = float(rng.uniform(0.2, 0.6)) * k("sharpen") if rng.random() < 0.15 else 0.0
55
+ p["defocus"] = float(rng.uniform(1, 2)) * k("defocus") * px if rng.random() < 0.10 else 0.0
56
+ if rng.random() < 0.30:
57
+ g = 0.30 * k("light_shadow")
58
+ p["light_shadow"] = [[(float(rng.uniform(0.1, 0.9)), float(rng.uniform(0.1, 0.9)),
59
+ float(rng.uniform(0.15, 0.35)), float(rng.uniform(1 - g, 1 + g)))
60
+ for _ in range(int(rng.integers(1, 3)))] for _ in range(views)]
61
+ else:
62
+ p["light_shadow"] = None
63
+ if rng.random() < 0.30:
64
+ g = 0.10 * k("wb_channel")
65
+ gains = []
66
+ for _ in range(views):
67
+ row = [1.0, 1.0, 1.0]
68
+ row[int(rng.integers(0, 3))] = float(rng.uniform(1 - g, 1 + g))
69
+ gains.append(row)
70
+ p["wb_channel"] = gains
71
+ else:
72
+ p["wb_channel"] = None
73
+ return p
74
+
75
+
76
+ def _conv(x, kernel):
77
+ n, c, h, w = x.shape
78
+ k = kernel.to(x)[None, None].repeat(c, 1, 1, 1)
79
+ pad = kernel.shape[-1] // 2
80
+ return F.conv2d(F.pad(x, (pad, pad, pad, pad), mode="reflect"), k, groups=c)
81
+
82
+
83
+ def motion_kernel(length, angle_deg):
84
+ size = max(3, int(round(length)) | 1)
85
+ k = torch.zeros(size, size)
86
+ c = size // 2
87
+ th = math.radians(angle_deg)
88
+ for s in np.linspace(-(length - 1) / 2, (length - 1) / 2, 8 * size):
89
+ k[int(round(c - s * math.sin(th))), int(round(c + s * math.cos(th)))] = 1.0
90
+ return k / k.sum()
91
+
92
+
93
+ def disk_kernel(r):
94
+ size = 2 * int(math.ceil(r)) + 1
95
+ c = size // 2
96
+ yy, xx = torch.meshgrid(torch.arange(size), torch.arange(size), indexing="ij")
97
+ k = (((yy - c) ** 2 + (xx - c) ** 2) <= (r + 0.5) ** 2).float()
98
+ return k / k.sum()
99
+
100
+
101
+ def apply(x, p, obs_mask=None, noise_seed=0):
102
+ t, v, c, h, w = x.shape
103
+ x = x.clone().clamp(0, 1)
104
+ for vi, (angle, scale) in enumerate(p["geometry"]):
105
+ x[:, vi] = TF.affine(x[:, vi], angle=angle, translate=[0, 0], scale=scale, shear=[0, 0],
106
+ interpolation=InterpolationMode.BILINEAR)
107
+ if p["light_shadow"] is not None:
108
+ yy, xx = torch.meshgrid(torch.linspace(0, 1, h), torch.linspace(0, 1, w), indexing="ij")
109
+ for vi, blobs in enumerate(p["light_shadow"]):
110
+ m = torch.ones(h, w)
111
+ for cx, cy, s, g in blobs:
112
+ m = m + (g - 1.0) * torch.exp(-((xx - cx) ** 2 + (yy - cy) ** 2) / (2 * s * s))
113
+ x[:, vi] = (x[:, vi] * m.clamp(min=0).to(x)).clamp(0, 1)
114
+ if p["wb_channel"] is not None:
115
+ for vi, g in enumerate(p["wb_channel"]):
116
+ x[:, vi] = (x[:, vi] * torch.tensor(g, dtype=x.dtype).view(1, 3, 1, 1)).clamp(0, 1)
117
+ y = x.reshape(t * v, c, h, w)
118
+ y = (y * float(2.0 ** p["exposure"])).clamp(0, 1)
119
+ d = p["temperature"]
120
+ y = (y * y.new_tensor([1.0 + d, 1.0, 1.0 - d]).view(1, 3, 1, 1)).clamp(0, 1)
121
+ y = ((y - 0.5) * p["contrast"] + 0.5).clamp(0, 1)
122
+ y = TF.adjust_saturation(y, p["saturation"])
123
+ y = TF.adjust_hue(y, p["hue"])
124
+ y = y.clamp(0, 1).pow(p["gamma"])
125
+ x = y.reshape(t, v, c, h, w)
126
+ mask = torch.ones(t, dtype=torch.bool) if obs_mask is None else torch.as_tensor(obs_mask, dtype=torch.bool)
127
+ if bool(mask.any()):
128
+ o = x[mask].reshape(-1, c, h, w)
129
+ if p["gaussian_blur"] > 0:
130
+ o = TF.gaussian_blur(o, [3, 3], [p["gaussian_blur"]] * 2)
131
+ if p["motion_blur"] is not None:
132
+ o = _conv(o, motion_kernel(*p["motion_blur"]))
133
+ if p["defocus"] > 0:
134
+ o = _conv(o, disk_kernel(p["defocus"]))
135
+ if p["sharpen"] > 0:
136
+ o = o + p["sharpen"] * (o - TF.gaussian_blur(o, [5, 5], [1.0, 1.0]))
137
+ o = o.clamp(0, 1)
138
+ if p["noise"] > 0:
139
+ gen = torch.Generator().manual_seed(int(noise_seed))
140
+ o = (o + torch.randn(o.shape, generator=gen, dtype=o.dtype) * p["noise"]).clamp(0, 1)
141
+ if p["jpeg"] is not None:
142
+ for i in range(o.shape[0]):
143
+ buf = io.BytesIO()
144
+ TF.to_pil_image(o[i]).save(buf, format="JPEG", quality=int(p["jpeg"]))
145
+ buf.seek(0)
146
+ with Image.open(buf) as im:
147
+ o[i] = TF.pil_to_tensor(im.convert("RGB")).to(o.dtype).div(255.0)
148
+ x[mask] = o.reshape(-1, v, c, h, w)
149
+ return x.clamp(0, 1)
150
+
151
+
152
+ def augment(x, rng, obs_mask=None):
153
+ p = sample(rng, x.shape[1], x.shape[-2])
154
+ return apply(x, p, obs_mask, noise_seed=int(rng.integers(0, 2 ** 62))), p
155
+
156
+
157
+ def perturb_joints(q, rng):
158
+ q = np.asarray(q, dtype=np.float64)
159
+ d = q.shape[-1]
160
+ bias = rng.uniform(-math.radians(2.0), math.radians(2.0), size=d)
161
+ sigma = math.radians(0.25)
162
+ noise = np.clip(rng.normal(0.0, sigma, size=q.shape), -3 * sigma, 3 * sigma)
163
+ return (q + bias + noise).astype(np.float32)
code/hb_pi05/hb_g1_config.py ADDED
@@ -0,0 +1,144 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """OpenPI TrainConfigs for the HB tasks. Importing this module registers in openpi.training.config._CONFIGS_DICT:
2
+ pi05_hb_g1_b6, pi05_hb_g1_basket, pi05_hb_g1_a3, pi05_hb_g1_a7 (G1: state 44, actions 76)
3
+ pi05_hb_h2_a18 (H2: state 78, actions 75 + 1 zero pad)
4
+ so get_config()/cli() and policy_config.create_trained_policy() can use them.
5
+
6
+ Model (all tasks): Pi0Config(pi05=True, action_dim=76, action_horizon=16) and one shared 76-D init checkpoint.
7
+ For pi05, ModelTransformFactory (training/config.py) runs TokenizePrompt(discrete_state_input=True) BEFORE
8
+ PadStatesAndActions(76): the normalized state is discretized into 256 bins and written into the prompt
9
+ ("Task: ..., State: s1 ... sS;\\nAction: "), then state/actions are zero-padded up to 76 (pad_to_dim never truncates, so
10
+ the H2 state stays 78 wide; the pi05 network only uses observation.state for its batch shape). H2 actions (75) get one
11
+ zero dim, sliced off again by HBOutputs. Each state value takes ~4 tokens, so max_token_len is per robot layout
12
+ (g1 256, h2 384; default pi05 is 200); verify_cpu.py checks nothing is truncated.
13
+ """
14
+ import dataclasses
15
+ import os
16
+ import pathlib
17
+
18
+ import numpy as np
19
+ from typing_extensions import override
20
+
21
+ import openpi.models.pi0_config as pi0_config
22
+ import openpi.training.config as _config
23
+ import openpi.training.optimizer as _optimizer
24
+ import openpi.transforms as _transforms
25
+
26
+ import hb_g1_data as D
27
+
28
+
29
+ def _uint8_hwc(img) -> np.ndarray: # same as libero_policy._parse_image
30
+ img = np.asarray(img)
31
+ if np.issubdtype(img.dtype, np.floating):
32
+ img = (255 * img).astype(np.uint8)
33
+ if img.shape[0] == 3:
34
+ img = np.moveaxis(img, 0, -1)
35
+ return img
36
+
37
+
38
+ @dataclasses.dataclass(frozen=True)
39
+ class HBInputs(_transforms.DataTransformFn):
40
+ """Training and inference input mapping (cf. policies/libero_policy.py:LiberoInputs). Inference clients send
41
+ observation/<cam> (raw 480x640 RGB uint8; ResizeImages(224, 224) does the identical resize_with_pad),
42
+ observation/state and prompt."""
43
+
44
+ layout: str = "g1"
45
+
46
+ def __call__(self, data: dict) -> dict:
47
+ L = D.LAYOUTS[self.layout]
48
+ base = _uint8_hwc(data[f"observation/{L.cameras[0]}"])
49
+ wrist = _uint8_hwc(data[f"observation/{L.cameras[1]}"])
50
+ state = np.asarray(data["observation/state"], dtype=np.float32)
51
+ if state.shape[-1] != L.state_dim:
52
+ raise ValueError(f"{self.layout}: state dim {state.shape} != {L.state_dim}")
53
+ out = {
54
+ "state": state,
55
+ "image": {"base_0_rgb": base, "left_wrist_0_rgb": wrist, "right_wrist_0_rgb": np.zeros_like(base)},
56
+ "image_mask": {"base_0_rgb": np.True_, "left_wrist_0_rgb": np.True_, "right_wrist_0_rgb": np.False_},
57
+ }
58
+ if "actions" in data:
59
+ actions = np.asarray(data["actions"], dtype=np.float32)
60
+ if actions.shape[-1] != L.action_dim:
61
+ raise ValueError(f"{self.layout}: action dim {actions.shape} != {L.action_dim}")
62
+ out["actions"] = actions
63
+ if "prompt" in data:
64
+ out["prompt"] = data["prompt"]
65
+ return out
66
+
67
+
68
+ @dataclasses.dataclass(frozen=True)
69
+ class HBOutputs(_transforms.DataTransformFn):
70
+ """Slices the 76-D model output back to the robot's raw action dim (h2: drops the zero pad dim). Runs after
71
+ Unnormalize, whose quantile branch leaves dims beyond the 75-D stats untouched."""
72
+
73
+ action_dim: int = D.MODEL_ACTION_DIM
74
+
75
+ def __call__(self, data: dict) -> dict:
76
+ return {"actions": np.asarray(data["actions"][..., : self.action_dim])}
77
+
78
+
79
+ @dataclasses.dataclass(frozen=True)
80
+ class HBDataConfig(_config.DataConfigFactory):
81
+ """repo_id 'hb_<layout>_<task>' is resolved to HBCachedDataset by hb_g1_data.install_dataset_hook (training only).
82
+ The dataset already yields the HBInputs keys and the prompt, so no repack transform / PromptFromLeRobotTask."""
83
+
84
+ repo_id: str = "hb_g1_b6"
85
+ layout: str = "g1"
86
+
87
+ @override
88
+ def create(self, assets_dirs: pathlib.Path, model_config) -> _config.DataConfig:
89
+ L = D.LAYOUTS[self.layout]
90
+ return dataclasses.replace(
91
+ self.create_base_config(assets_dirs, model_config), # loads assets/<asset_id>/norm_stats.json, quantile norm for pi05
92
+ data_transforms=_transforms.Group(inputs=[HBInputs(self.layout)], outputs=[HBOutputs(L.action_dim)]),
93
+ model_transforms=_config.ModelTransformFactory()(model_config),
94
+ )
95
+
96
+
97
+ def model_config(layout: str = "g1", **kw) -> pi0_config.Pi0Config:
98
+ return pi0_config.Pi0Config(pi05=True, action_dim=D.MODEL_ACTION_DIM, action_horizon=D.HORIZON,
99
+ max_token_len=D.LAYOUTS[layout].max_token_len, **kw)
100
+
101
+
102
+ def make_train_config(task: str, *, exp_name: str = "default", batch_size: int = 32 * 32, num_train_steps: int = 1000,
103
+ peak_lr: float | None = None, warmup_frac: float | None = None, num_workers: int = 8,
104
+ checkpoint_base_dir: str | None = None, pytorch_weight_path: str | None = "default",
105
+ resume: bool = False, log_interval: int = 10) -> _config.TrainConfig:
106
+ L = D.layout_of(task)
107
+ peak_lr = float(os.environ.get("HB_PEAK_LR", "5e-5")) if peak_lr is None else peak_lr
108
+ warmup_frac = float(os.environ.get("HB_WARMUP_FRAC", "0.05")) if warmup_frac is None else warmup_frac
109
+ if pytorch_weight_path == "default":
110
+ pytorch_weight_path = os.environ.get("HB_PI05_INIT", str(D.G / "models/openpi/pi05_base_g1a76_pytorch"))
111
+ runs = os.environ.get("HB_PI05_RUNS", str(D.G / "pi05-runs"))
112
+ return _config.TrainConfig(
113
+ name=D.config_name(task),
114
+ exp_name=exp_name,
115
+ model=model_config(L.name),
116
+ data=HBDataConfig(repo_id=D.repo_id(task), layout=L.name,
117
+ assets=_config.AssetsConfig(assets_dir=str(D.task_dir(task) / "assets"), asset_id=D.asset_id(task))),
118
+ pytorch_weight_path=pytorch_weight_path,
119
+ pytorch_training_precision="bfloat16",
120
+ # train_pytorch.py: linear warmup then cosine to decay_lr at decay_steps (= end of training here).
121
+ lr_schedule=_optimizer.CosineDecaySchedule(warmup_steps=max(1, round(warmup_frac * num_train_steps)),
122
+ peak_lr=peak_lr, decay_steps=num_train_steps, decay_lr=peak_lr / 10),
123
+ optimizer=_optimizer.AdamW(clip_gradient_norm=1.0),
124
+ ema_decay=None, # EMA is not supported by the PyTorch trainer
125
+ batch_size=batch_size, # GLOBAL batch; train_pytorch/data_loader divide by world size
126
+ num_workers=num_workers,
127
+ num_train_steps=num_train_steps,
128
+ log_interval=log_interval,
129
+ save_interval=10**9, # hb_train.py decides the save steps
130
+ keep_period=None,
131
+ wandb_enabled=False,
132
+ assets_base_dir=str(pathlib.Path(runs) / "assets"),
133
+ checkpoint_base_dir=checkpoint_base_dir or str(pathlib.Path(runs) / "checkpoints"),
134
+ resume=resume,
135
+ policy_metadata={"robot": L.name, "task": task, "state": L.state_desc, "actions": L.action_desc,
136
+ "action_dim": L.action_dim, "action_horizon": D.HORIZON},
137
+ )
138
+
139
+
140
+ for _task in D.TASKS:
141
+ _cfg = make_train_config(_task)
142
+ if _cfg.name not in _config._CONFIGS_DICT: # noqa: SLF001
143
+ _config._CONFIGS.append(_cfg) # noqa: SLF001
144
+ _config._CONFIGS_DICT[_cfg.name] = _cfg # noqa: SLF001
code/hb_pi05/hb_g1_data.py ADDED
@@ -0,0 +1,264 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """Robot layouts, task table, cached torch Dataset, HB augmentation v2, and the OpenPI hooks.
2
+ (The file name dates from the G1-only version; it now covers the G1 tasks b6/basket/a3/a7 and the H2 task a18.)
3
+
4
+ Sample produced by HBCachedDataset (consumed by hb_g1_config.HBInputs), per robot layout (LAYOUTS):
5
+ observation/<cam> uint8 [224, 224, 3] for the layout's two cameras (OpenPI resize_with_pad of 480x640, done in hb_prepare)
6
+ g1: ego_view -> base_0_rgb, chest -> left_wrist_0_rgb; h2: head -> base_0_rgb, chest -> left_wrist_0_rgb
7
+ observation/state float32 [44] g1 = body29 | inspire L6 | inspire R6 | gravity3
8
+ float32 [78] h2 = observation.state75 | gravity3
9
+ actions float32 [16, 76] g1 = rows t..t+15 of (motion_token64 | inspire L6 | inspire R6)
10
+ float32 [16, 75] h2 = rows t..t+15 of (action.body_target31 | action.hands44)
11
+ (clamped to the episode's last row, LeRobot delta_timestamps convention)
12
+ prompt str (meta/tasks.jsonl)
13
+ Augmentation (training only, augment=True), CLAUDE.md v2:
14
+ visual hb_aug_v2.augment on the two 224px views after the resize, letterbox rows re-zeroed.
15
+ proprio (sec. 3) with p=0.15 the whole state observation is taken from 1-2 control periods earlier
16
+ (round(k * fps / control_hz) rows, clamped to the episode start); then the leading n_body joints
17
+ (g1: 29 body, h2: 31 = body + head) get a per-trajectory bias U(+-2 deg), drawn deterministically from
18
+ (HB_AUG_SEED, loader pass, episode_index), so it is fixed within an episode for a pass and redrawn each pass,
19
+ plus per-sample N(0, 0.25 deg) noise clipped at 3 sigma. Same constants as hb_aug_v2.perturb_joints, which
20
+ draws bias and noise together and therefore cannot be used per trajectory.
21
+ History-drop and joint velocity (sec. 3/4) do not apply: pi0.5 has no state history / velocity input.
22
+ The loader pass reaches the workers through the index: hb_train's sampler yields i + pass * N.
23
+ Action labels and norm stats (hb_prepare.py) always use the raw values.
24
+ """
25
+ import json
26
+ import math
27
+ import os
28
+ import sys
29
+ from dataclasses import dataclass
30
+ from pathlib import Path
31
+
32
+ import numpy as np
33
+ import torch
34
+
35
+ H = Path(__file__).resolve().parent
36
+ G = Path(os.environ.get("HB_G", H.parent)) # $P/policies-20261109
37
+ P = Path(os.environ.get("HB_P", G.parent)) # .../preflight-20261104T1648Z (dataset parent)
38
+ try:
39
+ import hb_aug_v2 as aug
40
+ except ImportError: # fallback: GR00T folder next to us
41
+ sys.path.append(str(G / "hb_groot"))
42
+ import hb_aug_v2 as aug
43
+
44
+ HORIZON, IMG, MODEL_ACTION_DIM = 16, 224, 76 # one 76-D pi05 init checkpoint for all robots
45
+ BIAS_DEG, NOISE_DEG, DELAY_P = 2.0, 0.25, 0.15
46
+
47
+
48
+ @dataclass(frozen=True)
49
+ class Layout:
50
+ name: str
51
+ state_dim: int
52
+ action_dim: int # raw; zero-padded to MODEL_ACTION_DIM by OpenPI PadStatesAndActions after Normalize
53
+ cameras: tuple # (-> base_0_rgb, -> left_wrist_0_rgb)
54
+ n_body: int # leading state dims that get bias + noise
55
+ max_token_len: int # prompt + discretized state must fit (verify_cpu.py checks)
56
+ state_desc: str
57
+ action_desc: str
58
+
59
+
60
+ LAYOUTS = {
61
+ "g1": Layout("g1", 44, 76, ("ego_view", "chest"), 29, int(os.environ.get("HB_PI05_MAX_TOKEN_LEN_G1", "256")),
62
+ "body29(observation.state by name) | inspire_left6 | inspire_right6 | projected_gravity3",
63
+ "action.motion_token64 | inspire_left6 | inspire_right6 (same-row observed hands)"),
64
+ "h2": Layout("h2", 78, 75, ("head", "chest"), 31, int(os.environ.get("HB_PI05_MAX_TOKEN_LEN_H2", "384")),
65
+ "observation.state75 (legs12|waist3|arms14|head2|hands44) | projected_gravity3",
66
+ "action.body_target31 | action.hands44 (+1 zero pad dim -> 76)"),
67
+ }
68
+ G1_BODY = ["left_hip_pitch_joint", "left_hip_roll_joint", "left_hip_yaw_joint", "left_knee_joint", "left_ankle_pitch_joint",
69
+ "left_ankle_roll_joint", "right_hip_pitch_joint", "right_hip_roll_joint", "right_hip_yaw_joint",
70
+ "right_knee_joint", "right_ankle_pitch_joint", "right_ankle_roll_joint", "waist_yaw_joint", "waist_roll_joint",
71
+ "waist_pitch_joint", "left_shoulder_pitch_joint", "left_shoulder_roll_joint", "left_shoulder_yaw_joint",
72
+ "left_elbow_joint", "left_wrist_roll_joint", "left_wrist_pitch_joint", "left_wrist_yaw_joint",
73
+ "right_shoulder_pitch_joint", "right_shoulder_roll_joint", "right_shoulder_yaw_joint", "right_elbow_joint",
74
+ "right_wrist_roll_joint", "right_wrist_pitch_joint", "right_wrist_yaw_joint"]
75
+ BODY = G1_BODY # backwards-compatible name
76
+
77
+
78
+ @dataclass(frozen=True)
79
+ class Task:
80
+ root: Path # LeRobot v2.1 root
81
+ limit: int | None # episode limit (B6: always the fixed first 100)
82
+ episodes: int # expected episodes in the view
83
+ frames: int | None # expected frames (None = not known in advance)
84
+ layout: str
85
+ control_hz: float # assumption: the controller rate the recording was made at (used for the state delay)
86
+
87
+
88
+ TASKS = {
89
+ "b6": Task(P / "b6-v5t7-ready-start-60ep-20261109/data/deliveries/g1/b6-v5t7-ready-start-20261109", 100, 100, None, "g1", 50.0),
90
+ "basket": Task(P / "g1-basket-transfer-ed77719-40epochs/dataset", None, 100, 54337, "g1", 30.0),
91
+ "a3": Task(P / "a3-a7-20261109/a3-fridge-open-expert-v1/780fcf1e4bd7db947a8c2d8f7b92e61f24f71b45", None, 100, 28180, "g1", 30.0),
92
+ "a7": Task(P / "a3-a7-20261109/a7-cabinet-close-expert-v1/1935a98e8f55d4d1d2d1615ddb97c2679713595c", None, 100, 30664, "g1", 30.0),
93
+ "a18": Task(P / "a18-h2-dataset-audit-8e956a10/dataset", None, 100, 72752, "h2", 30.0),
94
+ }
95
+
96
+
97
+ def layout_of(task: str) -> Layout:
98
+ return LAYOUTS[TASKS[task].layout]
99
+
100
+
101
+ def config_name(task: str) -> str:
102
+ return f"pi05_hb_{TASKS[task].layout}_{task}"
103
+
104
+
105
+ def repo_id(task: str) -> str:
106
+ return f"hb_{TASKS[task].layout}_{task}"
107
+
108
+
109
+ def asset_id(task: str) -> str:
110
+ return f"{TASKS[task].layout}_{task}"
111
+
112
+
113
+ def data_root() -> Path:
114
+ return Path(os.environ.get("HB_PI05_DATA", G / "pi05-data"))
115
+
116
+
117
+ def task_dir(task: str) -> Path:
118
+ if task not in TASKS:
119
+ raise KeyError(f"unknown task {task!r}; known: {sorted(TASKS)}")
120
+ return data_root() / task
121
+
122
+
123
+ def trajectory_bias(seed: int, epoch: int, episode: int, n: int) -> np.ndarray:
124
+ """Per-joint +-2 deg bias, fixed for (seed, loader pass, episode) -> identical on every rank/worker."""
125
+ rng = np.random.default_rng([int(seed), int(epoch), int(episode)])
126
+ return rng.uniform(-math.radians(BIAS_DEG), math.radians(BIAS_DEG), size=n)
127
+
128
+
129
+ def joint_noise(rng: np.random.Generator, n: int) -> np.ndarray:
130
+ s = math.radians(NOISE_DEG)
131
+ return np.clip(rng.normal(0.0, s, size=n), -3 * s, 3 * s)
132
+
133
+
134
+ def delay_rows(rng: np.random.Generator, fps: float, control_hz: float) -> int:
135
+ if rng.random() >= DELAY_P:
136
+ return 0
137
+ k = int(rng.integers(1, 3)) # 1 or 2 control periods
138
+ return max(1, round(k * fps / control_hz))
139
+
140
+
141
+ class HBCachedDataset(torch.utils.data.Dataset):
142
+ def __init__(self, root, *, action_horizon: int = HORIZON, augment: bool = False, seed: int | None = None):
143
+ root = Path(root) if ("/" in str(root) or "\\" in str(root)) else task_dir(str(root))
144
+ if not (root / "READY").exists():
145
+ raise FileNotFoundError(f"{root}/READY missing; run hb_prepare.py first")
146
+ meta = json.loads((root / "meta.json").read_text())
147
+ self.layout = LAYOUTS[meta.get("layout", "g1")]
148
+ low = np.load(root / "lowdim.npz")
149
+ self.state = low["state"].astype(np.float32)
150
+ self.action = low["action"].astype(np.float32)
151
+ self.ep_end = low["ep_end"].astype(np.int64)
152
+ self.episode = low["episode_index"].astype(np.int64)
153
+ _, first = np.unique(self.episode, return_index=True)
154
+ self.ep_start = first[self.episode].astype(np.int64) # episodes are contiguous and indexed 0..n-1
155
+ L = self.layout
156
+ assert self.state.shape[1:] == (L.state_dim,) and self.action.shape[1:] == (L.action_dim,), (self.state.shape, self.action.shape)
157
+ self.n = len(self.state)
158
+ self.root, self.h, self.augment = str(root), int(action_horizon), bool(augment)
159
+ self.prompt = meta["prompt"]
160
+ self.box = tuple(meta["content_box"]) # top, bottom, left, right of the non-padded area in the 224 image
161
+ self.fps = float(meta["fps"])
162
+ self.control_hz = float(meta.get("control_hz", self.fps))
163
+ self.seed = int(os.environ.get("HB_AUG_SEED", "42")) if seed is None else int(seed)
164
+ self._frames = self._rng = self._pid = None
165
+
166
+ def __len__(self):
167
+ return self.n
168
+
169
+ def __getstate__(self): # never pickle memmaps / rng into DataLoader workers
170
+ d = dict(self.__dict__)
171
+ d["_frames"] = d["_rng"] = d["_pid"] = None
172
+ return d
173
+
174
+ def _lazy_init(self):
175
+ if self._pid == os.getpid():
176
+ return
177
+ self._frames = {c: np.load(f"{self.root}/frames_{c}.npy", mmap_mode="r") for c in self.layout.cameras}
178
+ if torch.utils.data.get_worker_info() is not None:
179
+ torch.set_num_threads(1)
180
+ # torch.initial_seed() differs per worker; OpenPI seeds every rank's loader identically, so mix in the global rank.
181
+ self._rng = np.random.default_rng([torch.initial_seed() % (2**63), int(os.environ.get("RANK", "0"))])
182
+ self._pid = os.getpid()
183
+
184
+ def proprio(self, i: int, epoch: int) -> tuple[np.ndarray, int]:
185
+ """Augmented state for row i (training): optional delay, then trajectory bias + per-sample noise."""
186
+ d = delay_rows(self._rng, self.fps, self.control_hz)
187
+ j = max(i - d, int(self.ep_start[i]))
188
+ state = self.state[j].copy()
189
+ nb = self.layout.n_body
190
+ state[:nb] += (trajectory_bias(self.seed, epoch, self.episode[i], nb) + joint_noise(self._rng, nb)).astype(np.float32)
191
+ return state, i - j
192
+
193
+ def __getitem__(self, index):
194
+ epoch, i = divmod(int(index), self.n)
195
+ self._lazy_init()
196
+ idx = np.minimum(np.arange(i, i + self.h), self.ep_end[i] - 1)
197
+ cams = self.layout.cameras
198
+ imgs = np.stack([np.asarray(self._frames[c][i]) for c in cams]) # [2, 224, 224, 3] uint8
199
+ actions = self.action[idx].copy()
200
+ if self.augment:
201
+ imgs = self._augment_images(imgs)
202
+ state, _ = self.proprio(i, epoch)
203
+ else:
204
+ state = self.state[i].copy()
205
+ return {f"observation/{cams[0]}": imgs[0], f"observation/{cams[1]}": imgs[1], "observation/state": state,
206
+ "actions": actions, "prompt": self.prompt}
207
+
208
+ def _augment_images(self, imgs: np.ndarray) -> np.ndarray:
209
+ x = torch.from_numpy(imgs).permute(0, 3, 1, 2).float().div_(255.0)[None] # [T=1, V=2, 3, 224, 224]
210
+ x, _ = aug.augment(x, self._rng)
211
+ out = x[0].mul(255.0).round_().clamp_(0, 255).to(torch.uint8).permute(0, 2, 3, 1).numpy().copy()
212
+ top, bottom, left, right = self.box # keep OpenPI's letterbox exactly black, as at inference
213
+ out[:, :top] = 0
214
+ out[:, bottom:] = 0
215
+ out[:, :, :left] = 0
216
+ out[:, :, right:] = 0
217
+ return out
218
+
219
+
220
+ G1CachedDataset = HBCachedDataset # backwards-compatible name
221
+
222
+
223
+ def disable_openpi_image_aug():
224
+ """PI0Pytorch._preprocess_observation (models_pytorch/pi0_pytorch.py) calls
225
+ preprocessing_pytorch.preprocess_observation_pytorch(observation, train=True) in forward(); with train=True it does
226
+ random crop 95% + rotate +-5 deg (non-wrist) and brightness/contrast/saturation jitter. Force train=False (the resize
227
+ and default-mask logic still run)."""
228
+ import openpi.models_pytorch.preprocessing_pytorch as pp
229
+
230
+ orig = pp.preprocess_observation_pytorch
231
+ if getattr(orig, "_hb_noaug", False):
232
+ return
233
+
234
+ def no_aug(observation, *, train=False, **kwargs): # noqa: ARG001
235
+ return orig(observation, train=False, **kwargs)
236
+
237
+ no_aug._hb_noaug = True
238
+ pp.preprocess_observation_pytorch = no_aug
239
+
240
+
241
+ def install_dataset_hook(*, augment: bool):
242
+ """openpi.training.data_loader.create_torch_data_loader calls create_torch_dataset(data_config, action_horizon,
243
+ model_config); route repo_id 'hb_<layout>_<task>' to HBCachedDataset. Everything after it (transform_dataset,
244
+ Normalize, model transforms, DistributedSampler, TorchDataLoader) is unchanged OpenPI code."""
245
+ import openpi.training.data_loader as dl
246
+
247
+ orig = getattr(dl.create_torch_dataset, "_hb_orig", dl.create_torch_dataset)
248
+
249
+ def create_torch_dataset(data_config, action_horizon, model_config):
250
+ rid = data_config.repo_id or ""
251
+ parts = rid.split("_", 2)
252
+ if len(parts) == 3 and parts[0] == "hb" and parts[2] in TASKS:
253
+ if action_horizon != HORIZON:
254
+ raise ValueError(f"action_horizon {action_horizon} != {HORIZON}")
255
+ return HBCachedDataset(parts[2], action_horizon=action_horizon, augment=augment)
256
+ return orig(data_config, action_horizon, model_config)
257
+
258
+ create_torch_dataset._hb_orig = orig
259
+ dl.create_torch_dataset = create_torch_dataset
260
+
261
+
262
+ def install_hooks(*, augment: bool = True):
263
+ disable_openpi_image_aug()
264
+ install_dataset_hook(augment=augment)
code/hb_pi05/hb_init_weights.py ADDED
@@ -0,0 +1,173 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """pi05_base (JAX) -> PyTorch -> 76-D G1 init checkpoint.
2
+
3
+ convert : runs OpenPI's own converter examples/convert_jax_model_to_pytorch.py:main (pi05 path, float32, stock
4
+ 32-D config `pi05_droid_finetune` = Pi0Config(pi05=True, action_dim=32, action_horizon=16)). That converter
5
+ calls PI0Pytorch.load_state_dict(..., strict=False) and ignores the result; we record it and FAIL if anything
6
+ except the unused/tied `lm_head` weights was not loaded.
7
+ init76 : builds PI0Pytorch for action_dim=76 in float32 (one checkpoint shared by all tasks: G1 uses 76 dims, H2 a18
8
+ uses 75 + 1 zero pad dim; max_token_len does not affect parameters), copies every converted tensor whose name and
9
+ shape match, keeps the fresh nn.Linear init for whole layers that mismatch (only action_in_proj and
10
+ action_out_proj are allowed to), logs every fresh/unloaded key, FAILS on any other mismatch, and writes
11
+ model.safetensors that train_pytorch.py loads strictly via TrainConfig.pytorch_weight_path.
12
+ --lerobot: GCS (gs://openpi-assets) is ~100 KB/s from the DLC nodes, so the base can instead be HF
13
+ lerobot/pi05_base @ 22c30e5 (public, fp32 PyTorch port of the same weights); its names are remapped to
14
+ PI0Pytorch and the same strict checks apply.
15
+ Usage: python hb_init_weights.py convert --jax-dir $HB_PI05_BASE_JAX --out $HB_PI05_BASE_PT
16
+ python hb_init_weights.py init76 --base $HB_PI05_BASE_PT --out $HB_PI05_INIT
17
+ python hb_init_weights.py init76 --base $HB_G/models/lerobot-pi05_base --lerobot --out $HB_PI05_INIT
18
+ """
19
+ import argparse
20
+ import importlib.util
21
+ import json
22
+ import os
23
+ import re
24
+ import sys
25
+ from pathlib import Path
26
+
27
+ import safetensors.torch
28
+ import torch
29
+
30
+ ALLOWED_FRESH_LAYERS = {"action_in_proj", "action_out_proj"}
31
+ ALLOWED_UNLOADED = re.compile(os.environ.get("HB_ALLOW_UNLOADED", r".*lm_head\.weight$"))
32
+
33
+
34
+ def cmd_convert(a):
35
+ import openpi.models_pytorch.pi0_pytorch as p0
36
+
37
+ jax_dir, out = Path(a.jax_dir).resolve(), Path(a.out).resolve()
38
+ if "pi05" not in str(jax_dir):
39
+ sys.exit(f"converter selects the pi05 (adaRMS) layout only if 'pi05' is in the checkpoint path: {jax_dir}")
40
+ if not (jax_dir / "params").is_dir():
41
+ sys.exit(f"{jax_dir}/params missing (download gs://openpi-assets/checkpoints/pi05_base/ there)")
42
+ results = []
43
+ orig = torch.nn.Module.load_state_dict
44
+
45
+ def recording_load(self, state_dict, strict=True, assign=False):
46
+ res = orig(self, state_dict, strict=strict, assign=assign)
47
+ results.append(res)
48
+ return res
49
+
50
+ p0.PI0Pytorch.load_state_dict = recording_load
51
+ path = Path(os.environ["HB_OPENPI"]) / "examples/convert_jax_model_to_pytorch.py"
52
+ spec = importlib.util.spec_from_file_location("openpi_convert_jax_model_to_pytorch", path)
53
+ mod = importlib.util.module_from_spec(spec)
54
+ spec.loader.exec_module(mod)
55
+ mod.main(checkpoint_dir=str(jax_dir), config_name="pi05_droid_finetune", output_path=str(out), precision="float32")
56
+ p0.PI0Pytorch.load_state_dict = orig
57
+ assert len(results) == 1, results
58
+ missing, unexpected = list(results[0].missing_keys), list(results[0].unexpected_keys)
59
+ bad = [k for k in missing if not ALLOWED_UNLOADED.fullmatch(k)]
60
+ report = {"jax_dir": str(jax_dir), "missing_keys": missing, "unexpected_keys": unexpected, "bad": bad}
61
+ (out / "hb_convert_report.json").write_text(json.dumps(report, indent=2))
62
+ print("converter load_state_dict: missing", missing, "unexpected", unexpected)
63
+ if bad or unexpected:
64
+ sys.exit(f"CONVERT_FAIL: unloaded {bad} unexpected {unexpected}")
65
+ print("CONVERT_OK", out)
66
+
67
+
68
+ def remap_lerobot(src, target_keys):
69
+ """lerobot/pi05_base (LeRobot PI05Policy, PyTorch port of OpenPI's PI0Pytorch) -> OpenPI PI0Pytorch names.
70
+ Exact name, then without the policy's leading 'model.', then a unique suffix match. Normalizer buffers are dropped;
71
+ anything else unmatched fails, so every tensor is either mapped 1:1 or reported."""
72
+ tset = set(target_keys)
73
+ out, dropped, unmatched, renamed = {}, [], [], {}
74
+ for k, v in src.items():
75
+ cands = [k] + ([k[len("model."):]] if k.startswith("model.") else [])
76
+ hit = next((c for c in cands if c in tset), None)
77
+ if hit is None:
78
+ b = cands[-1]
79
+ suf = [t for t in tset if t.endswith("." + b) or b.endswith("." + t)]
80
+ hit = suf[0] if len(suf) == 1 else None
81
+ if hit is None:
82
+ (dropped if re.search(r"normaliz", k) else unmatched).append(k)
83
+ continue
84
+ if hit in out:
85
+ unmatched.append(f"duplicate target {hit} <- {k}")
86
+ continue
87
+ out[hit] = v
88
+ if hit != k:
89
+ renamed[k] = hit
90
+ return out, dropped, unmatched, renamed
91
+
92
+
93
+ def cmd_init76(a):
94
+ import openpi.models_pytorch.pi0_pytorch as p0
95
+
96
+ import hb_g1_config as C
97
+
98
+ base, out = Path(a.base), Path(a.out)
99
+ src = safetensors.torch.load_file(str(base / "model.safetensors"))
100
+ cfg = C.model_config(dtype="float32", pytorch_compile_mode=None)
101
+ torch.manual_seed(a.seed)
102
+ model = p0.PI0Pytorch(cfg)
103
+ target = model.state_dict()
104
+ lerobot_info = None
105
+ if a.lerobot:
106
+ src = {k: v.float() for k, v in src.items()}
107
+ src, dropped, unmatched, renamed = remap_lerobot(src, target.keys())
108
+ print(f"lerobot remap: {len(src)} mapped, {len(renamed)} renamed, dropped {dropped}")
109
+ if unmatched:
110
+ sys.exit("INIT76_FAIL lerobot keys without an OpenPI match:\n" + "\n".join(unmatched[:200]))
111
+ lerobot_info = {"dropped": dropped, "renamed_count": len(renamed),
112
+ "renamed_examples": dict(list(renamed.items())[:20])}
113
+ # tied parameters share one tensor; save_model keeps only one of their names
114
+ alias = {}
115
+ for name, p in model.named_parameters(remove_duplicate=False):
116
+ alias.setdefault(id(p), []).append(name)
117
+ tied = {n: names for names in alias.values() if len(names) > 1 for n in names}
118
+
119
+ fresh_layers, problems, unloaded = set(), [], []
120
+ for k, v in target.items():
121
+ layer = k.rsplit(".", 1)[0]
122
+ if k in src:
123
+ if tuple(src[k].shape) != tuple(v.shape):
124
+ if layer in ALLOWED_FRESH_LAYERS:
125
+ fresh_layers.add(layer)
126
+ else:
127
+ problems.append(f"shape {k}: base {tuple(src[k].shape)} vs new {tuple(v.shape)}")
128
+ elif any(t in src for t in tied.get(k, [])):
129
+ continue # loaded through its tied alias
130
+ elif ALLOWED_UNLOADED.fullmatch(k):
131
+ unloaded.append(k)
132
+ else:
133
+ problems.append(f"missing in base: {k}")
134
+ problems += [f"not in new model: {k}" for k in src if k not in target]
135
+ if problems:
136
+ sys.exit("INIT76_FAIL\n" + "\n".join(problems))
137
+ load = {k: t for k, t in src.items() if k.rsplit(".", 1)[0] not in fresh_layers}
138
+ res = model.load_state_dict(load, strict=False)
139
+ fresh_keys = sorted(k for k in target if k.rsplit(".", 1)[0] in fresh_layers)
140
+ allowed_missing = set(fresh_keys) | set(unloaded) | {k for k in tied if k not in src}
141
+ assert not res.unexpected_keys and set(res.missing_keys) <= allowed_missing, res
142
+ for k in fresh_keys:
143
+ print(f"FRESH {k}: new {tuple(target[k].shape)} (base {tuple(src[k].shape) if k in src else None})")
144
+ for k in unloaded:
145
+ print(f"UNLOADED (kept init, unused) {k}")
146
+ out.mkdir(parents=True, exist_ok=True)
147
+ safetensors.torch.save_model(model, str(out / "model.safetensors"))
148
+ report = {"base": str(base), "base_format": "lerobot" if a.lerobot else "openpi", "lerobot_remap": lerobot_info,
149
+ "seed": a.seed, "model_config": {"pi05": cfg.pi05, "action_dim": cfg.action_dim,
150
+ "action_horizon": cfg.action_horizon, "max_token_len": cfg.max_token_len}, "fresh_keys": fresh_keys,
151
+ "unloaded_keys": unloaded, "loaded_tensors": len(load), "total_tensors": len(target)}
152
+ (out / "hb_init_report.json").write_text(json.dumps(report, indent=2))
153
+ print(f"INIT76_OK {out}: loaded {len(load)} tensors, fresh {fresh_keys}")
154
+
155
+
156
+ def main():
157
+ ap = argparse.ArgumentParser()
158
+ sub = ap.add_subparsers(dest="cmd", required=True)
159
+ c = sub.add_parser("convert")
160
+ c.add_argument("--jax-dir", required=True)
161
+ c.add_argument("--out", required=True)
162
+ i = sub.add_parser("init76")
163
+ i.add_argument("--base", required=True)
164
+ i.add_argument("--out", required=True)
165
+ i.add_argument("--seed", type=int, default=0)
166
+ i.add_argument("--lerobot", action="store_true",
167
+ help="--base is lerobot/pi05_base (HF, PyTorch); remap its names to OpenPI PI0Pytorch")
168
+ a = ap.parse_args()
169
+ {"convert": cmd_convert, "init76": cmd_init76}[a.cmd](a)
170
+
171
+
172
+ if __name__ == "__main__":
173
+ main()
code/hb_pi05/hb_prepare.py ADDED
@@ -0,0 +1,278 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """Build the pi0.5 training view of the G1/H2 tasks: decoded 224px frame cache, state / action arrays, norm stats.
2
+
3
+ Usage (openpi env, CPU only, one node): python hb_prepare.py b6 basket a3 a7 a18 [--workers 16] [--overwrite]
4
+ Output per task in $HB_PI05_DATA/<task>/:
5
+ frames_<cam>.npy uint8 [N, 224, 224, 3] for the layout's two cameras; openpi_client.image_tools.resize_with_pad
6
+ (= OpenPI's ResizeImages(224, 224) transform, so the transform is a no-op in training)
7
+ lowdim.npz state [N,S], action [N,A] (raw), ep_end [N] (exclusive end row of the row's episode), episode_index, frame_index
8
+ meta.json provenance, layout, fps/control_hz, prompt, content box; assets/<layout>_<task>/norm_stats.json; READY
9
+ g1 (b6 basket a3 a7): state 44 = body29 by name | inspire L6 | inspire R6 | gravity3, action 76 = motion64 | L6 | R6
10
+ h2 (a18) : state 78 = observation.state75 | gravity3, action 75 = body_target31 | hands44
11
+ Why not OpenPI's LeRobotDataset path: the pinned lerobot (0cf8648, CODEBASE_VERSION v2.1) can read these v2.1 folders from
12
+ a local root, but openpi.training.data_loader.create_torch_dataset passes neither `root` nor `episodes` (needed for the B6
13
+ first-100 subset) and would fall back to the HF hub on any miss; and random-access h264 decoding per sample is slow.
14
+ Norm stats follow scripts/compute_norm_stats.py: RunningStats over the pre-Normalize "state" [N,S] and the "actions"
15
+ chunks [N,16,A] (same chunk/clamp as training), raw values (no augmentation). For h2 the actions stats are 75-D; the 76th
16
+ (padding) dim is added after Normalize by OpenPI's PadStatesAndActions, so it is exactly 0 and never normalized.
17
+ """
18
+ import argparse
19
+ import json
20
+ import os
21
+ import shutil
22
+ import sys
23
+ import time
24
+ from concurrent.futures import ProcessPoolExecutor
25
+ from pathlib import Path
26
+
27
+ import numpy as np
28
+ import pyarrow.parquet as pq
29
+
30
+ import hb_g1_data as D
31
+
32
+ COMMON = ["observation.projected_gravity", "timestamp", "frame_index", "episode_index", "task_index"]
33
+ FEATURES = { # feature -> required shape, per layout
34
+ "g1": {"action.motion_token": 64, "observation.inspire_left_hand_state": 6, "observation.inspire_right_hand_state": 6,
35
+ "observation.projected_gravity": 3},
36
+ "h2": {"observation.state": 75, "observation.projected_gravity": 3, "action.body_target": 31, "action.hands": 44},
37
+ }
38
+
39
+
40
+ def jl(path):
41
+ return [json.loads(x) for x in Path(path).read_text().splitlines() if x.strip()]
42
+
43
+
44
+ def col2d(table, name):
45
+ arr = table.column(name).combine_chunks()
46
+ vals = np.asarray(arr.flatten().to_numpy(zero_copy_only=False), dtype=np.float64)
47
+ return vals.reshape(len(arr), -1)
48
+
49
+
50
+ def feature_names(info, key):
51
+ names = info["features"][key].get("names")
52
+ if isinstance(names, dict):
53
+ names = next(iter(names.values()))
54
+ return list(names) if names else None
55
+
56
+
57
+ def state_names(info):
58
+ return feature_names(info, "observation.state")
59
+
60
+
61
+ def content_box(h, w, size=D.IMG):
62
+ """Same arithmetic as openpi_client.image_tools._resize_with_pad_pil -> (top, bottom, left, right)."""
63
+ ratio = max(w / size, h / size)
64
+ rh, rw = int(h / ratio), int(w / ratio)
65
+ top, left = max(0, int((size - rh) / 2)), max(0, int((size - rw) / 2))
66
+ return top, top + rh, left, left + rw
67
+
68
+
69
+ def read_episode(src, info, ep, layout):
70
+ """One episode -> raw state/action rows for the layout (training and verify_cpu use this)."""
71
+ rel = info["data_path"].format(episode_chunk=ep // int(info["chunks_size"]), episode_index=ep)
72
+ hand_diff = None
73
+ if layout == "g1":
74
+ cols = ["observation.state", "action.motion_token", "observation.inspire_left_hand_state",
75
+ "observation.inspire_right_hand_state", *COMMON]
76
+ t = pq.read_table(src / rel, columns=cols)
77
+ st = col2d(t, "observation.state")
78
+ body_idx = [state_names(info).index(n) for n in D.G1_BODY]
79
+ left, right = col2d(t, "observation.inspire_left_hand_state"), col2d(t, "observation.inspire_right_hand_state")
80
+ grav, motion = col2d(t, "observation.projected_gravity"), col2d(t, "action.motion_token")
81
+ state = np.concatenate([st[:, body_idx], left, right, grav], axis=1)
82
+ action = np.concatenate([motion, left, right], axis=1)
83
+ if st.shape[1] >= 41:
84
+ hand_diff = float(np.abs(st[:, 29:41] - np.concatenate([left, right], 1)).max())
85
+ elif layout == "h2":
86
+ t = pq.read_table(src / rel, columns=["observation.state", "action.body_target", "action.hands", *COMMON])
87
+ st = col2d(t, "observation.state")
88
+ state = np.concatenate([st, col2d(t, "observation.projected_gravity")], axis=1)
89
+ action = np.concatenate([col2d(t, "action.body_target"), col2d(t, "action.hands")], axis=1)
90
+ else:
91
+ raise ValueError(layout)
92
+ n = t.num_rows
93
+ L = D.LAYOUTS[layout]
94
+ assert state.shape == (n, L.state_dim) and action.shape == (n, L.action_dim), (ep, state.shape, action.shape)
95
+ fi = t.column("frame_index").to_numpy()
96
+ assert (fi == np.arange(n)).all(), f"episode {ep}: frame_index not 0..n-1"
97
+ assert (t.column("episode_index").to_numpy() == ep).all(), ep
98
+ return dict(state=state, action=action, ts=t.column("timestamp").to_numpy().astype(np.float64), fi=fi,
99
+ task_index=np.unique(t.column("task_index").to_numpy()), hand_diff=hand_diff, raw_state_dim=st.shape[1])
100
+
101
+
102
+ def decode_into(job):
103
+ """Decode one episode video sequentially (PyAV), pick the frame nearest to each parquet timestamp, resize_with_pad."""
104
+ import av
105
+ from openpi_client import image_tools
106
+
107
+ video, npy, start, ts, fps, hw = job
108
+ frames, times = [], []
109
+ with av.open(str(video)) as container:
110
+ stream = container.streams.video[0]
111
+ stream.thread_type = "AUTO"
112
+ for f in container.decode(stream):
113
+ arr = f.to_ndarray(format="rgb24")
114
+ if arr.shape[:2] != tuple(hw):
115
+ raise ValueError(f"{video}: frame {arr.shape} != {hw}")
116
+ frames.append(image_tools.resize_with_pad(arr[None], D.IMG, D.IMG)[0])
117
+ times.append(float(f.pts * f.time_base))
118
+ times = np.asarray(times)
119
+ j = np.clip(np.searchsorted(times, ts), 1, len(times) - 1)
120
+ j = np.where(np.abs(times[j - 1] - ts) <= np.abs(times[j] - ts), j - 1, j)
121
+ err = float(np.abs(times[j] - ts).max())
122
+ if err > 0.5 / fps:
123
+ raise ValueError(f"{video}: timestamp mismatch {err:.4f}s (decoded {len(times)} frames, need {len(ts)})")
124
+ out = np.load(npy, mmap_mode="r+")
125
+ out[start:start + len(ts)] = np.stack(frames)[j]
126
+ out.flush()
127
+ return len(times), len(ts), err
128
+
129
+
130
+ def guard_quantiles(stats, raw, key, log):
131
+ """pi0.5 uses quantile norm (q01->-1, q99->+1, no clipping). If a dim's q01..q99 range is degenerate relative to its
132
+ full range (e.g. a hand that moves in <1% of frames), outliers would map to huge values; use min/max for that dim."""
133
+ flat = raw.reshape(-1, raw.shape[-1])
134
+ lo, hi = flat.min(0), flat.max(0)
135
+ q01, q99 = np.asarray(stats.q01, np.float64).copy(), np.asarray(stats.q99, np.float64).copy()
136
+ bad = (q99 - q01) < 0.05 * (hi - lo)
137
+ q01[bad], q99[bad] = lo[bad], hi[bad]
138
+ log[key] = {"dims_minmax_fallback": np.flatnonzero(bad).tolist(),
139
+ "constant_dims": np.flatnonzero((hi - lo) < 1e-9).tolist()}
140
+ stats.q01, stats.q99 = q01, q99
141
+ return stats
142
+
143
+
144
+ def check_features(task, info, layout):
145
+ feats = info["features"]
146
+ for key, dim in FEATURES[layout].items():
147
+ assert key in feats and list(feats[key]["shape"]) == [dim], (task, key, feats.get(key, {}).get("shape"))
148
+ notes = {}
149
+ if layout == "g1":
150
+ missing = [n for n in D.G1_BODY if n not in (state_names(info) or [])]
151
+ assert not missing, (task, "observation.state names lack", missing)
152
+ else: # h2: action.body_target / action.hands are documented to carry the names of state[0:31] / state[31:75]
153
+ sn = state_names(info)
154
+ for key, sl in (("action.body_target", slice(0, 31)), ("action.hands", slice(31, 75))):
155
+ an = feature_names(info, key)
156
+ if sn and an and list(an) != list(sn[sl]):
157
+ diff = [(a, b) for a, b in zip(an, sn[sl], strict=False) if a != b][:5]
158
+ print(f"WARNING {task}: {key} names differ from observation.state[{sl.start}:{sl.stop}] e.g. {diff}")
159
+ notes[key] = diff
160
+ return notes
161
+
162
+
163
+ def prepare(task, *, out_root=None, max_episodes=None, check_expected=True, workers=16, overwrite=False):
164
+ import openpi.shared.normalize as normalize # lightweight (pydantic/numpy)
165
+
166
+ spec = D.TASKS[task]
167
+ src, layout = spec.root, spec.layout
168
+ L = D.LAYOUTS[layout]
169
+ out_root = Path(out_root or D.data_root())
170
+ dst = out_root / task
171
+ if (dst / "READY").exists() and not overwrite:
172
+ print(f"{task}: {dst} already READY (use --overwrite)")
173
+ return dst
174
+ info = json.loads((src / "meta/info.json").read_text())
175
+ assert str(info.get("codebase_version", "")).startswith("v2"), info.get("codebase_version")
176
+ fps = float(info["fps"])
177
+ name_notes = check_features(task, info, layout)
178
+ feats = info["features"]
179
+ hw = {}
180
+ for cam in L.cameras:
181
+ f = feats[f"observation.images.{cam}"]
182
+ assert f["dtype"] == "video", (cam, f["dtype"])
183
+ shp = list(f["shape"])
184
+ hw[cam] = (shp[0], shp[1]) if shp[-1] == 3 else (shp[1], shp[2])
185
+ episodes = jl(src / "meta/episodes.jsonl")
186
+ n_eps = len(episodes) if spec.limit is None else spec.limit
187
+ if max_episodes:
188
+ n_eps = min(n_eps, max_episodes)
189
+ episodes = episodes[:n_eps]
190
+ assert [e["episode_index"] for e in episodes] == list(range(n_eps)), "episodes.jsonl not 0..n-1"
191
+ if check_expected:
192
+ assert n_eps == spec.episodes, (task, n_eps, spec.episodes)
193
+ tasks = {r["task_index"]: r["task"] for r in jl(src / "meta/tasks.jsonl")}
194
+
195
+ rows, task_ids, hand_diff, raw_dim = [], set(), [], None
196
+ for e in episodes:
197
+ r = read_episode(src, info, e["episode_index"], layout)
198
+ assert len(r["state"]) == int(e["length"]), (e, len(r["state"]))
199
+ rows.append(r)
200
+ task_ids.update(int(x) for x in r["task_index"])
201
+ hand_diff.append(r["hand_diff"])
202
+ raw_dim = r["raw_state_dim"]
203
+ assert len(task_ids) == 1, f"expected a single task, got {task_ids}"
204
+ prompt = tasks[next(iter(task_ids))]
205
+ lengths = np.array([len(r["state"]) for r in rows])
206
+ starts = np.concatenate([[0], np.cumsum(lengths)[:-1]])
207
+ n = int(lengths.sum())
208
+ if check_expected and spec.frames is not None:
209
+ assert n == spec.frames, (task, n, spec.frames)
210
+ state = np.concatenate([r["state"] for r in rows]).astype(np.float32)
211
+ action = np.concatenate([r["action"] for r in rows]).astype(np.float32)
212
+ ep_end = np.repeat(starts + lengths, lengths)
213
+ assert np.isfinite(state).all() and np.isfinite(action).all()
214
+
215
+ tmp = out_root / f".{task}.tmp-{os.getpid()}"
216
+ shutil.rmtree(tmp, ignore_errors=True)
217
+ tmp.mkdir(parents=True)
218
+ jobs = []
219
+ for cam in L.cameras:
220
+ npy = tmp / f"frames_{cam}.npy"
221
+ np.lib.format.open_memmap(npy, mode="w+", dtype=np.uint8, shape=(n, D.IMG, D.IMG, 3)).flush()
222
+ for e, s, r in zip(episodes, starts, rows, strict=True):
223
+ ep = e["episode_index"]
224
+ rel = info["video_path"].format(episode_chunk=ep // int(info["chunks_size"]),
225
+ video_key=f"observation.images.{cam}", episode_index=ep)
226
+ jobs.append((src / rel, npy, int(s), r["ts"], fps, hw[cam]))
227
+ t0 = time.time()
228
+ with ProcessPoolExecutor(max_workers=workers) as ex:
229
+ res = list(ex.map(decode_into, jobs, chunksize=1))
230
+ decoded_minus_rows = sorted({a - b for a, b, _ in res})
231
+ print(f"{task}: decoded {len(jobs)} videos in {time.time() - t0:.0f}s; decoded-rows diffs {decoded_minus_rows}; "
232
+ f"max |dt| {max(x[2] for x in res):.4f}s")
233
+
234
+ np.savez(tmp / "lowdim.npz", state=state, action=action, ep_end=ep_end,
235
+ episode_index=np.repeat(np.arange(n_eps), lengths), frame_index=np.concatenate([r["fi"] for r in rows]))
236
+ chunk_idx = np.minimum(np.arange(n)[:, None] + np.arange(D.HORIZON)[None], ep_end[:, None] - 1)
237
+ chunks = action[chunk_idx] # [N, 16, A] exactly as the training sampler builds it
238
+ norm_log = {}
239
+ stats = {}
240
+ for key, arr in (("state", state), ("actions", chunks)):
241
+ rs = normalize.RunningStats()
242
+ rs.update(arr)
243
+ stats[key] = guard_quantiles(rs.get_statistics(), arr, key, norm_log)
244
+ normalize.save(tmp / "assets" / D.asset_id(task), stats)
245
+ cam_hw = hw[L.cameras[0]]
246
+ assert all(hw[c] == cam_hw for c in L.cameras), hw
247
+ meta = dict(task=task, layout=layout, source=str(src), codebase_version=info.get("codebase_version"),
248
+ fps=fps, control_hz=spec.control_hz,
249
+ control_hz_note="assumed controller rate (hb_g1_data.TASKS); state delay = round(k*fps/control_hz) rows, k in {1,2}",
250
+ episodes=n_eps, frames=n, episode_lengths=lengths.tolist(), prompt=prompt, raw_state_dim=raw_dim,
251
+ max_abs_diff_state29_41_vs_inspire=max([h for h in hand_diff if h is not None], default=None),
252
+ name_notes=name_notes, state_layout=L.state_desc, action_layout=L.action_desc + ", rows t..t+15",
253
+ cameras={"base_0_rgb": f"observation.images.{L.cameras[0]}",
254
+ "left_wrist_0_rgb": f"observation.images.{L.cameras[1]}",
255
+ "right_wrist_0_rgb": "zeros, image_mask False"},
256
+ raw_hw=list(cam_hw), content_box=list(content_box(*cam_hw)), norm_guard=norm_log,
257
+ created=time.strftime("%Y-%m-%dT%H:%M:%SZ", time.gmtime()))
258
+ (tmp / "meta.json").write_text(json.dumps(meta, indent=2))
259
+ (tmp / "READY").write_text("ok\n")
260
+ if dst.exists():
261
+ shutil.rmtree(dst)
262
+ tmp.rename(dst)
263
+ print(f"PREPARE_OK {task} ({layout}): {n_eps} episodes, {n} frames, fps {fps}, prompt={prompt!r}, norm guard={norm_log}")
264
+ return dst
265
+
266
+
267
+ def main():
268
+ ap = argparse.ArgumentParser()
269
+ ap.add_argument("tasks", nargs="+", choices=sorted(D.TASKS))
270
+ ap.add_argument("--workers", type=int, default=16)
271
+ ap.add_argument("--overwrite", action="store_true")
272
+ a = ap.parse_args()
273
+ for t in a.tasks:
274
+ prepare(t, workers=a.workers, overwrite=a.overwrite)
275
+
276
+
277
+ if __name__ == "__main__":
278
+ sys.exit(main())
code/hb_pi05/hb_train.py ADDED
@@ -0,0 +1,153 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """torchrun entry point: OpenPI scripts/train_pytorch.py:train_loop, unmodified, with G1 hooks installed.
2
+
3
+ Hooks (module attributes that OpenPI looks up at call time; no OpenPI file is edited):
4
+ openpi.training.data_loader.create_torch_dataset -> hb_g1_data.HBCachedDataset + HB aug v2 (training only)
5
+ models_pytorch.preprocessing_pytorch.preprocess_observation_pytorch -> train=False (OpenPI crop/rotate/jitter off)
6
+ torch DistributedSampler -> advances its epoch on every pass (DataLoaderImpl has no set_epoch, so upstream would
7
+ replay the same shuffle every epoch) and encodes the pass into the index (index + pass*N)
8
+ train_pytorch.set_seed -> seeded by global rank (upstream uses local rank: identical noise/time draws across nodes)
9
+ train_pytorch.save_checkpoint -> saves only at the requested steps (epoch 10 and the last step), then barrier;
10
+ older checkpoints lose optimizer.pt (resume state only for the latest one)
11
+ Budget: num_train_steps = ceil(epochs * frames / (per_gpu_batch * world_size)); frames = rows of the task's training view.
12
+ Probe mode (--probe-dir): run a few steps without base weights, write ok_rank<r>.json / oom_rank<r>.json markers.
13
+ """
14
+ import argparse
15
+ import dataclasses
16
+ import datetime
17
+ import importlib.util
18
+ import json
19
+ import math
20
+ import os
21
+ import sys
22
+ import time
23
+ from pathlib import Path
24
+
25
+ import torch
26
+ import torch.distributed as dist
27
+
28
+ sys.path.insert(0, str(Path(__file__).resolve().parent))
29
+ import hb_g1_config as C # noqa: E402 (registers the G1 configs)
30
+ import hb_g1_data as D # noqa: E402
31
+
32
+
33
+ class _EpochAdvancingSampler(torch.utils.data.distributed.DistributedSampler):
34
+ """Reshuffles every pass, and passes the pass number to the dataset as index + pass * N (HBCachedDataset decodes
35
+ it with divmod) so the per-trajectory joint bias can be redrawn per pass identically on all ranks/workers."""
36
+
37
+ def __iter__(self):
38
+ epoch, n = self.epoch, len(self.dataset)
39
+ indices = list(super().__iter__())
40
+ self.set_epoch(epoch + 1)
41
+ return iter([i + epoch * n for i in indices])
42
+
43
+
44
+ def load_train_pytorch():
45
+ path = Path(os.environ["HB_OPENPI"]) / "scripts/train_pytorch.py"
46
+ spec = importlib.util.spec_from_file_location("openpi_train_pytorch", path)
47
+ mod = importlib.util.module_from_spec(spec)
48
+ sys.modules[spec.name] = mod
49
+ spec.loader.exec_module(mod)
50
+ return mod
51
+
52
+
53
+ def is_oom(e: BaseException) -> bool:
54
+ return isinstance(e, torch.cuda.OutOfMemoryError) or "out of memory" in str(e).lower()
55
+
56
+
57
+ def marker(d: str, name: str, payload: dict):
58
+ Path(d).mkdir(parents=True, exist_ok=True)
59
+ tmp = Path(d) / f".{name}.tmp"
60
+ tmp.write_text(json.dumps(payload))
61
+ tmp.rename(Path(d) / name)
62
+
63
+
64
+ def main():
65
+ ap = argparse.ArgumentParser()
66
+ ap.add_argument("--task", required=True, choices=sorted(D.TASKS))
67
+ ap.add_argument("--exp-name", required=True)
68
+ ap.add_argument("--per-gpu-batch", type=int, required=True)
69
+ ap.add_argument("--epochs", type=float, default=float(os.environ.get("HB_EPOCHS", "20")))
70
+ ap.add_argument("--save-epochs", default=os.environ.get("HB_SAVE_EPOCHS", "10"), help="comma list; last step always saved")
71
+ ap.add_argument("--num-train-steps", type=int, default=None, help="override (probe)")
72
+ ap.add_argument("--ckpt-base", default=None)
73
+ ap.add_argument("--num-workers", type=int, default=int(os.environ.get("HB_NUM_WORKERS", "8")))
74
+ ap.add_argument("--resume", action="store_true")
75
+ ap.add_argument("--no-aug", action="store_true")
76
+ ap.add_argument("--probe-dir", default=None)
77
+ ap.add_argument("--probe-max-frac", type=float, default=float(os.environ.get("HB_PROBE_MAX_FRAC", "0.95")))
78
+ ap.add_argument("--nccl-timeout-min", type=int, default=60)
79
+ a = ap.parse_args()
80
+
81
+ rank, world = int(os.environ.get("RANK", "0")), int(os.environ.get("WORLD_SIZE", "1"))
82
+ local = int(os.environ.get("LOCAL_RANK", "0"))
83
+ tp = load_train_pytorch()
84
+ tp.init_logging()
85
+
86
+ D.install_hooks(augment=not a.no_aug)
87
+ torch.utils.data.distributed.DistributedSampler = _EpochAdvancingSampler
88
+ orig_seed = tp.set_seed
89
+ tp.set_seed = lambda seed, local_rank: orig_seed(seed, rank) # noqa: ARG005
90
+
91
+ frames = json.loads((D.task_dir(a.task) / "meta.json").read_text())["frames"]
92
+ global_batch = a.per_gpu_batch * world
93
+ steps = a.num_train_steps or math.ceil(a.epochs * frames / global_batch)
94
+ save_steps = set() if a.probe_dir else {steps} | {
95
+ math.ceil(float(e) * frames / global_batch) for e in a.save_epochs.split(",") if e.strip()}
96
+ orig_save = tp.save_checkpoint
97
+
98
+ def save_checkpoint(model, optimizer, global_step, config, is_main, data_config):
99
+ if global_step in save_steps:
100
+ # upstream condition `global_step % save_interval == 0` is made true for exactly this step
101
+ orig_save(model, optimizer, global_step, dataclasses.replace(config, save_interval=global_step), is_main,
102
+ data_config)
103
+ if is_main: # resume state only for the latest checkpoint: older steps keep weights, lose optimizer.pt
104
+ for d in Path(config.checkpoint_dir).iterdir():
105
+ if d.is_dir() and d.name.isdigit() and int(d.name) != global_step and (d / "optimizer.pt").exists():
106
+ (d / "optimizer.pt").unlink()
107
+ print(f"[hb] dropped resume state {d}/optimizer.pt", flush=True)
108
+ if dist.is_initialized():
109
+ dist.barrier()
110
+
111
+ tp.save_checkpoint = save_checkpoint
112
+
113
+ cfg = C.make_train_config(a.task, exp_name=a.exp_name, batch_size=global_batch, num_train_steps=steps,
114
+ num_workers=a.num_workers, checkpoint_base_dir=a.ckpt_base, resume=a.resume,
115
+ pytorch_weight_path=None if a.probe_dir else "default",
116
+ log_interval=1 if a.probe_dir else 10)
117
+ plan = dict(task=a.task, layout=D.TASKS[a.task].layout, config=cfg.name, exp_name=a.exp_name, frames=frames, world_size=world, per_gpu_batch=a.per_gpu_batch,
118
+ global_batch=global_batch, epochs=a.epochs, num_train_steps=steps, save_steps=sorted(save_steps),
119
+ warmup_steps=cfg.lr_schedule.warmup_steps, peak_lr=cfg.lr_schedule.peak_lr,
120
+ weights=cfg.pytorch_weight_path, checkpoint_dir=str(cfg.checkpoint_dir), augment=not a.no_aug,
121
+ aug_version=D.aug.VERSION, max_token_len=cfg.model.max_token_len)
122
+ if rank == 0:
123
+ print("HB_PLAN", json.dumps(plan), flush=True)
124
+ if not a.probe_dir:
125
+ cfg.checkpoint_dir.mkdir(parents=True, exist_ok=True)
126
+ (cfg.checkpoint_dir / "hb_plan.json").write_text(json.dumps(plan, indent=2))
127
+ if not a.probe_dir and cfg.pytorch_weight_path and not Path(cfg.pytorch_weight_path, "model.safetensors").exists():
128
+ sys.exit(f"missing init weights {cfg.pytorch_weight_path}/model.safetensors (run convert_base.sh)")
129
+
130
+ if world > 1 and not dist.is_initialized(): # train_pytorch.setup_ddp skips init if already initialized
131
+ torch.cuda.set_device(local)
132
+ dist.init_process_group("nccl", timeout=datetime.timedelta(minutes=a.nccl_timeout_min))
133
+
134
+ t0 = time.time()
135
+ try:
136
+ tp.train_loop(cfg)
137
+ except BaseException as e:
138
+ if a.probe_dir and is_oom(e):
139
+ marker(a.probe_dir, f"oom_rank{rank}.json", {"error": str(e)[:500]})
140
+ raise
141
+ if a.probe_dir:
142
+ peak = torch.cuda.max_memory_reserved(local)
143
+ total = torch.cuda.get_device_properties(local).total_memory
144
+ name = "ok" if peak / total <= a.probe_max_frac else "oom"
145
+ marker(a.probe_dir, f"{name}_rank{rank}.json", {"per_gpu_batch": a.per_gpu_batch, "peak_reserved_gb": peak / 1e9,
146
+ "total_gb": total / 1e9, "seconds": time.time() - t0,
147
+ "steps": steps})
148
+ elif rank == 0:
149
+ (cfg.checkpoint_dir / "HB_DONE").write_text(json.dumps({"seconds": time.time() - t0, **plan}))
150
+
151
+
152
+ if __name__ == "__main__":
153
+ main()
code/hb_pi05/verify_cpu.py ADDED
@@ -0,0 +1,222 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """CPU-only checks of the HB pi0.5 pipeline for one task (no GPU needed). Tasks: b6 basket a3 a7 (g1), a18 (h2).
2
+
3
+ python verify_cpu.py --task a18 # uses the prepared view $HB_PI05_DATA/a18 (hb_prepare.py a18)
4
+ python verify_cpu.py --task a18 --mini # first builds a 2-episode view in a temp dir (no full prep needed)
5
+ python verify_cpu.py --task b6 --forward # + full-size PI0Pytorch fp32 forward on CPU with $HB_PI05_INIT (~35 GB RAM)
6
+ Checks: raw sample vs parquet, action chunk/clamp, visual augmentation (images changed, letterbox black, labels
7
+ untouched), proprio augmentation (delay <= 2 control periods within the episode, per-trajectory bias, per-sample noise,
8
+ only the n_body joints touched, delay rate ~15%), norm stats dims, OpenPI transform chain (image slots/masks, state
9
+ tokens not truncated for this layout's max_token_len, actions padded to 76 with zeros for h2), Unnormalize + HBOutputs
10
+ back to the raw action dim, torch loader batch -> Observation, OpenPI image aug disabled, PI0Pytorch structure on the
11
+ meta device, init checkpoint header shapes.
12
+ """
13
+ import argparse
14
+ import json
15
+ import math
16
+ import os
17
+ import sys
18
+ import tempfile
19
+ from pathlib import Path
20
+
21
+ import numpy as np
22
+
23
+ sys.path.insert(0, str(Path(__file__).resolve().parent))
24
+ FAILS = []
25
+
26
+
27
+ def check(cond, msg):
28
+ print(("PASS " if cond else "FAIL ") + msg, flush=True)
29
+ if not cond:
30
+ FAILS.append(msg)
31
+
32
+
33
+ def main():
34
+ ap = argparse.ArgumentParser()
35
+ ap.add_argument("--task", default="b6")
36
+ ap.add_argument("--mini", action="store_true")
37
+ ap.add_argument("--forward", action="store_true")
38
+ ap.add_argument("--samples", type=int, default=200)
39
+ a = ap.parse_args()
40
+ if a.mini:
41
+ os.environ["HB_PI05_DATA"] = tempfile.mkdtemp(prefix="hb_pi05_mini_")
42
+ import torch
43
+
44
+ import hb_g1_data as D
45
+ import hb_prepare
46
+
47
+ if a.mini:
48
+ hb_prepare.prepare(a.task, max_episodes=2, check_expected=False, workers=4)
49
+ import hb_g1_config as C # noqa: E402
50
+ import openpi.models.model as _model
51
+ import openpi.training.config as _config
52
+ import openpi.training.data_loader as _data
53
+ import openpi.transforms as _transforms
54
+
55
+ tok = Path(os.environ.get("OPENPI_DATA_HOME", "~/.cache/openpi")).expanduser() / "big_vision/paligemma_tokenizer.model"
56
+ check(tok.exists(), f"tokenizer present at {tok}")
57
+ rng = np.random.default_rng(0)
58
+ L = D.layout_of(a.task)
59
+ cam0, cam1 = (f"observation/{c}" for c in L.cameras)
60
+ print(f"task {a.task}: layout {L.name} state {L.state_dim} actions {L.action_dim} cams {L.cameras} "
61
+ f"n_body {L.n_body} max_token_len {L.max_token_len}")
62
+
63
+ # 1. raw sample vs parquet
64
+ ds = D.HBCachedDataset(a.task, augment=False)
65
+ n = len(ds)
66
+ s = ds[5]
67
+ check(s[cam0].shape == (224, 224, 3) and s[cam0].dtype == np.uint8 and s[cam1].shape == (224, 224, 3),
68
+ f"{L.cameras} uint8 224x224x3")
69
+ check(s["observation/state"].shape == (L.state_dim,) and s["actions"].shape == (16, L.action_dim),
70
+ f"state {s['observation/state'].shape} actions {s['actions'].shape}")
71
+ check(isinstance(s["prompt"], str) and len(s["prompt"]) > 0, f"prompt {s['prompt']!r}")
72
+ src = D.TASKS[a.task].root
73
+ info = json.loads((src / "meta/info.json").read_text())
74
+ ep0 = hb_prepare.read_episode(src, info, 0, L.name)
75
+ check(np.allclose(s["observation/state"], ep0["state"][5], atol=1e-6), f"state[5] == parquet ({L.state_desc})")
76
+ check(np.allclose(s["actions"][0], ep0["action"][5], atol=1e-6) and np.allclose(s["actions"][15], ep0["action"][20], atol=1e-6),
77
+ f"actions rows t..t+15 == parquet ({L.action_desc})")
78
+ ln = len(ep0["state"])
79
+ tail = ds[ln - 3]
80
+ check(np.allclose(tail["actions"][2:], ep0["action"][ln - 1], atol=1e-6), "chunk clamped to the episode's last row")
81
+ q = ds.state[:, :L.n_body]
82
+ check(np.abs(q).max() < 2 * math.pi, f"perturbed joints look like radians (max |q| {np.abs(q).max():.2f})")
83
+ print(f" fps {ds.fps} control_hz {ds.control_hz} -> delay rows for 1/2 periods: "
84
+ f"{max(1, round(ds.fps / ds.control_hz))}/{max(1, round(2 * ds.fps / ds.control_hz))}")
85
+
86
+ # 2. visual augmentation (+ labels untouched)
87
+ dsa = D.HBCachedDataset(a.task, augment=True)
88
+ top, bottom, _, _ = ds.box
89
+ changed, pad_black, act_ok = 0, True, True
90
+ for i in rng.integers(0, n, 16):
91
+ r, g = ds[int(i)], dsa[int(i)]
92
+ changed += int(not np.array_equal(r[cam0], g[cam0]))
93
+ for k in (cam0, cam1):
94
+ pad_black &= bool(g[k][:top].max(initial=0) == 0 and g[k][bottom:].max(initial=0) == 0)
95
+ act_ok &= bool(np.array_equal(g["actions"], r["actions"]))
96
+ check(changed >= 12, f"augmentation changes images ({changed}/16)")
97
+ check(pad_black, f"letterbox rows stay black (content rows {top}:{bottom})")
98
+ check(act_ok, "action labels untouched by augmentation")
99
+
100
+ # 3. proprio augmentation (CLAUDE.md sec. 3): delay, per-trajectory bias, per-sample noise
101
+ dsa._lazy_init() # noqa: SLF001
102
+ nb, max_d = L.n_body, max(1, round(2 * ds.fps / ds.control_hz))
103
+ bound = math.radians(D.BIAS_DEG + 3 * D.NOISE_DEG) + 1e-5
104
+ delays, ok = [], True
105
+ for i in rng.integers(0, n, 400):
106
+ i = int(i)
107
+ st, d = dsa.proprio(i, epoch=0)
108
+ j = i - d
109
+ ok &= bool(0 <= d <= max_d and j >= dsa.ep_start[i] and np.array_equal(st[nb:], ds.state[j][nb:])
110
+ and np.abs(st[:nb] - ds.state[j][:nb]).max() <= bound)
111
+ delays.append(d)
112
+ rate = np.mean(np.asarray(delays) > 0)
113
+ check(ok, f"delay <= {max_d} rows within the episode; only state[:{nb}] perturbed within 2 deg + 3 sigma")
114
+ check(0.08 <= rate <= 0.23, f"delayed fraction {rate:.3f} ~ 0.15 (clamped delays at episode start count as 0)")
115
+ ep = int(ds.episode[0])
116
+ b1, b2 = D.trajectory_bias(42, 0, ep, nb), D.trajectory_bias(42, 0, ep, nb)
117
+ check(np.array_equal(b1, b2) and np.abs(b1).max() <= math.radians(D.BIAS_DEG), "bias deterministic per (seed, pass, episode), |b| <= 2 deg")
118
+ check(not np.array_equal(b1, D.trajectory_bias(42, 1, ep, nb)) and not np.array_equal(b1, D.trajectory_bias(42, 0, ep + 1, nb)),
119
+ "bias redrawn per pass and per episode")
120
+ # same episode, same pass, no delay: difference between two samples' offsets is noise only (<= 2 * 3 sigma)
121
+ same = [k for k in range(n) if ds.episode[k] == ep][:50]
122
+ offs = []
123
+ for k in same:
124
+ st, d = dsa.proprio(k, epoch=0)
125
+ if d == 0:
126
+ offs.append(st[:nb] - ds.state[k][:nb])
127
+ spread = np.ptp(np.stack(offs), axis=0).max() if len(offs) > 1 else 0.0
128
+ check(spread <= 2 * math.radians(3 * D.NOISE_DEG) + 1e-5, f"bias fixed within the episode (offset spread {math.degrees(spread):.2f} deg)")
129
+ check(dsa[n + 5]["actions"].shape == (16, L.action_dim), "index + pass*N decodes to row 5 of pass 1")
130
+
131
+ # 4. config + norm stats
132
+ cfg = C.make_train_config(a.task, batch_size=2, num_workers=0, num_train_steps=10)
133
+ check(_config.get_config(D.config_name(a.task)).name == cfg.name, f"config {cfg.name} registered")
134
+ m = cfg.model
135
+ check(m.pi05 and m.action_dim == 76 and m.action_horizon == 16 and m.discrete_state_input
136
+ and m.max_token_len == L.max_token_len, f"model {m}")
137
+ dc = cfg.data.create(cfg.assets_dirs, cfg.model)
138
+ check(dc.use_quantile_norm, "pi05 uses quantile norm")
139
+ ns = dc.norm_stats or {}
140
+ check(set(ns) == {"state", "actions"} and np.asarray(ns["state"].q01).shape == (L.state_dim,)
141
+ and np.asarray(ns["actions"].q99).shape == (L.action_dim,),
142
+ f"norm stats state[{L.state_dim}] actions[{L.action_dim}] (raw dims, no pad dim) with q01/q99")
143
+
144
+ # 5. OpenPI transform chain
145
+ tds = _data.transform_dataset(dsa, dc)
146
+ t = tds[7]
147
+ check(set(t["image"]) == set(_model.IMAGE_KEYS), f"image slots {sorted(t['image'])}")
148
+ check(t["image"]["base_0_rgb"].shape == (224, 224, 3) and t["image"]["base_0_rgb"].dtype == np.uint8, "base_0_rgb uint8 224")
149
+ check(bool(t["image_mask"]["base_0_rgb"]) and bool(t["image_mask"]["left_wrist_0_rgb"])
150
+ and not bool(t["image_mask"]["right_wrist_0_rgb"]), "masks base/left True, right False")
151
+ sw = max(L.state_dim, 76)
152
+ check(t["state"].shape == (sw,) and np.all(t["state"][L.state_dim:] == 0), f"state after pad ({sw},)")
153
+ check(t["actions"].shape == (16, 76) and np.all(t["actions"][:, L.action_dim:] == 0),
154
+ f"actions (16,76); dims {L.action_dim}: are zero padding")
155
+ lens = [int(tds[int(i)]["tokenized_prompt_mask"].sum()) for i in rng.integers(0, n, a.samples)]
156
+ check(max(lens) < m.max_token_len, f"prompt+state tokens max {max(lens)} < max_token_len {m.max_token_len} (mean {np.mean(lens):.0f})")
157
+ from openpi.models.tokenizer import PaligemmaTokenizer
158
+
159
+ sp = PaligemmaTokenizer(m.max_token_len)._tokenizer # noqa: SLF001
160
+ print(" decoded prompt:", sp.decode([int(x) for x, k in zip(t["tokenized_prompt"], t["tokenized_prompt_mask"]) if k]))
161
+ out = _transforms.compose([_transforms.Unnormalize(dc.norm_stats, use_quantiles=True), *dc.data_transforms.outputs])(
162
+ {"actions": np.asarray(t["actions"], np.float64), "state": np.asarray(t["state"])})
163
+ check(out["actions"].shape == (16, L.action_dim), f"Unnormalize + HBOutputs -> (16,{L.action_dim})")
164
+ raw_rows = ds.action[np.minimum(np.arange(7, 23), ds.ep_end[7] - 1)]
165
+ check(np.allclose(out["actions"], raw_rows, atol=1e-3 * (1 + np.abs(raw_rows).max())), "normalize -> unnormalize round trip")
166
+
167
+ # 6. loader + OpenPI aug disabled
168
+ D.install_hooks(augment=True)
169
+ loader = _data.create_data_loader(cfg, framework="pytorch", shuffle=True, num_batches=1)
170
+ obs, act = next(iter(loader))
171
+ check(tuple(obs.images["base_0_rgb"].shape) == (2, 3, 224, 224) and obs.images["base_0_rgb"].dtype == torch.float32,
172
+ f"batch image {tuple(obs.images['base_0_rgb'].shape)} float [-1,1]")
173
+ check(tuple(obs.state.shape) == (2, sw) and tuple(act.shape) == (2, 16, 76), f"batch state (2,{sw}) actions (2,16,76)")
174
+ import openpi.models_pytorch.preprocessing_pytorch as pp
175
+
176
+ torch.manual_seed(0)
177
+ o1 = pp.preprocess_observation_pytorch(obs, train=True)
178
+ o2 = pp.preprocess_observation_pytorch(obs, train=False)
179
+ check(all(torch.equal(o1.images[k], o2.images[k]) for k in o1.images), "OpenPI train-time image aug disabled")
180
+
181
+ # 7. model structure (meta device) + init checkpoint header
182
+ import openpi.models_pytorch.pi0_pytorch as p0
183
+
184
+ mcfg = C.model_config(L.name, dtype="float32", pytorch_compile_mode=None)
185
+ with torch.device("meta"):
186
+ net = p0.PI0Pytorch(mcfg)
187
+ check(tuple(net.action_in_proj.weight.shape) == (1024, 76) and tuple(net.action_out_proj.weight.shape) == (76, 1024),
188
+ "PI0Pytorch action_in/out_proj sized for 76")
189
+ print(f" params: {sum(p.numel() for p in net.parameters()) / 1e9:.3f} B")
190
+ init = Path(os.environ.get("HB_PI05_INIT", "")) / "model.safetensors"
191
+ if init.exists():
192
+ from safetensors import safe_open
193
+
194
+ sd = net.state_dict()
195
+ groups = {}
196
+ for name, p in net.named_parameters(remove_duplicate=False):
197
+ groups.setdefault(id(p), []).append(name)
198
+ alias = {k: g for g in groups.values() for k in g}
199
+ with safe_open(str(init), "pt") as f:
200
+ keys = set(f.keys())
201
+ bad = [k for k in keys if k not in sd or tuple(f.get_slice(k).get_shape()) != tuple(sd[k].shape)]
202
+ missing = [k for k in sd if k not in keys and not any(x in keys for x in alias.get(k, []))]
203
+ check(not bad and not missing, f"init checkpoint matches model ({len(keys)} tensors; bad {bad[:5]} missing {missing[:5]})")
204
+ else:
205
+ print(f" (skip) {init} not found")
206
+
207
+ if a.forward:
208
+ import safetensors.torch
209
+
210
+ net = p0.PI0Pytorch(mcfg)
211
+ safetensors.torch.load_model(net, str(init))
212
+ net.eval()
213
+ with torch.no_grad():
214
+ loss = net(obs, act.to(torch.float32))
215
+ check(tuple(loss.shape) == (2, 16, 76) and torch.isfinite(loss).all().item(), f"fp32 CPU forward loss {loss.mean().item():.4f}")
216
+
217
+ print("VERIFY_" + ("OK" if not FAILS else f"FAILED ({len(FAILS)})"))
218
+ sys.exit(1 if FAILS else 0)
219
+
220
+
221
+ if __name__ == "__main__":
222
+ main()