File size: 59,093 Bytes
cfb943d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
diff --git a/src/openpi/models/gemma.py b/src/openpi/models/gemma.py
index d1623d1..364d8cb 100644
--- a/src/openpi/models/gemma.py
+++ b/src/openpi/models/gemma.py
@@ -1,3 +1,4 @@
+import os
 # Copyright 2024 Big Vision Authors.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
@@ -213,22 +214,26 @@ class Attention(nn.Module):
             k = jnp.concatenate([cache_k, k], axis=1)
             v = jnp.concatenate([cache_v, v], axis=1)
 
-        q = einops.rearrange(q, "B T (K G) H -> B T K G H", K=self.configs[0].num_kv_heads)
-        logits = jnp.einsum("BTKGH,BSKH->BKGTS", q, k, preferred_element_type=jnp.float32)
-
         if attn_mask.shape != (q.shape[0], 1, q.shape[1], k.shape[1]):
             raise ValueError(
                 f"Attention mask with shape {attn_mask.shape} but shapes for q and k are: {q.shape} and {k.shape}"
             )
 
-        # big_neg = jnp.finfo(logits.dtype).min
-        big_neg = -2.3819763e38  # See gemma/modules.py
-        masked_logits = jnp.where(attn_mask[:, :, None, :, :], logits, big_neg)
+        if os.environ.get("OPENPI_ATTN_IMPL") == "cudnn":
+            # Fused cuDNN attention (same math; q is already scaled above, so scale=1). Opt-in via env var.
+            encoded = jax.nn.dot_product_attention(q, k, v, mask=attn_mask, scale=1.0, implementation="cudnn")
+        else:
+            q = einops.rearrange(q, "B T (K G) H -> B T K G H", K=self.configs[0].num_kv_heads)
+            logits = jnp.einsum("BTKGH,BSKH->BKGTS", q, k, preferred_element_type=jnp.float32)
+
+            # big_neg = jnp.finfo(logits.dtype).min
+            big_neg = -2.3819763e38  # See gemma/modules.py
+            masked_logits = jnp.where(attn_mask[:, :, None, :, :], logits, big_neg)
 
-        probs = jax.nn.softmax(masked_logits, axis=-1).astype(dtype)
+            probs = jax.nn.softmax(masked_logits, axis=-1).astype(dtype)
 
-        encoded = jnp.einsum("BKGTS,BSKH->BTKGH", probs, v)
-        encoded = einops.rearrange(encoded, "B T K G H -> B T (K G) H")
+            encoded = jnp.einsum("BKGTS,BSKH->BTKGH", probs, v)
+            encoded = einops.rearrange(encoded, "B T K G H -> B T (K G) H")
 
         out = []
         start = 0
diff --git a/src/openpi/models/pi0.py b/src/openpi/models/pi0.py
index ae7c459..483b164 100644
--- a/src/openpi/models/pi0.py
+++ b/src/openpi/models/pi0.py
@@ -1,4 +1,5 @@
 import logging
+import os
 
 import einops
 import flax.nnx as nnx
@@ -110,7 +111,13 @@ class Pi0(_model.BaseModel):
         ar_mask = []
         tokens = []
         # embed images
+        # OPENPI_SKIP_IMAGE_KEYS: comma-separated image keys that are ALWAYS masked out in this dataset (e.g. the
+        # zero-filled right wrist view of MobileManipBench). Their tokens never contribute (input_mask=False removes
+        # them from attention and from the cumsum positions), so skipping them is exact and saves their compute.
+        skip = {k for k in os.environ.get("OPENPI_SKIP_IMAGE_KEYS", "").split(",") if k}
         for name in obs.images:
+            if name in skip:
+                continue
             image_tokens, _ = self.PaliGemma.img(obs.images[name], train=False)
 
             tokens.append(image_tokens)
diff --git a/src/openpi/training/config.py b/src/openpi/training/config.py
index 4ca47e1..eb9874c 100644
--- a/src/openpi/training/config.py
+++ b/src/openpi/training/config.py
@@ -20,6 +20,7 @@ import openpi.models.tokenizer as _tokenizer
 import openpi.policies.aloha_policy as aloha_policy
 import openpi.policies.droid_policy as droid_policy
 import openpi.policies.libero_policy as libero_policy
+import openpi.policies.mmb_policy as mmb_policy
 import openpi.shared.download as _download
 import openpi.shared.normalize as _normalize
 import openpi.training.droid_rlds_dataset as droid_rlds_dataset
@@ -86,6 +87,9 @@ class DataConfig:
     # sequence is defined by the `action_horizon` field in the model config. This should be adjusted if your
     # LeRobot dataset is using different keys to represent the action.
     action_sequence_keys: Sequence[str] = ("actions",)
+    # If set, this key is loaded for frames t .. t+action_horizon (action_horizon + 1 frames); used when the action
+    # chunk is derived from future states (MobileManipBench v2: actions = f(state(t), state(t+1 .. t+H))).
+    state_chunk_key: str | None = None
 
     # If true, will use the LeRobot dataset task to define the prompt.
     prompt_from_task: bool = False
@@ -355,6 +359,61 @@ class LeRobotLiberoDataConfig(DataConfigFactory):
         )
 
 
+@dataclasses.dataclass(frozen=True)
+class LeRobotMMBDataConfig(DataConfigFactory):
+    """MobileManipBench (all robots) for pi0.5 warm-up. Per-robot quantile normalisation happens in
+    mmb_policy.PerRobotNormalize; the global Normalize step receives identity statistics.
+    canonical_v2: Canonical Action Protocol v2 dataset (convert_merge.py --v2); chunks become canonical 12D actions."""
+
+    canonical_v2: bool = False
+    # v2 EEF action source: "state" = future measured EEF (state chunk), "target" = expert sparse targets
+    action_source: str = "state"
+    # v2 action rotation (relative to the current EEF): "rot6d" (14D action) or legacy "rpy" (11D: pi05_mmb_v2 / v2t)
+    action_rot: str = "rot6d"
+    # v2 EEF action reference: None = module default (rot6d: "base" = target pose in the body frame); "eef" = relative
+    action_ref: str | None = None
+
+    @override
+    def create(self, assets_dirs: pathlib.Path, model_config: _model.BaseModelConfig) -> DataConfig:
+        base = self.create_base_config(assets_dirs, model_config)
+        stats_dir = pathlib.Path(self.assets.assets_dir or assets_dirs) / (self.assets.asset_id or self.repo_id)
+        per_robot = mmb_policy.load_per_robot_stats(stats_dir / mmb_policy.PER_ROBOT_STATS_FILE)
+        repack_transform = _transforms.Group(
+            inputs=[
+                _transforms.RepackTransform(
+                    {
+                        "observation/image": "observation.images.head",
+                        "observation/wrist_image": "observation.images.wrist",
+                        # v2: canonical 14D state chunk t .. t+H (base frame, no world coordinates); actions derived from it
+                        "observation/state": "observation.state_canonical" if self.canonical_v2 else "observation.state",
+                        "robot_id": "robot_id",
+                        **({"commands": "action.command"} if self.canonical_v2 else {"actions": "actions"}),
+                        **({"targets": "action.eef_target"} if self.canonical_v2 and self.action_source == "target" else {}),
+                        "prompt": "prompt",
+                    }
+                )
+            ]
+        )
+        data_transforms = _transforms.Group(
+            inputs=[mmb_policy.MMBInputs(model_type=model_config.model_type),
+                    *([mmb_policy.MMBActionsFromTargets(rot=self.action_rot, ref=self.action_ref) if self.action_source == "target"
+                       else mmb_policy.MMBActionsFromStates(rot=self.action_rot, ref=self.action_ref)]
+                      if self.canonical_v2 else []),
+                    mmb_policy.PerRobotNormalize(per_robot)],
+        )
+        return dataclasses.replace(
+            base,
+            repack_transforms=repack_transform,
+            data_transforms=data_transforms,
+            model_transforms=ModelTransformFactory()(model_config),
+            # v2: no action column; the chunk is built from state(t .. t+H) of the canonical state column
+            **(({"action_sequence_keys": ("action.command", "action.eef_target"), "state_chunk_key": None}
+                if self.action_source == "target" else
+                {"action_sequence_keys": ("action.command",), "state_chunk_key": "observation.state_canonical"})
+               if self.canonical_v2 else {}),
+        )
+
+
 @dataclasses.dataclass(frozen=True)
 class RLDSDroidDataConfig(DataConfigFactory):
     """
@@ -740,6 +799,137 @@ _CONFIGS = [
         # Turn off EMA for LoRA finetuning.
         ema_decay=None,
     ),
+    TrainConfig(
+        # pi0.5 full fine-tuning "warm-up" on all of prlq/mobilemanipbench-data (rev 356ec43), 3 robots, 80k steps.
+        # Recipe = pi05_libero except: steps, batch (multiple of 6 GPUs), loader workers, checkpoint cadence.
+        name="pi05_mmb_warmup",
+        model=pi0_config.Pi0Config(pi05=True, action_horizon=10, discrete_state_input=False),
+        data=LeRobotMMBDataConfig(
+            repo_id="mmb/all",
+            base_config=DataConfig(prompt_from_task=True),
+        ),
+        batch_size=252,
+        lr_schedule=_optimizer.CosineDecaySchedule(
+            warmup_steps=10_000,
+            peak_lr=5e-5,
+            decay_steps=1_000_000,
+            decay_lr=5e-5,
+        ),
+        optimizer=_optimizer.AdamW(clip_gradient_norm=1.0),
+        ema_decay=0.999,
+        weight_loader=weight_loaders.CheckpointWeightLoader("gs://openpi-assets/checkpoints/pi05_base/params"),
+        num_train_steps=80_000,
+        num_workers=80,  # 48 saturated (~37 cores busy, occasional 3 s steps); node has ~70% CPU idle
+        save_interval=2000,  # ~45 min at 1.3 s/step (each 42 GB save stalls training ~1 min)
+        keep_period=10_000,
+        wandb_enabled=False,
+    ),
+    TrainConfig(
+        # Canonical Action Protocol v2 (docs/writing-block.md): same recipe as pi05_mmb_warmup, fresh from pi05_base,
+        # on the v2 dataset (raw chunks -> canonical 12D chunk-relative actions at load time, new per-robot stats).
+        name="pi05_mmb_v2",
+        # discrete_state_input=True: the canonical 12D state is discretised (256 bins) into the prompt, the native pi0.5
+        # way; with False (pi05_libero / pi05_mmb_warmup) pi0.5 ignores the state entirely.
+        model=pi0_config.Pi0Config(pi05=True, action_horizon=10, discrete_state_input=True),
+        data=LeRobotMMBDataConfig(
+            repo_id="mmb/v2",
+            base_config=DataConfig(prompt_from_task=True),
+            canonical_v2=True,
+            action_rot="rpy",
+        ),
+        batch_size=252,
+        lr_schedule=_optimizer.CosineDecaySchedule(
+            warmup_steps=10_000,
+            peak_lr=5e-5,
+            decay_steps=1_000_000,
+            decay_lr=5e-5,
+        ),
+        optimizer=_optimizer.AdamW(clip_gradient_norm=1.0),
+        ema_decay=0.999,
+        weight_loader=weight_loaders.CheckpointWeightLoader("gs://openpi-assets/checkpoints/pi05_base/params"),
+        num_train_steps=80_000,
+        num_workers=80,
+        save_interval=2000,
+        keep_period=10_000,
+        wandb_enabled=False,
+    ),
+    TrainConfig(
+        # v2 with the EEF action from the expert sparse targets (yiqi 11-13); otherwise identical to pi05_mmb_v2.
+        # Canonical Action Protocol v2 (docs/writing-block.md): same recipe as pi05_mmb_warmup, fresh from pi05_base,
+        # on the v2 dataset (raw chunks -> canonical 12D chunk-relative actions at load time, new per-robot stats).
+        name="pi05_mmb_v2t",
+        # discrete_state_input=True: the canonical 12D state is discretised (256 bins) into the prompt, the native pi0.5
+        # way; with False (pi05_libero / pi05_mmb_warmup) pi0.5 ignores the state entirely.
+        model=pi0_config.Pi0Config(pi05=True, action_horizon=10, discrete_state_input=True),
+        data=LeRobotMMBDataConfig(
+            repo_id="mmb/v2t",
+            base_config=DataConfig(prompt_from_task=True),
+            canonical_v2=True,
+            action_source="target",
+            action_rot="rpy",
+        ),
+        batch_size=252,
+        lr_schedule=_optimizer.CosineDecaySchedule(
+            warmup_steps=10_000,
+            peak_lr=5e-5,
+            decay_steps=1_000_000,
+            decay_lr=5e-5,
+        ),
+        optimizer=_optimizer.AdamW(clip_gradient_norm=1.0),
+        ema_decay=0.999,
+        weight_loader=weight_loaders.CheckpointWeightLoader("gs://openpi-assets/checkpoints/pi05_base/params"),
+        num_train_steps=80_000,
+        num_workers=80,
+        save_interval=2000,
+        keep_period=10_000,
+        wandb_enabled=False,
+    ),
+    TrainConfig(
+        # v3 (yiqi 11-13/14): EEF action = the expert sparse target pose seen from the body frame (the base projected to
+        # the floor, heading only - same reference on all robots), all rotations rot6d (first two columns) - the 14D
+        # action [base cmd 4 | target xyz 3 | target rot6d 6 | gripper cmd 1]; dataset merged_v2t (patch_frame.py
+        # --frame ground); same recipe as pi05_mmb_v2; stats assets/pi05_mmb_v3/mmb/v2t (compute_stats_v3.py: base +
+        # EEF xyz one q01/q99 set for all robots and for state and action; rot6d and gripper not normalised).
+        name="pi05_mmb_v3",
+        # discrete_state_input=True: the canonical 14D state is discretised (256 bins) into the prompt, the native pi0.5
+        # way; with False pi0.5 ignores the state entirely.
+        model=pi0_config.Pi0Config(pi05=True, action_horizon=10, discrete_state_input=True),
+        data=LeRobotMMBDataConfig(
+            repo_id="mmb/v2t",
+            base_config=DataConfig(prompt_from_task=True),
+            canonical_v2=True,
+            action_source="target",
+            action_rot="rot6d",
+        ),
+        batch_size=252,
+        lr_schedule=_optimizer.CosineDecaySchedule(
+            warmup_steps=10_000,
+            peak_lr=5e-5,
+            decay_steps=1_000_000,
+            decay_lr=5e-5,
+        ),
+        optimizer=_optimizer.AdamW(clip_gradient_norm=1.0),
+        ema_decay=0.999,
+        weight_loader=weight_loaders.CheckpointWeightLoader("gs://openpi-assets/checkpoints/pi05_base/params"),
+        num_train_steps=80_000,
+        num_workers=80,
+        save_interval=2000,
+        keep_period=10_000,
+        wandb_enabled=False,
+    ),
+    TrainConfig(
+        # Compute-only speed benchmark for pi05_mmb_warmup (same model/batch, fake data, random init).
+        name="pi05_mmb_bench",
+        model=pi0_config.Pi0Config(pi05=True, action_horizon=10, discrete_state_input=False),
+        data=FakeDataConfig(),
+        batch_size=252,
+        num_workers=16,
+        num_train_steps=45,
+        log_interval=10,
+        save_interval=100000,
+        overwrite=True,
+        wandb_enabled=False,
+    ),
     TrainConfig(
         name="pi05_libero",
         model=pi0_config.Pi0Config(pi05=True, action_horizon=10, discrete_state_input=False),
diff --git a/src/openpi/training/data_loader.py b/src/openpi/training/data_loader.py
index e2ee7dd..8a69450 100644
--- a/src/openpi/training/data_loader.py
+++ b/src/openpi/training/data_loader.py
@@ -141,8 +141,13 @@ def create_torch_dataset(
     dataset = lerobot_dataset.LeRobotDataset(
         data_config.repo_id,
         delta_timestamps={
-            key: [t / dataset_meta.fps for t in range(action_horizon)] for key in data_config.action_sequence_keys
+            **{key: [t / dataset_meta.fps for t in range(action_horizon)] for key in data_config.action_sequence_keys},
+            # actions derived from future states: frames t .. t+H of this key (see DataConfig.state_chunk_key)
+            **({data_config.state_chunk_key: [t / dataset_meta.fps for t in range(action_horizon + 1)]}
+               if data_config.state_chunk_key else {}),
         },
+        # B300 cluster: torchcodec has no FFmpeg shared libs here; PyAV bundles FFmpeg (with libdav1d for AV1).
+        video_backend=os.environ.get("OPENPI_VIDEO_BACKEND") or None,
     )
 
     if data_config.prompt_from_task:
diff --git a/src/openpi/policies/mmb_canonical_action.py b/src/openpi/policies/mmb_canonical_action.py
new file mode 100644
index 0000000..a2eaae4
--- /dev/null
+++ b/src/openpi/policies/mmb_canonical_action.py
@@ -0,0 +1,485 @@
+"""MobileManipBench canonical state and action chunks (v3, yiqi 11-13/14) - framework-agnostic, numpy only, shared
+by pi0.5, GR00T N1.7 and OpenWAM. The dataset (convert_merge.py --v2, patch_frame.py) stores per frame the canonical
+state (observation.state_canonical), the expert's EEF target seen from the body (action.eef_target) and the recorded
+commands (action.command); the data loader only builds chunks (action_chunk_from_targets) and normalises.
+
+Body frame (BODY_FRAME = "ground", same reference on all robots): origin = the measured base position projected to the
+floor (x_B, y_B, 0), z up, x = base heading (yaw only; Go2 gait roll / pitch and height are not part of the frame).
+Raw recordings (world frame, quaternions w-x-y-z): observation.eef_pose = measured EEF pose; base pose = measured, from
+sim_info/0/sim_states.npz (base_pose_from_sim_states: npz row sim_step_idx - 1; constant asset-frame offsets removed,
+e.g. Sawyer N051/N053/N058/N059 Ry(-90 deg)).
+
+STATE, 14D (canonical_state):  [vx, vy, wz, height | x, y, z | r6 (6) | gripper]
+  vx, vy, wz = measured base velocity in the body frame from this and the previous frame (0 at frame 0);
+  height     = measured Go2 base height;  x, y, z = EEF position in the body frame (z = height above the floor);
+  r6         = rot6d of the EEF orientation in the body frame = first two COLUMNS [R[:,0], R[:,1]] (Zhou et al.);
+  gripper    = measured finger closure 0 (open) .. 1 (closed), continuous.
+  Placeholders: Sawyer height; Stretch vy, height are exactly 0.
+
+ACTION chunk, 14D per step k = 0 .. H-1 (action_chunk_from_targets, ref "base", rot "rot6d"):
+  [vx, vy, wz, height | x, y, z | r6 (6) | gripper]
+  base part = recorded base command at t+k;  gripper = recorded gripper command at t+k (0 open / 1 close);
+  x, y, z, r6 = the expert's (planner's) sparse EEF target at t+k seen from the body frame at t+k: where the EEF has
+              to be relative to the body (quasi-static, decoupled from base motion); nothing is subtracted.
+  Steps past the episode end repeat the last frame (LeRobot padding).
+Execution (decode_action_chunk): EEF target pose = (x, y, z, R(r6) orthonormalised), placed in the world with the body
+frame of the base pose measured when the step is executed, solved by the low-level IK; base command = the base part;
+gripper command = gripper > 0.5.
+Legacy layouts (older checkpoints only): rot="rpy" (11D, relative to the current EEF, roll/pitch/yaw), ref="eef"
+(relative to the current EEF), frame="base" (measured base pose as the body frame), future measured states as the EEF
+source (action_chunk_from_states).
+"""
+import numpy as np
+
+ROBOTS = ("go2_arm", "sawyer_umi", "stretch3")
+RAW_DIM = {"go2_arm": 12, "sawyer_umi": 11, "stretch3": 11}
+STATE_DIM, ACTION_DIM = 14, 14
+S_POS, S_ROT, S_GRIP = slice(4, 7), slice(7, 13), 13
+A_POS, A_ROT, A_GRIP = slice(4, 7), slice(7, 13), 13            # rot = "rot6d" (default)
+ACTION_DIM_RPY, A_ROT_RPY, A_GRIP_RPY = 11, slice(7, 10), 10      # rot = "rpy" (legacy: pi05_v2, GR00T v2 / v3)
+ROT_REPRS = ("rot6d", "rpy")
+EEF_REFS = ("base", "eef")      # EEF action: pose in the body frame ("base") or relative to the current EEF ("eef")
+DEFAULT_EEF_REF = "base"
+# body frame the EEF is seen from: "base" = measured base pose (Go2: torso, 0.31-0.42 m above the floor, with gait
+# roll/pitch; Sawyer / Stretch: 0.06 m); "ground" = the base projected to the floor: origin (x_B, y_B, 0), z up, base
+# heading (yaw) only - the same reference on all robots
+BODY_FRAMES = ("base", "ground")
+BODY_FRAME = "ground"           # yiqi 11-14: unified frame        # rot6d default (yiqi 11-13: "相对于当前底盘,我的末端要走到什么位置去?什么角度去?")
+GRIPPER_CLOSE_THRESHOLD = 0.5   # execution: closure > 0.5 -> close command (1)
+_BASE_MASK = {"go2_arm": [1, 1, 1, 1], "sawyer_umi": [1, 1, 1, 0], "stretch3": [1, 0, 1, 0]}
+STATE_MASK = {r: np.array(m + [1] * (STATE_DIM - 4), bool) for r, m in _BASE_MASK.items()}
+ACTION_MASK = {r: np.array(m + [1] * (ACTION_DIM - 4), bool) for r, m in _BASE_MASK.items()}
+ACTION_MASK_RPY = {r: np.array(m + [1] * (ACTION_DIM_RPY - 4), bool) for r, m in _BASE_MASK.items()}
+ROT6D_LAYOUT = "cols"   # first two COLUMNS [R[:,0], R[:,1]] (Zhou et al.; OpenWAM). "rows" = pytorch3d / GR00T N1.7 pose.py
+# observation.state slots of the base (x, y, yaw, z); verified against the measured base pose (all tasks):
+STATE_BASE_SLOTS = {"go2_arm": (20, 21, 22, 23), "sawyer_umi": (0, 1, 2, None), "stretch3": (0, 1, 2, None)}
+# finger joints in observation.state and their fully-open magnitude (closed = 0 on all robots; verified per task)
+GRIPPER_JOINTS = {"go2_arm": ((24, 25), 0.035), "sawyer_umi": ((11, 12), 0.021), "stretch3": ((13, 14), 0.25)}
+IDENTITY_R6 = np.array([1.0, 0.0, 0.0, 0.0, 1.0, 0.0])
+
+
+# ---------------------------------------------------------------- quaternion helpers (w-x-y-z unless noted)
+def qnorm(q):
+    q = np.asarray(q, np.float64)
+    return q / np.linalg.norm(q, axis=-1, keepdims=True)
+
+
+def qconj(q):
+    return q * np.array([1.0, -1.0, -1.0, -1.0])
+
+
+def qmul(a, b):
+    aw, ax, ay, az = np.moveaxis(a, -1, 0)
+    bw, bx, by, bz = np.moveaxis(b, -1, 0)
+    return np.stack([aw * bw - ax * bx - ay * by - az * bz,
+                     aw * bx + ax * bw + ay * bz - az * by,
+                     aw * by - ax * bz + ay * bw + az * bx,
+                     aw * bz + ax * by - ay * bx + az * bw], -1)
+
+
+def qrotmat(q):
+    w, x, y, z = np.moveaxis(qnorm(q), -1, 0)
+    return np.stack([
+        np.stack([1 - 2 * (y * y + z * z), 2 * (x * y - w * z), 2 * (x * z + w * y)], -1),
+        np.stack([2 * (x * y + w * z), 1 - 2 * (x * x + z * z), 2 * (y * z - w * x)], -1),
+        np.stack([2 * (x * z - w * y), 2 * (y * z + w * x), 1 - 2 * (x * x + y * y)], -1)], -2)
+
+
+def wxyz_to_xyzw(q):
+    return np.concatenate([q[..., 1:], q[..., :1]], -1)
+
+
+def xyzw_to_wxyz(q):
+    return np.concatenate([q[..., 3:], q[..., :3]], -1)
+
+
+def geodesic_deg(qa, qb):
+    d = np.abs(np.sum(qnorm(qa) * qnorm(qb), -1)).clip(0, 1)
+    return np.degrees(2 * np.arccos(d))
+
+
+def rotmat_to_quat(R):
+    """Rotation matrices [..., 3, 3] -> unit quaternions w-x-y-z (w >= 0); Shepperd's method (numerically stable,
+    also for near-identity and near-180-degree rotations)."""
+    R = np.asarray(R, np.float64)
+    m00, m11, m22 = R[..., 0, 0], R[..., 1, 1], R[..., 2, 2]
+    tr = m00 + m11 + m22
+    cand = []
+    w = np.sqrt(np.maximum(1 + tr, 1e-300)) / 2                                     # case w largest
+    cand.append(np.stack([w, (R[..., 2, 1] - R[..., 1, 2]) / (4 * w), (R[..., 0, 2] - R[..., 2, 0]) / (4 * w), (R[..., 1, 0] - R[..., 0, 1]) / (4 * w)], -1))
+    x = np.sqrt(np.maximum(1 + m00 - m11 - m22, 1e-300)) / 2                        # case x largest
+    cand.append(np.stack([(R[..., 2, 1] - R[..., 1, 2]) / (4 * x), x, (R[..., 0, 1] + R[..., 1, 0]) / (4 * x), (R[..., 0, 2] + R[..., 2, 0]) / (4 * x)], -1))
+    y = np.sqrt(np.maximum(1 - m00 + m11 - m22, 1e-300)) / 2                        # case y largest
+    cand.append(np.stack([(R[..., 0, 2] - R[..., 2, 0]) / (4 * y), (R[..., 0, 1] + R[..., 1, 0]) / (4 * y), y, (R[..., 1, 2] + R[..., 2, 1]) / (4 * y)], -1))
+    z = np.sqrt(np.maximum(1 - m00 - m11 + m22, 1e-300)) / 2                        # case z largest
+    cand.append(np.stack([(R[..., 1, 0] - R[..., 0, 1]) / (4 * z), (R[..., 0, 2] + R[..., 2, 0]) / (4 * z), (R[..., 1, 2] + R[..., 2, 1]) / (4 * z), z], -1))
+    pick = np.argmax(np.stack([tr, m00, m11, m22], -1), -1)[..., None, None]
+    q = np.take_along_axis(np.stack(cand, -2), pick, -2)[..., 0, :]
+    q = np.where(q[..., :1] < 0, -q, q)
+    return qnorm(q)
+
+
+def rotmat_to_rot6d(R):
+    """[..., 3, 3] -> [..., 6]: the first two rows (ROT6D_LAYOUT = "rows") or columns ("cols") of R."""
+    R = np.asarray(R, np.float64)
+    if ROT6D_LAYOUT == "rows":
+        return R[..., :2, :].reshape(R.shape[:-2] + (6,))
+    return np.concatenate([R[..., :, 0], R[..., :, 1]], -1)
+
+
+def rot6d_to_rotmat(r6):
+    """[..., 6] -> [..., 3, 3] by Gram-Schmidt (also valid for non-orthonormal network outputs)."""
+    r6 = np.asarray(r6, np.float64)
+    a1, a2 = r6[..., :3], r6[..., 3:6]
+    b1 = a1 / np.linalg.norm(a1, axis=-1, keepdims=True)
+    b2 = a2 - np.sum(b1 * a2, -1, keepdims=True) * b1
+    b2 = b2 / np.linalg.norm(b2, axis=-1, keepdims=True)
+    b3 = np.cross(b1, b2)
+    return np.stack([b1, b2, b3], -2) if ROT6D_LAYOUT == "rows" else np.stack([b1, b2, b3], -1)
+
+
+def rotmat_to_rotvec(R):
+    """[..., 3, 3] -> rotation vectors [..., 3] (axis * angle, angle in [0, pi])."""
+    q = rotmat_to_quat(R)
+    s = np.linalg.norm(q[..., 1:], axis=-1, keepdims=True)
+    ang = 2 * np.arctan2(s, q[..., :1])
+    return np.where(s > 1e-12, q[..., 1:] / np.maximum(s, 1e-12) * ang, 2 * q[..., 1:])
+
+
+def rotvec_to_rotmat(v):
+    v = np.asarray(v, np.float64)
+    ang = np.linalg.norm(v, axis=-1, keepdims=True)
+    axis = v / np.maximum(ang, 1e-12)
+    q = np.concatenate([np.cos(ang / 2), axis * np.sin(ang / 2)], -1)
+    return qrotmat(q)
+
+
+
+
+def rotmat_to_rpy(R):
+    """[..., 3, 3] -> [..., 3] = (roll, pitch, yaw), R = Rz(yaw) Ry(pitch) Rx(roll) (ROS / URDF convention)."""
+    R = np.asarray(R, np.float64)
+    pitch = np.arcsin(np.clip(-R[..., 2, 0], -1.0, 1.0))
+    return np.stack([np.arctan2(R[..., 2, 1], R[..., 2, 2]), pitch, np.arctan2(R[..., 1, 0], R[..., 0, 0])], -1)
+
+
+def rpy_to_rotmat(rpy):
+    r, p, y = np.moveaxis(np.asarray(rpy, np.float64), -1, 0)
+    cr, sr, cp, sp, cy, sy = np.cos(r), np.sin(r), np.cos(p), np.sin(p), np.cos(y), np.sin(y)
+    return np.stack([np.stack([cy * cp, cy * sp * sr - sy * cr, cy * sp * cr + sy * sr], -1),
+                     np.stack([sy * cp, sy * sp * sr + cy * cr, sy * sp * cr - cy * sr], -1),
+                     np.stack([-sp, cp * sr, cp * cr], -1)], -2)
+
+
+def rot_angle_deg(R):
+    """Rotation angle of rotation matrices [..., 3, 3] in degrees."""
+    return np.degrees(np.arccos(np.clip((np.trace(R, axis1=-2, axis2=-1) - 1) / 2, -1, 1)))
+
+
+def yaw_quat(yaw):
+    yaw = np.asarray(yaw, np.float64)
+    z = np.zeros_like(yaw)
+    return np.stack([np.cos(yaw / 2), z, z, np.sin(yaw / 2)], -1)
+
+
+def base_heading_pose(robot, state):
+    """observation.state [..., D] -> base heading pose [..., 7] = [x, y, z, qw, 0, 0, qz] (world frame)."""
+    state = np.asarray(state, np.float64)
+    ix, iy, iyaw, iz = STATE_BASE_SLOTS[robot]
+    z = state[..., iz] if iz is not None else np.zeros(state.shape[:-1])
+    return np.concatenate([np.stack([state[..., ix], state[..., iy], z], -1), yaw_quat(state[..., iyaw])], -1)
+
+
+def _axis_aligned_quats():
+    """The 24 rotations that map coordinate axes onto coordinate axes (w-x-y-z, w >= 0)."""
+    out = []
+    for perm in ((0, 1, 2), (0, 2, 1), (1, 0, 2), (1, 2, 0), (2, 0, 1), (2, 1, 0)):
+        for signs in np.ndindex(2, 2, 2):
+            M = np.zeros((3, 3))
+            for r, (c, sg) in enumerate(zip(perm, signs)):
+                M[r, c] = -1.0 if sg else 1.0
+            if np.linalg.det(M) < 0:
+                continue
+            w = np.sqrt(max(0.0, 1 + np.trace(M))) / 2
+            if w > 1e-6:
+                q = np.array([w, (M[2, 1] - M[1, 2]) / (4 * w), (M[0, 2] - M[2, 0]) / (4 * w), (M[1, 0] - M[0, 1]) / (4 * w)])
+            else:  # 180 deg rotations
+                axis = np.sqrt(np.clip((np.diag(M) + 1) / 2, 0, 1))
+                i = int(np.argmax(axis))
+                for j in range(3):
+                    if j != i and M[i, j] + M[j, i] < 0:
+                        axis[j] = -axis[j]
+                q = np.r_[0.0, axis]
+            out.append(q)
+    return np.array(out)
+
+
+AXIS_ALIGNED_QUATS = _axis_aligned_quats()
+
+
+def base_pose_from_sim_states(sim_states, sim_step_idx, robot, state, return_offset=False):
+    """Full world base pose [T, 7] = [x, y, z, qw, qx, qy, qz] for parquet frames with the given sim_step_idx [T].
+    sim_states: path to sim_info/0/sim_states.npz or the loaded npz mapping (keys robot__<Name>__base_pos/base_quat,
+    50 Hz control steps; parquet keeps every 2nd step). Frame s <- npz row with sim_step_idx == s - 1.
+    A constant asset-frame offset in base_quat (e.g. Ry(-90 deg) for Sawyer in N051/N053/N058/N059) is removed: the
+    mean of heading(state)^-1 * base_quat is snapped to the nearest axis-aligned rotation c and base_quat * c^-1 is
+    returned, so the body frame is x forward / z up everywhere (Go2 keeps its real roll / pitch)."""
+    z = np.load(sim_states) if isinstance(sim_states, (str, bytes)) or hasattr(sim_states, "__fspath__") else sim_states
+    key = [k for k in z.files if k.startswith("robot__") and k.endswith("__base_pos")]
+    assert len(key) == 1, key
+    key = key[0][: -len("__base_pos")]
+    row = {int(v): i for i, v in enumerate(np.asarray(z["sim_step_idx"]))}
+    idx = np.array([row.get(int(s) - 1, -1) for s in np.asarray(sim_step_idx).reshape(-1)])
+    assert (idx >= 0).all(), f"{(idx < 0).sum()} frames have no sim state at sim_step_idx - 1"
+    q = qnorm(np.asarray(z[key + "__base_quat"], np.float64)[idx])
+    off = qmul(qconj(qnorm(base_heading_pose(robot, state)[..., 3:7])), q)              # [T, 4]
+    c = AXIS_ALIGNED_QUATS[int(np.argmax(np.abs(off @ AXIS_ALIGNED_QUATS.T).sum(0)))]
+    q = qnorm(qmul(q, qconj(c)[None]))
+    q = np.where(q[:, :1] < 0, -q, q)
+    pose = np.concatenate([np.asarray(z[key + "__base_pos"], np.float64)[idx], q], -1)
+    return (pose, c) if return_offset else pose
+
+
+def asset_offset(robot, base_quat, state):
+    """Online / per-episode: the constant asset-frame offset c (axis-aligned rotation, identity except e.g. Sawyer
+    N051/N053/N058/N059 Ry(-90 deg)) from the simulator base_quat [..., 4] (w-x-y-z, same link as sim_states.npz) and
+    the raw observation.state [..., D] of one or more frames of the episode. Compute once per episode, then use
+    base_pose_online for every frame. Same rule as base_pose_from_sim_states."""
+    q = qnorm(np.atleast_2d(np.asarray(base_quat, np.float64)))
+    off = qmul(qconj(qnorm(base_heading_pose(robot, np.atleast_2d(np.asarray(state, np.float64)))[..., 3:7])), q)
+    return AXIS_ALIGNED_QUATS[int(np.argmax(np.abs(off @ AXIS_ALIGNED_QUATS.T).sum(0)))]
+
+
+def base_pose_online(base_pos, base_quat, offset):
+    """Online: measured base pose [..., 7] = [x, y, z, qw, qx, qy, qz] with the asset offset removed (as stored in
+    observation.base_pose), from the simulator base_pos [..., 3] / base_quat [..., 4] taken at the observation time."""
+    q = qnorm(qmul(qnorm(np.asarray(base_quat, np.float64)), qconj(np.asarray(offset, np.float64))))
+    q = np.where(q[..., :1] < 0, -q, q)
+    return np.concatenate([np.asarray(base_pos, np.float64), q], -1)
+
+
+def sim_states_alignment_error(sim_states, sim_step_idx, eef_pose):
+    """Consistency check of the npz <-> parquet alignment: max |npz EEF - observation.eef_pose| as (position m,
+    rotation deg) over the episode. Both are ~0 (float32 precision) when the npz belongs to the same recording."""
+    z = np.load(sim_states) if isinstance(sim_states, (str, bytes)) or hasattr(sim_states, "__fspath__") else sim_states
+    key = [k for k in z.files if k.startswith("robot__") and k.endswith("__eef_pos")][0][: -len("__eef_pos")]
+    row = {int(v): i for i, v in enumerate(np.asarray(z["sim_step_idx"]))}
+    idx = np.array([row[int(s) - 1] for s in np.asarray(sim_step_idx).reshape(-1)])
+    eef_pose = np.asarray(eef_pose, np.float64)
+    pos = np.abs(np.asarray(z[key + "__eef_pos"], np.float64)[idx] - eef_pose[:, :3]).max()
+    rot = geodesic_deg(np.asarray(z[key + "__eef_quat"], np.float64)[idx], eef_pose[:, 3:7]).max()
+    return float(pos), float(rot)
+
+
+
+
+# ---------------------------------------------------------------- raw <-> parts
+def split_raw(robot, raw):
+    raw = np.asarray(raw, np.float64)
+    assert raw.shape[-1] == RAW_DIM[robot], (robot, raw.shape)
+    if robot == "go2_arm":
+        base = raw[..., 0:4]
+        eef = raw[..., 4:11]
+    else:
+        base = np.concatenate([raw[..., 0:3], np.zeros(raw.shape[:-1] + (1,))], -1)
+        eef = raw[..., 3:10]
+    if robot == "stretch3":
+        base = base.copy(); base[..., 1] = 0.0   # differential drive: vy is never executable
+    return base, eef[..., :3], eef[..., 3:7], raw[..., -1:]
+
+
+def join_raw(robot, base, pos, quat_wxyz, grip):
+    if robot == "go2_arm":
+        return np.concatenate([base, pos, quat_wxyz, grip], -1)
+    b = base[..., :3].copy()
+    if robot == "stretch3":
+        b[..., 1] = 0.0
+    return np.concatenate([b, pos, quat_wxyz, grip], -1)
+
+
+# ---------------------------------------------------------------- canonical state / actions
+def body_frame(base_poses, frame=None):
+    """Measured base poses [..., 7] -> pose of the body frame the EEF is seen from [..., 7] (BODY_FRAME by default)."""
+    B = np.asarray(base_poses, np.float64)
+    frame = frame or BODY_FRAME
+    if frame == "base":
+        return np.concatenate([B[..., :3], qnorm(B[..., 3:7])], -1)
+    if frame != "ground":
+        raise ValueError(f"frame must be one of {BODY_FRAMES}, got {frame!r}")
+    R = qrotmat(B[..., 3:7])
+    yaw = np.arctan2(R[..., 1, 0], R[..., 0, 0])                     # heading of the base x-axis in the horizontal plane
+    return np.concatenate([B[..., :2], np.zeros(B.shape[:-1] + (1,)), yaw_quat(yaw)], -1)
+
+
+def canonical_state(robot, base_poses, eef_poses, state, timestamps, frame=None):
+    """Per-frame canonical state [T, 14] of one episode (see module docstring); EEF and base velocity are expressed in
+    body_frame(base_poses, frame); height = measured base height."""
+    B0, E = np.asarray(base_poses, np.float64), np.asarray(eef_poses, np.float64)
+    B = body_frame(B0, frame)
+    S, ts = np.asarray(state, np.float64), np.asarray(timestamps, np.float64).reshape(-1)
+    T = len(B)
+    qb = qnorm(B[:, 3:7]); R = qrotmat(qb)
+    out = np.zeros((T, STATE_DIM))
+    if T > 1:
+        dt = np.diff(ts)[:, None]
+        v = np.einsum("ti,tij->tj", (B[1:, :3] - B[:-1, :3]) / dt, R[1:])           # world velocity -> current base frame
+        dq = qmul(qconj(qb[:-1]), qb[1:]); dq = np.where(dq[:, :1] < 0, -dq, dq)    # base rotation over the step
+        ang = 2 * np.arctan2(np.linalg.norm(dq[:, 1:], axis=1), dq[:, 0])
+        axis = dq[:, 1:] / np.maximum(np.linalg.norm(dq[:, 1:], axis=1, keepdims=True), 1e-12)
+        out[1:, 0:2] = v[:, :2]
+        out[1:, 2] = (axis[:, 2] * ang) / dt[:, 0]
+    out[:, 3] = B0[:, 2]
+    out[:, S_POS] = np.einsum("ti,tij->tj", E[:, :3] - B[:, :3], R)                    # R_B^T (p_E - p_B)
+    out[:, S_ROT] = rotmat_to_rot6d(np.einsum("tji,tjk->tik", R, qrotmat(E[:, 3:7])))   # R_B^T R_E
+    idx, open_mag = GRIPPER_JOINTS[robot]
+    out[:, S_GRIP] = np.clip(1 - np.abs(S[:, list(idx)]).mean(1) / open_mag, 0, 1)
+    out[:, ~STATE_MASK[robot]] = 0.0
+    return out
+
+
+def command_columns(robot, raw_actions):
+    """Recorded commands per frame [T, 5] = [vx, vy, wz, height, gripper] (dataset column action.command;
+    placeholders 0: Sawyer height, Stretch vy and height)."""
+    base, _, _, grip = split_raw(robot, raw_actions)
+    out = np.concatenate([base, grip], -1)
+    out[:, :4][:, ~ACTION_MASK[robot][:4]] = 0.0
+    return out
+
+
+def action_layout(rot="rot6d"):
+    """(action dim, rotation slice, gripper index, per-robot mask) of an action chunk with rotation representation rot."""
+    if rot == "rot6d":
+        return ACTION_DIM, A_ROT, A_GRIP, ACTION_MASK
+    if rot == "rpy":
+        return ACTION_DIM_RPY, A_ROT_RPY, A_GRIP_RPY, ACTION_MASK_RPY
+    raise ValueError(f"rot must be one of {ROT_REPRS}, got {rot!r}")
+
+
+def rot_from_action_dim(dim):
+    return {ACTION_DIM: "rot6d", ACTION_DIM_RPY: "rpy"}[dim]
+
+
+def relative_rotation(R_s, R_x, rot="rot6d"):
+    """R_s^T R_x (rotation of x relative to the current EEF, in the current EEF frame) as rot6d (cols) or roll/pitch/yaw.
+    Always a relative rotation - never a difference of absolute angles."""
+    R = np.einsum("...ji,...jk->...ik", R_s, R_x)
+    return rotmat_to_rot6d(R) if rot == "rot6d" else rotmat_to_rpy(R)
+
+
+def relative_rotation_matrix(a_rot, rot="rot6d"):
+    """Inverse of relative_rotation: rot6d (Gram-Schmidt, tolerates non-orthonormal model outputs) or rpy -> R."""
+    return rot6d_to_rotmat(a_rot) if rot == "rot6d" else rpy_to_rotmat(a_rot)
+
+
+def _eef_ref(rot, ref):
+    ref = ref or ("eef" if rot == "rpy" else DEFAULT_EEF_REF)
+    if ref not in EEF_REFS or (rot == "rpy" and ref != "eef"):
+        raise ValueError(f"unsupported EEF reference {ref!r} for rot {rot!r}")
+    return ref
+
+
+def eef_action(state_t, pos, R, rot="rot6d", ref=None):
+    """EEF part of the action steps from body-frame EEF poses (pos [..., H, 3], R [..., H, 3, 3]):
+    ref "base" (default, rot6d only): [pos | rot6d(R)] - where the EEF has to be, seen from the body;
+    ref "eef" (legacy): [pos - state(t).xyz | relative_rotation(R(state(t).r6), R)] - relative to the current EEF."""
+    ref = _eef_ref(rot, ref)
+    if ref == "base":
+        return np.concatenate([pos, rotmat_to_rot6d(R)], -1)
+    s = np.asarray(state_t, np.float64)
+    R_s = rot6d_to_rotmat(s[..., None, S_ROT])
+    return np.concatenate([pos - s[..., None, S_POS], relative_rotation(R_s, R, rot)], -1)
+
+
+def eef_pose_from_action(state_t, a_pos, a_rot, rot="rot6d", ref=None):
+    """Inverse of eef_action -> body-frame EEF pose (pos [..., H, 3], R [..., H, 3, 3]); rot6d outputs are
+    orthonormalised (Gram-Schmidt), so non-orthonormal model outputs give proper rotations."""
+    ref = _eef_ref(rot, ref)
+    if ref == "base":
+        return np.asarray(a_pos, np.float64), rot6d_to_rotmat(a_rot)
+    s = np.asarray(state_t, np.float64)
+    return s[..., None, S_POS] + a_pos, rot6d_to_rotmat(s[..., None, S_ROT]) @ relative_rotation_matrix(a_rot, rot)
+
+
+def action_chunk_from_states(robot, state_t, future_states, commands, rot="rot6d", ref=None):
+    """Action chunk from the future measured states: state(t) [..., 14], state(t+1 .. t+H) [..., H, 14] and the
+    recorded commands at t .. t+H-1 [..., H, 5] -> [..., H, 14] (rot6d) / [..., H, 11] (legacy rpy)."""
+    f, c = np.asarray(future_states, np.float64), np.asarray(commands, np.float64)
+    out = np.concatenate([c[..., 0:4], eef_action(state_t, f[..., S_POS], rot6d_to_rotmat(f[..., S_ROT]), rot, ref),
+                          c[..., 4:5]], -1)
+    out[..., ~action_layout(rot)[3][robot]] = 0.0
+    return out
+
+
+def expert_target_columns(robot, raw_actions, base_poses, frame=None):
+    """Per-frame expert EEF target seen from the body at the same frame [T, 9] = [x, y, z | rot6d] (dataset column
+    action.eef_target): R_B(t)^T (p_target(t) - p_B(t)) and rot6d(R_B(t)^T R_target(t))."""
+    _, p_tgt, q_tgt, _ = split_raw(robot, raw_actions)
+    B = body_frame(base_poses, frame)
+    R = qrotmat(B[..., 3:7])
+    pos = np.einsum("...i,...ij->...j", p_tgt - B[..., :3], R)
+    rot = rotmat_to_rot6d(np.einsum("...ji,...jk->...ik", R, qrotmat(q_tgt)))
+    return np.concatenate([pos, rot], -1)
+
+
+def action_chunk_from_targets(robot, state_t, targets, commands, rot="rot6d", ref=None):
+    """Action chunk from the expert targets at t .. t+H-1 [..., H, 9] (body frame of their own frame) and the recorded
+    commands at t .. t+H-1 [..., H, 5] -> [..., H, 14]: [base cmd | target xyz | target rot6d | gripper cmd] (ref
+    "base", default; state(t) is not used). Legacy: ref "eef" (relative to state(t)), rot "rpy" (11D)."""
+    g, c = np.asarray(targets, np.float64), np.asarray(commands, np.float64)
+    out = np.concatenate([c[..., 0:4], eef_action(state_t, g[..., 0:3], rot6d_to_rotmat(g[..., 3:9]), rot, ref),
+                          c[..., 4:5]], -1)
+    out[..., ~action_layout(rot)[3][robot]] = 0.0
+    return out
+
+
+def episode_target_chunks(robot, canon_states, targets, commands, horizon, rot="rot6d", ref=None):
+    """All training chunks of one episode [T, H, A] from expert targets and commands at t .. t+H-1 (clamped)."""
+    S, G, Cm = (np.asarray(x, np.float64) for x in (canon_states, targets, commands))
+    T = len(S)
+    idx = np.minimum(np.arange(T)[:, None] + np.arange(horizon)[None, :], T - 1)
+    return action_chunk_from_targets(robot, S, G[idx], Cm[idx], rot, ref)
+
+
+def future_states_from_chunk(robot, state_t, action_chunk, rot=None, ref=None):
+    """Inverse of action_chunk_from_states / _from_targets for the EEF: -> predicted EEF part of the state [..., H, 14]
+    (base / gripper slots carry the commands). rot defaults to the one implied by the action dim (14 rot6d, 11 rpy),
+    ref to DEFAULT_EEF_REF for rot6d ("eef" for rpy)."""
+    a = np.asarray(action_chunk, np.float64).copy()
+    rot = rot or rot_from_action_dim(a.shape[-1])
+    _, a_rot, a_grip, mask = action_layout(rot)
+    a[..., ~mask[robot]] = 0.0
+    pos, R = eef_pose_from_action(state_t, a[..., A_POS], a[..., a_rot], rot, ref)
+    out = np.zeros(a.shape[:-1] + (STATE_DIM,))
+    out[..., 0:4] = a[..., 0:4]
+    out[..., S_POS] = pos
+    out[..., S_ROT] = rotmat_to_rot6d(R)
+    out[..., S_GRIP] = a[..., a_grip]          # the gripper command (not a measured closure)
+    return out
+
+
+def decode_action_chunk(robot, action_chunk, state_t, base_poses_now, rot=None, ref=None, frame=None):
+    """Model output [H, 14] (legacy rpy [H, 11]) + current canonical state [14] (used only by the legacy ref "eef") +
+    body pose measured when each step is executed [H, 7] -> controller commands in the recorded format [H, D_robot]:
+    base velocity command, world EEF target (wxyz) and gripper command (closure > GRIPPER_CLOSE_THRESHOLD).
+    Unsupported dimensions are never emitted."""
+    a = np.asarray(action_chunk, np.float64)
+    rot = rot or rot_from_action_dim(a.shape[-1])
+    f = future_states_from_chunk(robot, state_t, a, rot, ref)
+    B = body_frame(base_poses_now, frame)                              # measured base pose -> body frame
+    R_b = qrotmat(B[..., 3:7])
+    p = B[..., :3] + np.einsum("...ij,...j->...i", R_b, f[..., S_POS])
+    q = rotmat_to_quat(R_b @ rot6d_to_rotmat(f[..., S_ROT]))
+    a_grip = action_layout(rot)[2]
+    grip = (a[..., a_grip:a_grip + 1] > GRIPPER_CLOSE_THRESHOLD).astype(np.float64)
+    return join_raw(robot, f[..., 0:4], p, q, grip)
+
+
+def episode_action_chunks(robot, canon_states, commands, horizon, rot="rot6d", ref=None):
+    """All training chunks of one episode [T, H, A] from its canonical states (future steps t+1 .. t+H) and the
+    recorded commands [T, 5] (steps t .. t+H-1); indices are clamped to the last frame like LeRobot."""
+    S = np.asarray(canon_states, np.float64)
+    Cm = np.asarray(commands, np.float64)
+    T = len(S)
+    k = np.arange(horizon)[None, :]
+    return action_chunk_from_states(robot, S, S[np.minimum(np.arange(T)[:, None] + k + 1, T - 1)],
+                                    Cm[np.minimum(np.arange(T)[:, None] + k, T - 1)], rot, ref)
diff --git a/src/openpi/policies/mmb_policy.py b/src/openpi/policies/mmb_policy.py
new file mode 100644
index 0000000..f4b3275
--- /dev/null
+++ b/src/openpi/policies/mmb_policy.py
@@ -0,0 +1,169 @@
+"""MobileManipBench (prlq/mobilemanipbench-data) inputs/outputs for pi0.5 warm-up on all three robots.
+
+Dataset (built by /data/changyi/mmb/convert_merge.py): actions[12] and observation.state[26] are each robot's
+raw vectors zero-padded (no semantic alignment across robots, as in pi0 pretraining); robot_id selects the
+robot's own normalisation statistics (PerRobotNormalize). The global openpi Normalize step is fed identity
+statistics, so all real normalisation happens here.
+
+v2 (dataset built with convert_merge.py --v2; definitions in mmb_canonical_action.py, shared with GR00T / OpenWAM):
+state   = observation.state_canonical, 14D [base vel + height | EEF xyz in base frame | EEF rot6d | gripper closure];
+actions = built by MMBActionsFromTargets (expert sparse targets action.eef_target at t .. t+H-1) or MMBActionsFromStates
+(future measured states t+1 .. t+H), 14D per step [recorded base command | EEF target xyz in the body frame | EEF
+target rot6d in the body frame | recorded gripper command] (rot="rot6d", ref="base", default; legacy ref="eef":
+relative to the current EEF; legacy rot="rpy": 11D, relative, roll/pitch/yaw) - the only training-time processing
+besides chunking and normalisation. Inactive dimensions (sawyer height; stretch vy, height) are exactly 0 before and after
+normalisation; like openpi's 32D padding they DO contribute to the flow-matching loss (target 0). MMBOutputs zeros
+them again at inference.
+"""
+import dataclasses
+import json
+import pathlib
+
+import einops
+import numpy as np
+
+from openpi import transforms
+from openpi.models import model as _model
+from openpi.policies import mmb_canonical_action as canon
+
+ROBOT_NAMES = {0: "go2_arm", 1: "sawyer_umi", 2: "stretch3"}
+ACTION_DIMS = {0: 12, 1: 11, 2: 11}
+PER_ROBOT_STATS_FILE = "norm_stats_per_robot.json"
+
+
+def _parse_image(image) -> np.ndarray:
+    image = np.asarray(image)
+    if np.issubdtype(image.dtype, np.floating):
+        image = (255 * image).astype(np.uint8)
+    if image.shape[0] == 3:
+        image = einops.rearrange(image, "c h w -> h w c")
+    return image
+
+
+@dataclasses.dataclass(frozen=True)
+class MMBInputs(transforms.DataTransformFn):
+    model_type: _model.ModelType
+
+    def __call__(self, data: dict) -> dict:
+        head = _parse_image(data["observation/image"])
+        wrist = _parse_image(data["observation/wrist_image"])
+        inputs = {
+            "state": np.asarray(data["observation/state"], np.float32),
+            "robot_id": int(np.asarray(data["robot_id"]).reshape(-1)[0]),
+            "image": {"base_0_rgb": head, "left_wrist_0_rgb": wrist, "right_wrist_0_rgb": np.zeros_like(head)},
+            "image_mask": {
+                "base_0_rgb": np.True_,
+                "left_wrist_0_rgb": np.True_,
+                "right_wrist_0_rgb": np.True_ if self.model_type == _model.ModelType.PI0_FAST else np.False_,
+            },
+        }
+        if "actions" in data:
+            inputs["actions"] = np.asarray(data["actions"], np.float32)
+        if "prompt" in data:
+            inputs["prompt"] = data["prompt"]
+        if "commands" in data:  # v2: recorded commands [vx, vy, wz, height, gripper] at t .. t+H-1
+            inputs["commands"] = np.asarray(data["commands"], np.float32)
+        if "targets" in data:   # v2 (expert-target variant): expert EEF targets [xyz | rot6d] at t .. t+H-1
+            inputs["targets"] = np.asarray(data["targets"], np.float32)
+        return inputs
+
+
+@dataclasses.dataclass(frozen=True)
+class MMBActionsFromStates(transforms.DataTransformFn):
+    """v2 (training): the state arrives as a chunk state(t .. t+H) [H+1, 14] and the recorded commands as [H, 5];
+    the action chunk [H, 14] (rot6d; [H, 11] rpy) is derived from them (canon.action_chunk_from_states) and the
+    state becomes state(t). Must run before normalisation."""
+
+    rot: str = "rot6d"
+    ref: str | None = None
+
+    def __call__(self, data: dict) -> dict:
+        st = np.asarray(data["state"], np.float64)
+        if st.ndim == 1:                                   # inference: only the current state
+            return data
+        out = dict(data)
+        robot = ROBOT_NAMES[int(np.asarray(data["robot_id"]).reshape(-1)[0])]
+        out["actions"] = canon.action_chunk_from_states(robot, st[0], st[1:], out.pop("commands"), self.rot, self.ref).astype(np.float32)
+        out["state"] = st[0].astype(np.float32)
+        return out
+
+
+@dataclasses.dataclass(frozen=True)
+class MMBActionsFromTargets(transforms.DataTransformFn):
+    """v2, expert-target variant (training): action chunk [H, 14] (rot6d; [H, 11] rpy) from state(t) [14], the expert
+    EEF targets [H, 9] and the recorded commands [H, 5] at t .. t+H-1 (canon.action_chunk_from_targets). Before
+    normalisation."""
+
+    rot: str = "rot6d"
+    ref: str | None = None
+
+    def __call__(self, data: dict) -> dict:
+        if "targets" not in data:                         # inference
+            return data
+        out = dict(data)
+        robot = ROBOT_NAMES[int(np.asarray(data["robot_id"]).reshape(-1)[0])]
+        out["actions"] = canon.action_chunk_from_targets(robot, np.asarray(data["state"], np.float64), out.pop("targets"),
+                                                         out.pop("commands"), self.rot, self.ref).astype(np.float32)
+        return out
+
+
+def load_per_robot_stats(path: str | pathlib.Path) -> dict:
+    raw = json.loads(pathlib.Path(path).read_text())
+    return {int(rid): {k: {s: np.asarray(v, np.float32) for s, v in d.items()} for k, d in keys.items()}
+            for rid, keys in raw.items()}
+
+
+def _quantile_norm(x, q01, q99):
+    span = q99 - q01
+    ok = span > 1e-6  # padded / constant dimensions stay exactly 0
+    n = x.shape[-1]
+    q01, span, ok = q01[:n], span[:n], ok[:n]
+    y = np.where(ok, (x - q01) / np.where(ok, span, 1.0) * 2.0 - 1.0, 0.0)
+    return y.astype(np.float32)
+
+
+def _quantile_unnorm(y, q01, q99):
+    span = q99 - q01
+    ok = span > 1e-6
+    n = y.shape[-1]
+    q01, span, ok = q01[:n], span[:n], ok[:n]
+    return np.where(ok, (y + 1.0) / 2.0 * span + q01, 0.0).astype(np.float32)
+
+
+@dataclasses.dataclass(frozen=True)
+class PerRobotNormalize(transforms.DataTransformFn):
+    """Quantile-normalise state and actions with the statistics of the sample's own robot."""
+
+    stats: dict  # robot_id -> {"state"|"actions": {"q01","q99",...}}
+
+    def __call__(self, data: dict) -> dict:
+        st = self.stats[int(data["robot_id"])]
+        out = dict(data)
+        out["state"] = _quantile_norm(np.asarray(data["state"], np.float32), st["state"]["q01"], st["state"]["q99"])
+        if "actions" in data:
+            out["actions"] = _quantile_norm(np.asarray(data["actions"], np.float32), st["actions"]["q01"], st["actions"]["q99"])
+        out.pop("robot_id")
+        return out
+
+
+@dataclasses.dataclass(frozen=True)
+class MMBOutputs(transforms.DataTransformFn):
+    """Inference only: undo the per-robot normalisation and return the robot's own action dimensions.
+    canonical_v2: return the 14D (rot6d; 11D rpy) action chunk (inactive dims zeroed); the execution adapter converts it to controller
+    commands with canon.decode_action_chunk(robot, actions, state_t, base_pose_at_execution), where state_t is the
+    (unnormalised) canonical state of the observation that produced the chunk."""
+
+    stats: dict
+    robot_id: int
+    canonical_v2: bool = False
+    rot: str = "rot6d"
+
+    def __call__(self, data: dict) -> dict:
+        st = self.stats[self.robot_id]
+        n = canon.action_layout(self.rot)[0] if self.canonical_v2 else 12
+        a = np.asarray(data["actions"])[..., :n]
+        a = _quantile_unnorm(a, st["actions"]["q01"], st["actions"]["q99"])
+        if not self.canonical_v2:
+            return {"actions": a[..., : ACTION_DIMS[self.robot_id]]}
+        robot = ROBOT_NAMES[self.robot_id]
+        return {"actions": np.where(canon.action_layout(self.rot)[3][robot], a, 0.0).astype(np.float32)}
diff --git a/assets/pi05_mmb_v3/mmb/v2t/norm_stats_per_robot.json b/assets/pi05_mmb_v3/mmb/v2t/norm_stats_per_robot.json
new file mode 100644
index 0000000..9ce5298
--- /dev/null
+++ b/assets/pi05_mmb_v3/mmb/v2t/norm_stats_per_robot.json
@@ -0,0 +1 @@
+{"0": {"state": {"q01": [-0.30000001192092896, -0.2997402250766754, -0.7999288439750671, 0.28999999165534973, -0.160398930311203, -0.7775490880012512, 0.16975568234920502, -1.0, -1.0, -1.0, -1.0, -1.0, -1.0, -1.0], "q99": [0.3499999940395355, 0.16914814397692687, 0.800000011920929, 0.41369785636663436, 0.9592860096693039, 0.2369463916122912, 1.3730802536010742, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0], "mean": [0.06909525805149054, -0.003308204971539885, 0.019374634107881785, 0.3640675575727502, 0.34041552462654767, -0.15113465219859984, 0.7745604179620226, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0], "std": [0.138943109349931, 0.04951923437360962, 0.20712037634151662, 0.025289301915518446, 0.2849244620208932, 0.24881026083948993, 0.21464410493892133, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0]}, "actions": {"q01": [-0.30000001192092896, -0.2997402250766754, -0.7999288439750671, 0.28999999165534973, -0.160398930311203, -0.7775490880012512, 0.16975568234920502, -1.0, -1.0, -1.0, -1.0, -1.0, -1.0, -1.0], "q99": [0.3499999940395355, 0.16914814397692687, 0.800000011920929, 0.41369785636663436, 0.9592860096693039, 0.2369463916122912, 1.3730802536010742, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0], "mean": [0.06909525805149054, -0.003308204971539885, 0.019374634107881785, 0.3640675575727502, 0.34041552462654767, -0.15113465219859984, 0.7745604179620226, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0], "std": [0.138943109349931, 0.04951923437360962, 0.20712037634151662, 0.025289301915518446, 0.2849244620208932, 0.24881026083948993, 0.21464410493892133, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0]}}, "1": {"state": {"q01": [-0.30000001192092896, -0.2997402250766754, -0.7999288439750671, 0.0, -0.160398930311203, -0.7775490880012512, 0.16975568234920502, -1.0, -1.0, -1.0, -1.0, -1.0, -1.0, -1.0], "q99": [0.3499999940395355, 0.16914814397692687, 0.800000011920929, 0.0, 0.9592860096693039, 0.2369463916122912, 1.3730802536010742, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0], "mean": [0.06909525805149054, -0.003308204971539885, 0.019374634107881785, 0.0, 0.34041552462654767, -0.15113465219859984, 0.7745604179620226, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0], "std": [0.138943109349931, 0.04951923437360962, 0.20712037634151662, 0.0, 0.2849244620208932, 0.24881026083948993, 0.21464410493892133, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0]}, "actions": {"q01": [-0.30000001192092896, -0.2997402250766754, -0.7999288439750671, 0.0, -0.160398930311203, -0.7775490880012512, 0.16975568234920502, -1.0, -1.0, -1.0, -1.0, -1.0, -1.0, -1.0], "q99": [0.3499999940395355, 0.16914814397692687, 0.800000011920929, 0.0, 0.9592860096693039, 0.2369463916122912, 1.3730802536010742, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0], "mean": [0.06909525805149054, -0.003308204971539885, 0.019374634107881785, 0.0, 0.34041552462654767, -0.15113465219859984, 0.7745604179620226, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0], "std": [0.138943109349931, 0.04951923437360962, 0.20712037634151662, 0.0, 0.2849244620208932, 0.24881026083948993, 0.21464410493892133, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0]}}, "2": {"state": {"q01": [-0.30000001192092896, 0.0, -0.7999288439750671, 0.0, -0.160398930311203, -0.7775490880012512, 0.16975568234920502, -1.0, -1.0, -1.0, -1.0, -1.0, -1.0, -1.0], "q99": [0.3499999940395355, 0.0, 0.800000011920929, 0.0, 0.9592860096693039, 0.2369463916122912, 1.3730802536010742, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0], "mean": [0.06909525805149054, 0.0, 0.019374634107881785, 0.0, 0.34041552462654767, -0.15113465219859984, 0.7745604179620226, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0], "std": [0.138943109349931, 0.0, 0.20712037634151662, 0.0, 0.2849244620208932, 0.24881026083948993, 0.21464410493892133, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0]}, "actions": {"q01": [-0.30000001192092896, 0.0, -0.7999288439750671, 0.0, -0.160398930311203, -0.7775490880012512, 0.16975568234920502, -1.0, -1.0, -1.0, -1.0, -1.0, -1.0, -1.0], "q99": [0.3499999940395355, 0.0, 0.800000011920929, 0.0, 0.9592860096693039, 0.2369463916122912, 1.3730802536010742, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0], "mean": [0.06909525805149054, 0.0, 0.019374634107881785, 0.0, 0.34041552462654767, -0.15113465219859984, 0.7745604179620226, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0], "std": [0.138943109349931, 0.0, 0.20712037634151662, 0.0, 0.2849244620208932, 0.24881026083948993, 0.21464410493892133, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0]}}}
\ No newline at end of file
diff --git a/assets/pi05_mmb_v3/mmb/v2t/norm_stats.json b/assets/pi05_mmb_v3/mmb/v2t/norm_stats.json
new file mode 100644
index 0000000..f60d7f3
--- /dev/null
+++ b/assets/pi05_mmb_v3/mmb/v2t/norm_stats.json
@@ -0,0 +1,280 @@
+{
+ "norm_stats": {
+  "state": {
+   "mean": [
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0
+   ],
+   "std": [
+    1.0,
+    1.0,
+    1.0,
+    1.0,
+    1.0,
+    1.0,
+    1.0,
+    1.0,
+    1.0,
+    1.0,
+    1.0,
+    1.0,
+    1.0,
+    1.0,
+    1.0,
+    1.0,
+    1.0,
+    1.0,
+    1.0,
+    1.0,
+    1.0,
+    1.0,
+    1.0,
+    1.0,
+    1.0,
+    1.0,
+    1.0,
+    1.0,
+    1.0,
+    1.0,
+    1.0,
+    1.0
+   ],
+   "q01": [
+    -1.0,
+    -1.0,
+    -1.0,
+    -1.0,
+    -1.0,
+    -1.0,
+    -1.0,
+    -1.0,
+    -1.0,
+    -1.0,
+    -1.0,
+    -1.0,
+    -1.0,
+    -1.0,
+    -1.0,
+    -1.0,
+    -1.0,
+    -1.0,
+    -1.0,
+    -1.0,
+    -1.0,
+    -1.0,
+    -1.0,
+    -1.0,
+    -1.0,
+    -1.0,
+    -1.0,
+    -1.0,
+    -1.0,
+    -1.0,
+    -1.0,
+    -1.0
+   ],
+   "q99": [
+    1.0,
+    1.0,
+    1.0,
+    1.0,
+    1.0,
+    1.0,
+    1.0,
+    1.0,
+    1.0,
+    1.0,
+    1.0,
+    1.0,
+    1.0,
+    1.0,
+    1.0,
+    1.0,
+    1.0,
+    1.0,
+    1.0,
+    1.0,
+    1.0,
+    1.0,
+    1.0,
+    1.0,
+    1.0,
+    1.0,
+    1.0,
+    1.0,
+    1.0,
+    1.0,
+    1.0,
+    1.0
+   ]
+  },
+  "actions": {
+   "mean": [
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0
+   ],
+   "std": [
+    1.0,
+    1.0,
+    1.0,
+    1.0,
+    1.0,
+    1.0,
+    1.0,
+    1.0,
+    1.0,
+    1.0,
+    1.0,
+    1.0,
+    1.0,
+    1.0,
+    1.0,
+    1.0,
+    1.0,
+    1.0,
+    1.0,
+    1.0,
+    1.0,
+    1.0,
+    1.0,
+    1.0,
+    1.0,
+    1.0,
+    1.0,
+    1.0,
+    1.0,
+    1.0,
+    1.0,
+    1.0
+   ],
+   "q01": [
+    -1.0,
+    -1.0,
+    -1.0,
+    -1.0,
+    -1.0,
+    -1.0,
+    -1.0,
+    -1.0,
+    -1.0,
+    -1.0,
+    -1.0,
+    -1.0,
+    -1.0,
+    -1.0,
+    -1.0,
+    -1.0,
+    -1.0,
+    -1.0,
+    -1.0,
+    -1.0,
+    -1.0,
+    -1.0,
+    -1.0,
+    -1.0,
+    -1.0,
+    -1.0,
+    -1.0,
+    -1.0,
+    -1.0,
+    -1.0,
+    -1.0,
+    -1.0
+   ],
+   "q99": [
+    1.0,
+    1.0,
+    1.0,
+    1.0,
+    1.0,
+    1.0,
+    1.0,
+    1.0,
+    1.0,
+    1.0,
+    1.0,
+    1.0,
+    1.0,
+    1.0,
+    1.0,
+    1.0,
+    1.0,
+    1.0,
+    1.0,
+    1.0,
+    1.0,
+    1.0,
+    1.0,
+    1.0,
+    1.0,
+    1.0,
+    1.0,
+    1.0,
+    1.0,
+    1.0,
+    1.0,
+    1.0
+   ]
+  }
+ }
+}
\ No newline at end of file