Spaces:
Sleeping
Sleeping
File size: 94,018 Bytes
34f3bc9 | 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 1273 1274 1275 1276 1277 1278 1279 1280 1281 1282 1283 1284 1285 1286 1287 1288 1289 1290 1291 1292 1293 1294 1295 1296 1297 1298 1299 1300 1301 1302 1303 1304 1305 1306 1307 1308 1309 1310 1311 1312 1313 1314 1315 1316 1317 1318 1319 1320 1321 1322 1323 1324 1325 1326 1327 1328 1329 1330 1331 1332 1333 1334 1335 1336 1337 1338 1339 1340 1341 1342 1343 1344 1345 1346 1347 1348 1349 1350 1351 1352 1353 1354 1355 1356 1357 1358 1359 1360 1361 1362 1363 1364 1365 1366 1367 1368 1369 1370 1371 1372 1373 1374 1375 1376 1377 1378 1379 1380 1381 1382 1383 1384 1385 1386 1387 1388 1389 1390 1391 1392 1393 1394 1395 1396 1397 1398 1399 1400 1401 1402 1403 1404 1405 1406 1407 1408 1409 1410 1411 1412 1413 1414 1415 1416 1417 1418 1419 1420 1421 1422 1423 1424 1425 1426 1427 1428 1429 1430 1431 1432 1433 1434 1435 1436 1437 1438 1439 1440 1441 1442 1443 1444 1445 1446 1447 1448 1449 1450 1451 1452 1453 1454 1455 1456 1457 1458 1459 1460 1461 1462 1463 1464 1465 1466 1467 1468 1469 1470 1471 1472 1473 1474 1475 1476 1477 1478 1479 1480 1481 1482 1483 1484 1485 1486 1487 1488 1489 1490 1491 1492 1493 1494 1495 1496 1497 1498 1499 1500 1501 1502 1503 1504 1505 1506 1507 1508 1509 1510 1511 1512 1513 1514 1515 1516 1517 1518 1519 1520 1521 1522 1523 1524 1525 1526 1527 1528 1529 1530 1531 1532 1533 1534 1535 1536 1537 1538 1539 1540 1541 1542 1543 1544 1545 1546 1547 1548 1549 1550 1551 1552 1553 1554 1555 1556 1557 1558 1559 1560 1561 1562 1563 1564 1565 1566 1567 1568 1569 1570 1571 1572 1573 1574 1575 1576 1577 1578 1579 1580 1581 1582 1583 1584 1585 1586 1587 1588 1589 1590 1591 1592 1593 1594 1595 1596 1597 1598 1599 1600 1601 1602 1603 1604 1605 1606 1607 1608 1609 1610 1611 1612 1613 1614 1615 1616 1617 1618 1619 1620 1621 1622 1623 1624 1625 1626 1627 1628 1629 1630 1631 1632 1633 1634 1635 1636 1637 1638 1639 1640 1641 1642 1643 1644 1645 1646 1647 1648 1649 1650 1651 1652 1653 1654 1655 1656 1657 1658 1659 1660 1661 1662 1663 1664 1665 1666 1667 1668 1669 1670 1671 1672 1673 1674 1675 1676 1677 1678 1679 1680 1681 1682 1683 1684 1685 1686 1687 1688 1689 1690 1691 1692 1693 1694 1695 1696 1697 1698 1699 1700 1701 1702 1703 1704 1705 1706 1707 1708 1709 1710 1711 1712 1713 1714 1715 1716 1717 1718 1719 1720 1721 1722 1723 1724 1725 1726 1727 1728 1729 1730 1731 1732 1733 1734 1735 1736 1737 1738 1739 1740 1741 1742 1743 1744 1745 1746 1747 1748 1749 1750 1751 1752 1753 1754 1755 1756 1757 1758 1759 1760 1761 1762 1763 1764 1765 1766 1767 1768 1769 1770 1771 1772 1773 1774 1775 1776 1777 1778 1779 1780 1781 1782 1783 1784 1785 1786 1787 1788 1789 1790 1791 1792 1793 1794 1795 1796 1797 1798 1799 1800 1801 1802 1803 1804 1805 1806 1807 1808 1809 1810 1811 1812 1813 1814 1815 1816 1817 1818 1819 1820 1821 1822 1823 1824 1825 1826 1827 1828 1829 1830 1831 1832 1833 1834 1835 1836 1837 1838 1839 1840 1841 1842 1843 1844 1845 1846 1847 1848 1849 1850 1851 1852 1853 1854 1855 1856 1857 1858 1859 1860 1861 1862 1863 1864 1865 1866 1867 1868 1869 1870 1871 1872 1873 1874 1875 1876 1877 1878 1879 1880 1881 1882 1883 1884 1885 1886 1887 1888 1889 1890 1891 1892 1893 1894 1895 1896 1897 1898 1899 1900 1901 1902 1903 1904 1905 1906 1907 1908 1909 1910 1911 1912 1913 1914 1915 1916 1917 1918 1919 1920 1921 1922 1923 1924 1925 1926 1927 1928 1929 1930 1931 1932 1933 1934 1935 1936 1937 1938 1939 1940 1941 1942 1943 1944 1945 1946 1947 1948 1949 1950 1951 1952 1953 1954 1955 1956 1957 1958 1959 1960 1961 1962 1963 1964 1965 1966 1967 1968 1969 1970 1971 1972 1973 1974 1975 1976 1977 1978 1979 1980 1981 1982 1983 1984 1985 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020 2021 2022 2023 2024 2025 2026 2027 2028 2029 2030 2031 2032 2033 2034 2035 2036 2037 2038 2039 2040 2041 2042 2043 2044 2045 2046 2047 2048 2049 2050 2051 2052 2053 2054 2055 2056 2057 2058 2059 2060 2061 2062 2063 2064 2065 2066 2067 2068 2069 2070 2071 2072 2073 2074 2075 2076 2077 2078 2079 2080 2081 2082 2083 2084 2085 2086 2087 2088 2089 2090 2091 2092 2093 2094 2095 2096 2097 2098 2099 | from __future__ import annotations
from typing import Any, Optional, runtime_checkable
from dataclasses import dataclass, field, replace
import abc
import os
import draccus
import logging
import torch
import torchvision
import torch.nn.functional as F
import numpy as np
from src.transforms.utils import resize_with_pad, resize_center_crop
from src.utils import env_flags as _env_flags # registered LABVLA_* flags
from src.utils.constants import OBS_IMAGE, OBS_IMAGES, OBS_STATE, ACTION, NUM_IMAGE_SLOTS
# All hydration goes through hydrate_all() with a DatasetSchema.
DataDict = dict[str, Any]
_SOURCE_FAST_MAX_LENGTHS: dict[str, int] = {
# Keep schema-explicit budget hooks, but do not tighten 4ds action sources
# below the launcher union cap. Memory savings come from ``trim_to_mask``
# emitting the real per-sample token length; lowering the tokenizer cap
# itself can truncate FAST labels, which changes the CE target.
"robointer_droid_v2": 224,
"robointer_droid_anno_v1": 224,
"oxe_auge_v1": 224,
"agibot_dual_arm_v1": 224,
}
def source_fast_max_length(schema_id: str, configured_max_length: int) -> int:
"""Return the Phase-C FAST cap for a schema without widening user config."""
budget = _SOURCE_FAST_MAX_LENGTHS.get(str(schema_id))
if budget is None:
return int(configured_max_length)
return min(int(configured_max_length), int(budget))
def _dual_arm_canonical_indices(
layout, source_dims: tuple[int, ...]
) -> "list[int | None] | None":
"""Raw [left joints, right joints, left grip, right grip] β 14-dim indices.
Returns a length-14 list of indices into the raw concatenated vector, one
per canonical slot:
0..5 left joints, 6 left gripper, 7..12 right joints, 13 right gripper.
Sub-6-DoF arms produce zero-pad canonical slots with NO raw index β those
entries are ``None`` (sentinel stats are substituted by the caller,
mirroring the single-arm path), so the index list always has the canonical
14-dim width rather than silently producing narrower stats arrays.
"""
if layout is None:
return None
try:
from src.schema.arm_layout import ArmCount
except Exception:
return None
if getattr(layout, "arm_count", None) != ArmCount.DUAL:
return None
left_dof = int(layout.left_arm_dof)
right_dof = int(layout.right_arm_dof)
raw_width = sum(int(d) for d in source_dims)
if raw_width <= 0:
return None
left_grip = int(layout.left_gripper_index_in_raw)
right_grip = int(layout.right_gripper_index_in_raw)
right_start = left_dof
left_keep = min(left_dof, 6)
right_keep = min(right_dof, 6)
indices: list[int | None] = (
list(range(0, left_keep))
+ [None] * (6 - left_keep) # zero-pad slots (sentinel stats)
+ [left_grip]
+ list(range(right_start, right_start + right_keep))
+ [None] * (6 - right_keep) # zero-pad slots (sentinel stats)
+ [right_grip]
)
real = [i for i in indices if i is not None]
if not real or max(real) >= raw_width:
return None
return indices
def _canonicalize_dual_arm_stats(stats: dict | None, schema) -> dict | None:
"""Reorder raw dual-arm stats to the schema's 14-dim canonical layout."""
if not stats or schema is None:
return stats
layout = getattr(schema, "arm_layout", None)
state_idx = _dual_arm_canonical_indices(
layout, tuple(getattr(schema, "source_state_dims", ()) or ())
)
action_idx = _dual_arm_canonical_indices(
layout, tuple(getattr(schema, "source_action_dims", ()) or ())
)
if state_idx is None and action_idx is None:
return stats
# Sentinel stats for zero-pad canonical slots (mirrors the single-arm
# path): std=1 avoids div-by-zero in mean_std, q01=-1/q99=+1 maps the
# constant 0 input to a stable normalized 0.
_SENTINEL = {"mean": 0.0, "std": 1.0, "q01": -1.0, "q99": 1.0}
def _remap_entry(entry: dict, indices: "list[int | None] | None") -> dict:
if indices is None:
return entry
out = dict(entry)
real = [i for i in indices if i is not None]
if not real:
return entry
max_idx = max(real)
for stat_name, value in entry.items():
try:
if len(value) <= max_idx:
continue
except TypeError:
continue
arr = np.asarray(value)
if arr.ndim != 1:
continue
new = []
for idx in indices:
if idx is None:
# zero-pad slot β no raw index to copy; substitute the
# neutral sentinel so the output keeps canonical width.
new.append(float(_SENTINEL.get(stat_name, 0.0)))
else:
new.append(float(arr[idx]))
out[stat_name] = new
return out
canonicalized = dict(stats)
for key, indices in (
(OBS_STATE, state_idx),
(ACTION, action_idx),
("action_abs", action_idx),
):
entry = stats.get(key)
if isinstance(entry, dict):
canonicalized[key] = _remap_entry(entry, indices)
return canonicalized
_LAYOUT_CANON_MARKER = "_layout_canonicalization"
_WARNED_LEGACY_LAYOUT_MARKER = False
_GRIPPER_CANON_MARKER = "_gripper_canonicalization"
def _canonicalize_single_arm_stats(stats: dict | None, schema) -> dict | None:
"""Remap raw single-arm stats to the canonical 8-dim layout.
Mirror of ``_canonicalize_dual_arm_stats`` for the single-arm path
that ``CanonicalSingleArmLayoutTransformFn`` builds at training time.
Raw layout: dims 0..arm_dof-1 are arm joints, dim ``gripper_index_in_raw``
is the scalar gripper, any trailing dims are redundant mirrors.
Canonical layout: dims 0..6 = arm (zero-padded if arm_dof<7),
dim 7 = gripper.
For zero-pad dims (e.g. UR/festo 6-DoF β canonical dim 6) there is
no raw index to copy from; the transform fills dim 6 with 0.0 at
every frame. The stats for that dim are set to a neutral sentinel
``mean=0, std=1, q01=-1, q99=+1`` so q01_q99 normalization of the
constant 0 input yields a stable 0 (without div-by-zero).
No-op when:
- schema has no arm_layout or arm_count != SINGLE
- schema has no ``source_state_keys`` (raw == canonical already, no
remap needed; e.g. Franka 8-dim)
- stats already carry ``_layout_canonicalization`` marker (pre-canonicalized
on disk by an offline stats-canonicalization tool).
"""
if not stats or schema is None:
return stats
# Idempotency: skip if stats were already layout-canonicalized on disk.
# The marker is trusted as a behavior switch, so a structured (dict) marker
# must match the ACTIVE schema's identity and source geometry; a mismatched
# sidecar fails loud instead of silently feeding wrong-layout quantiles to
# normalization. Legacy boolean markers (pre-binding artifacts on disk) are
# accepted with a one-time warning.
_marker = stats.get(_LAYOUT_CANON_MARKER)
if _marker:
if isinstance(_marker, dict):
_mk_sid = _marker.get("schema_id")
if _mk_sid and _mk_sid != str(getattr(schema, "schema_id", "")):
raise ValueError(
f"[stats] _layout_canonicalization marker was produced for "
f"schema_id={_mk_sid!r} but the active schema is "
f"{getattr(schema, 'schema_id', None)!r} β refusing to "
f"trust pre-canonicalized stats across schemas (H26)."
)
for _fld, _attr in (("source_state_dims", "source_state_dims"),
("source_action_dims", "source_action_dims")):
_mk_v = _marker.get(_fld)
_cur = [int(d) for d in (getattr(schema, _attr, ()) or ())]
if _mk_v is not None and list(_mk_v) != _cur:
raise ValueError(
f"[stats] _layout_canonicalization marker {_fld}="
f"{_mk_v} does not match the active schema's "
f"{_cur} (schema_id={schema.schema_id!r}) β the "
f"sidecar was computed for a different source "
f"layout (H26)."
)
else:
global _WARNED_LEGACY_LAYOUT_MARKER
if not _WARNED_LEGACY_LAYOUT_MARKER:
logging.getLogger(__name__).warning(
"[stats] legacy boolean _layout_canonicalization marker "
"(no schema/geometry binding) β accepted for backward "
"compatibility; regenerate stats with the current "
"`python -m data_process stats` to get a bound marker "
"(H26). Warned once."
)
_WARNED_LEGACY_LAYOUT_MARKER = True
return stats
layout = getattr(schema, "arm_layout", None)
if layout is None:
return stats
try:
from src.schema.arm_layout import ArmCount
except Exception:
return stats
if getattr(layout, "arm_count", None) != ArmCount.SINGLE:
return stats
# If schema doesn't declare a raw source layout, stats are already
# in canonical orientation (e.g. Franka 8-dim).
if not getattr(schema, "source_state_keys", None) and not getattr(
schema, "source_action_keys", None
):
return stats
arm_dof = int(getattr(layout, "arm_dof", 0))
grip_idx_raw = getattr(layout, "gripper_index_in_raw", None)
if grip_idx_raw is None or arm_dof <= 0:
return stats
grip_idx = int(grip_idx_raw)
arm_keep = min(arm_dof, 7)
pad_count = 7 - arm_keep
# Per-slot index into the raw vector. None = zero-pad slot (no raw idx).
# Length is always 8 (canonical width).
indices: list[int | None] = (
list(range(arm_keep)) # canonical 0..arm_keep-1 β raw 0..arm_keep-1
+ [None] * pad_count # canonical arm_keep..6 (zero-pad β sentinel stats)
+ [grip_idx] # canonical 7 β raw[grip_idx]
)
# Sentinel stats for zero-pad dim. Keep std=1 to avoid div-by-zero in
# the eventual (x-mean)/std path; q01/q99 spans 2 so (0 - q01)/(q99-q01)
# = 0.5 normalized β neutral.
_SENTINEL = {"mean": 0.0, "std": 1.0, "q01": -1.0, "q99": 1.0}
def _remap_entry(entry: dict, raw_width: int) -> dict:
out = dict(entry)
for stat_name, value in entry.items():
try:
if len(value) <= max(i for i in indices if i is not None):
# raw stats too narrow to slice with our grip_idx β bail
continue
except TypeError:
continue
arr = np.asarray(value)
if arr.ndim != 1:
continue
new = []
for idx in indices:
if idx is None:
new.append(float(_SENTINEL.get(stat_name, 0.0)))
else:
if idx >= arr.shape[0]:
# Stats narrower than expected raw β bail without
# corrupting; the dataset will fail its strict checks
# downstream with a clearer error.
return entry
new.append(float(arr[idx]))
out[stat_name] = new
return out
canonicalized = dict(stats)
for key in (OBS_STATE, ACTION, "action_abs"):
entry = stats.get(key)
if isinstance(entry, dict):
# raw_width here is informational β we let _remap_entry inspect
# each stat array's length individually.
canonicalized[key] = _remap_entry(entry, raw_width=0)
return canonicalized
class DataTransformFn(draccus.ChoiceRegistry, abc.ABC):
@abc.abstractmethod
def __call__(self, data: DataDict) -> DataDict: ...
def hydrate(self, ctx: "HydrateContext") -> "DataTransformFn":
"""Return a schema/stats-hydrated copy of self (default: unchanged).
Called once per transform in chain order; implementations may READ AND
UPDATE ``ctx.stats`` (``GripperSemanticCanonicalizeFn`` rewrites
gripper-dim stats that downstream Normalize/Snap hydration must
observe). Transforms with no schema needs simply inherit this no-op.
"""
return self
@dataclass(frozen=True)
class TransformGroup:
"""A group of transforms."""
# Transforms that are applied to the model input data.
inputs: list[DataTransformFn] = field(default_factory=list)
# Transforms that are applied to the model output data.
outputs: list[DataTransformFn] = field(default_factory=list)
def push(self,
*,
inputs: list[DataTransformFn] = None,
outputs: list[DataTransformFn] = None) -> TransformGroup:
"""Append transforms to the group and return a new group.
Both inputs and outputs are appended to the *end* (FIFO order):
earlier pushes run first.
Returns:
A new group with the appended transforms.
"""
if inputs is None: inputs = []
if outputs is None: outputs = []
return TransformGroup(
inputs=[*self.inputs, *inputs],
outputs=[*self.outputs, *outputs],
)
@DataTransformFn.register_subclass("composite")
@dataclass(frozen=True)
class CompositeTransform(DataTransformFn):
"""A composite transform that applies a sequence of transforms in order."""
transforms: list[DataTransformFn]
def __call__(self, data: DataDict) -> DataDict:
for transform in self.transforms:
data = transform(data)
return data
def compose(transforms: list[DataTransformFn]) -> DataTransformFn:
"""Compose a sequence of transforms into a single transform."""
return CompositeTransform(transforms)
@DataTransformFn.register_subclass("identity")
@dataclass(frozen=True)
class IdentityTransformFn(DataTransformFn):
def __call__(self, data: DataDict) -> DataDict:
return data
@DataTransformFn.register_subclass("gripper_semantic_canonicalize")
@dataclass
class GripperSemanticCanonicalizeFn(DataTransformFn):
"""Convert gripper dim from source semantic (width / position) to the
canonical continuous ``open_fraction`` β [0, 1] target.
This transform is currently DORMANT β it is intentionally NOT inserted
into any chain (q01/q99 normalization already aligns width/open_fraction
sources, and per-robot calibration for UR/Festo/Rizon4 is unknown). The
hydrate branch and the stats canonicalizer below are maintained so the
transform is correct if re-enabled with proper calibration entries.
Mathematical mapping:
open_fraction = clip((x - closed) / (open - closed), 0, 1)
if direction == -1: open_fraction = 1 - open_fraction
where ``closed``, ``open``, ``direction`` are schema-level calibration
constants for the source robot family. After this transform every
source's gripper dim shares the SAME normalized semantic β fully
closed = 0.0, fully open = 1.0 β so cross-source posttrain MSE on
the gripper dim is no longer a contradictory target. The downstream
``NormalizeTransformFn`` should then use open_fraction-domain stats
(q01=0, q99=1) on the gripper dim.
Enabled only when the schema declares a non-empty
``source_semantic`` distinct from ``open_fraction``. For
``open_fraction`` sources (e.g. OXE-Auge) this transform is a no-op.
"""
enabled: bool = False
source_semantic: str = ""
target_semantic: str = "open_fraction"
closed: float = 0.0
open: float = 1.0
direction: int = 1 # +1: xβ β ofβ; -1: xβ β ofβ
gripper_dim: int = 7
state_keys: tuple = ()
action_keys: tuple = ()
def __call__(self, data: DataDict) -> DataDict:
if not self.enabled:
return data
if self.source_semantic == self.target_semantic:
return data
span = float(self.open) - float(self.closed)
if not (span > 0 or span < 0):
return data
for key in (*self.state_keys, *self.action_keys):
tensor = data.get(key)
if tensor is None or not hasattr(tensor, "shape"):
continue
if tensor.shape[-1] <= self.gripper_dim:
continue
new_tensor = tensor.clone() if hasattr(tensor, "clone") else tensor.copy()
raw = tensor[..., self.gripper_dim]
of = (raw - float(self.closed)) / span
if int(self.direction) < 0:
of = 1.0 - of
of = of.clamp(0.0, 1.0) if hasattr(of, "clamp") else of.clip(0.0, 1.0)
new_tensor[..., self.gripper_dim] = of
data[key] = new_tensor
return data
def hydrate(self, ctx: "HydrateContext") -> "GripperSemanticCanonicalizeFn":
# Convert source gripper semantic (width / position) to the canonical
# continuous open_fraction target so multi-source posttrain MSE sees a
# unified gripper signal across all repos. NOTE: rewrites ctx.stats
# when enabled β downstream Normalize/Snap hydration must see the
# open_fraction-domain gripper stats.
schema = ctx.schema
src_sem = getattr(schema, "gripper_semantic", None)
grip_dims = tuple(getattr(schema, "gripper_action_dims", ()) or ())
grip_dim = int(grip_dims[0]) if grip_dims else 7
# Hardcoded calibration tables (extend here as new sources
# come in with calibrated endpoints):
# width β open_fraction with Franka spec [0, 0.04]
# open_fraction β open_fraction is no-op (enabled=False)
_CALIBRATION = {
"width": dict(closed=0.0, open=0.04, direction=+1),
"open_fraction": None, # already canonical
}
cal = _CALIBRATION.get(src_sem)
# A declared-but-uncalibrated semantic ("position", "velocity",
# "binary") disables the transform. Surface it once so operators know
# the gripper stays in its source semantic (the cross-repo mix guard in
# scripts/utils/dataset_helpers.py is then the only protection).
if src_sem and src_sem not in _CALIBRATION:
from src.utils.logging_utils import warn_once
warn_once(
logging.getLogger(__name__),
("gripper_semantic_no_calibration", src_sem, schema.schema_id),
"[hydrate_all] schema %s declares gripper_semantic=%r but "
"GripperSemanticCanonicalizeFn has no calibration entry for "
"it β the transform stays DISABLED and the gripper remains "
"in its source semantic. Add a _CALIBRATION entry if "
"canonicalization to open_fraction is intended.",
schema.schema_id, src_sem,
)
enabled = bool(src_sem and cal is not None and src_sem != "open_fraction")
if enabled:
t = replace(
self,
enabled=True,
source_semantic=src_sem,
target_semantic="open_fraction",
closed=float(cal["closed"]),
open=float(cal["open"]),
direction=int(cal["direction"]),
gripper_dim=grip_dim,
state_keys=tuple(schema.state_keys),
action_keys=tuple(schema.action_keys),
)
# Apply the same affine to gripper-dim stats so normalize
# downstream uses open_fraction-domain q01/q99/mean/std.
_calibration_for_stats = dict(
enabled=True,
source_semantic=src_sem,
target_semantic="open_fraction",
closed=float(cal["closed"]),
open=float(cal["open"]),
direction=int(cal["direction"]),
gripper_dim=grip_dim,
)
ctx.stats = _canonicalize_gripper_semantic_stats(
ctx.stats, schema, _calibration_for_stats
)
else:
t = replace(self, enabled=False, source_semantic=src_sem or "")
logging.info(
f"Hydrated {t.__class__.__name__} enabled={t.enabled} "
f"source_semantic={t.source_semantic} target=open_fraction "
f"({schema.schema_id})"
)
return t
def _canonicalize_gripper_semantic_stats(
stats: dict | None, schema, calibration: dict | None
) -> dict | None:
"""Rescale gripper-dim stats from source semantic to open_fraction.
Mirrors what GripperSemanticCanonicalizeFn does to the data, but on
the q01/q99/mean/std arrays. Must run AFTER any layout canonicalization
so the gripper is already at canonical dim 7.
For source_semantic="width" (LabUtopia Franka):
new[7] = (old[7] - closed) / (open - closed)
For source_semantic="open_fraction" (OXE-Auge): no-op
"""
if not stats or schema is None or not calibration:
return stats
if not calibration.get("enabled"):
return stats
# Idempotency: skip if stats were already gripper-canonicalized on disk.
if stats.get(_GRIPPER_CANON_MARKER):
return stats
src = calibration.get("source_semantic", "")
tgt = calibration.get("target_semantic", "open_fraction")
if src == tgt or not src:
return stats
closed = float(calibration["closed"])
open_v = float(calibration["open"])
direction = int(calibration.get("direction", 1))
span = open_v - closed
if span == 0:
return stats
grip_dim = int(calibration.get("gripper_dim", 7))
def _remap_value(v):
of = (float(v) - closed) / span
if direction < 0:
of = 1.0 - of
return of
def _remap_entry(entry: dict) -> dict:
out = dict(entry)
def _has(arr) -> bool:
return isinstance(arr, (list, tuple)) and len(arr) > grip_dim
# Location stats get the affine map. min/max are included too, so a
# min_max-mode consumer is not left with source-domain bounds after
# canonicalization.
for stat_name in ("q01", "q99", "mean", "min", "max"):
arr = entry.get(stat_name)
if not _has(arr):
continue
new = list(arr)
new[grip_dim] = _remap_value(new[grip_dim])
out[stat_name] = new
# std scales by |1/span| (affine multiplier); direction sign drops
std = entry.get("std")
if _has(std):
new = list(std)
new[grip_dim] = float(new[grip_dim]) / abs(span)
out["std"] = new
# A DECREASING affine (direction=-1) swaps order statistics β the
# mapped q01 becomes the upper quantile and vice versa. Without the
# swap, stored q01 > q99 and downstream q01/q99 normalization inverts
# the gripper axis relative to direction=+1 sources, which is exactly
# the cross-source contradiction this canonicalizer exists to remove.
# Same for min/max.
if direction < 0:
for lo_name, hi_name in (("q01", "q99"), ("min", "max")):
lo, hi = out.get(lo_name), out.get(hi_name)
if _has(lo) and _has(hi):
lo, hi = list(lo), list(hi)
lo[grip_dim], hi[grip_dim] = hi[grip_dim], lo[grip_dim]
out[lo_name], out[hi_name] = lo, hi
return out
canon = dict(stats)
for key in (OBS_STATE, ACTION, "action_abs"):
e = stats.get(key)
if isinstance(e, dict):
canon[key] = _remap_entry(e)
# Persist the idempotency marker the entry guard checks β without it a
# second application would re-apply the affine rescale to already-canonical
# stats.
canon[_GRIPPER_CANON_MARKER] = True
return canon
@DataTransformFn.register_subclass("canonical_arm_layout")
@dataclass
class CanonicalArmLayoutTransformFn(DataTransformFn):
"""Materialize canonical state/action vectors from raw source columns.
AgiBot stores raw dual-arm action as 14 joints plus 2 effectors:
``[left_7, right_7] + [left_gripper, right_gripper]``. LabVLA's existing
dual-arm canonical layout is 14 dims:
``[left_6, left_gripper, right_6, right_gripper]``. This transform creates
the canonical ``observation.state`` and ``action`` tensors before
normalization, state discretization, FAST tokenization, and padding.
"""
enabled: bool = False
left_arm_dof: int = 0
right_arm_dof: int = 0
left_gripper_index_in_raw: int = -1
right_gripper_index_in_raw: int = -1
state_source_keys: tuple[str, ...] = ()
action_source_keys: tuple[str, ...] = ()
def __call__(self, data: DataDict) -> DataDict:
if not self.enabled:
return data
self._materialize(data, self.state_source_keys, OBS_STATE)
self._materialize(data, self.action_source_keys, ACTION)
pad_keys = [f"{k}_is_pad" for k in self.action_source_keys if f"{k}_is_pad" in data]
if ACTION in data and pad_keys and "action_is_pad" not in data:
merged = data[pad_keys[0]].clone()
for key in pad_keys[1:]:
merged = merged | data[key]
data["action_is_pad"] = merged
return data
def _materialize(self, data: DataDict, source_keys: tuple[str, ...], target_key: str) -> None:
if not source_keys or any(k not in data for k in source_keys):
return
pieces = []
for key in source_keys:
value = data[key]
if not isinstance(value, torch.Tensor):
value = torch.as_tensor(value, dtype=torch.float32)
pieces.append(value.to(dtype=torch.float32))
raw = torch.cat(pieces, dim=-1)
left_grip = int(self.left_gripper_index_in_raw)
right_grip = int(self.right_gripper_index_in_raw)
if raw.shape[-1] <= max(left_grip, right_grip):
raise ValueError(
f"CanonicalArmLayoutTransformFn raw width {raw.shape[-1]} is "
f"too small for gripper indices ({left_grip}, {right_grip}) "
f"from source_keys={source_keys!r}"
)
left_keep = min(self.left_arm_dof, 6)
right_keep = min(self.right_arm_dof, 6)
right_start = int(self.left_arm_dof)
right_end = right_start + right_keep
if raw.shape[-1] < right_end:
raise ValueError(
f"CanonicalArmLayoutTransformFn raw width {raw.shape[-1]} is "
f"too small for right arm slice [{right_start}:{right_end}] "
f"from source_keys={source_keys!r}"
)
out_shape = raw.shape[:-1] + (14,)
out = raw.new_zeros(out_shape)
out[..., :left_keep] = raw[..., :left_keep]
out[..., 6] = raw[..., left_grip]
out[..., 7:7 + right_keep] = raw[..., right_start:right_end]
out[..., 13] = raw[..., right_grip]
data[target_key] = out
def hydrate(self, ctx: "HydrateContext") -> "CanonicalArmLayoutTransformFn":
schema = ctx.schema
layout = getattr(schema, "arm_layout", None)
try:
from src.schema.arm_layout import ArmCount
is_dual_arm_layout = getattr(layout, "arm_count", None) == ArmCount.DUAL
except Exception:
is_dual_arm_layout = False
left_gripper_index = getattr(layout, "left_gripper_index_in_raw", None)
right_gripper_index = getattr(layout, "right_gripper_index_in_raw", None)
enabled = bool(
layout is not None
and is_dual_arm_layout
and getattr(schema, "source_state_keys", ())
and getattr(schema, "source_action_keys", ())
and left_gripper_index is not None
and right_gripper_index is not None
)
t = replace(
self,
enabled=enabled,
left_arm_dof=int(getattr(layout, "left_arm_dof", 0) or 0),
right_arm_dof=int(getattr(layout, "right_arm_dof", 0) or 0),
left_gripper_index_in_raw=int(left_gripper_index or -1),
right_gripper_index_in_raw=int(right_gripper_index or -1),
state_source_keys=tuple(getattr(schema, "source_state_keys", ()) or ()),
action_source_keys=tuple(getattr(schema, "source_action_keys", ()) or ()),
)
logging.info(
f"Hydrated {t.__class__.__name__} enabled={t.enabled} "
f"state_sources={t.state_source_keys} action_sources={t.action_source_keys} "
f"({schema.schema_id})"
)
return t
@DataTransformFn.register_subclass("canonical_single_arm_layout")
@dataclass
class CanonicalSingleArmLayoutTransformFn(DataTransformFn):
"""Materialize canonical 8-dim single-arm state/action from raw N-dim source.
Counterpart to ``CanonicalArmLayoutTransformFn`` for single-arm datasets
whose raw layout has either a sub-7-DoF arm (e.g. UR/festo 6-DoF) or
trailing redundant multi-finger gripper joints that aggregate to one
scalar (e.g. UR/festo 11-dim, rizon4 12-dim where dims after
the gripper are perfectly correlated mirror copies of the single
open-width signal).
Canonical 8-dim layout (per src/schema/arm_layout.py):
dim 0..6 = arm joints (zero-padded if raw arm DoF < 7)
dim 7 = gripper (scalar copied from raw[..., raw_gripper_index_in_raw])
Trailing raw dims past ``raw_gripper_index_in_raw`` are intentionally
DROPPED β they are duplicate finger-joint copies of the single gripper
scalar (verified by cross-correlation == Β±1.0 against dim
``raw_gripper_index_in_raw``).
Mirrors the dual-arm transform's pad-aggregation behavior: if any
``<source_action_key>_is_pad`` tensors are present, they are OR-merged
into ``action_is_pad`` so downstream MSE masking stays correct.
"""
enabled: bool = False
raw_arm_dof: int = 0
raw_gripper_index_in_raw: int = -1
state_source_keys: tuple[str, ...] = ()
action_source_keys: tuple[str, ...] = ()
def __call__(self, data: DataDict) -> DataDict:
if not self.enabled:
return data
self._materialize(data, self.state_source_keys, OBS_STATE)
self._materialize(data, self.action_source_keys, ACTION)
pad_keys = [
f"{k}_is_pad" for k in self.action_source_keys
if f"{k}_is_pad" in data
]
if ACTION in data and pad_keys and "action_is_pad" not in data:
merged = data[pad_keys[0]].clone()
for key in pad_keys[1:]:
merged = merged | data[key]
data["action_is_pad"] = merged
return data
def _materialize(
self,
data: DataDict,
source_keys: tuple[str, ...],
target_key: str,
) -> None:
if not source_keys or any(k not in data for k in source_keys):
return
pieces = []
for key in source_keys:
value = data[key]
if not isinstance(value, torch.Tensor):
value = torch.as_tensor(value, dtype=torch.float32)
pieces.append(value.to(dtype=torch.float32))
raw = torch.cat(pieces, dim=-1)
grip = int(self.raw_gripper_index_in_raw)
if raw.shape[-1] <= grip:
raise ValueError(
f"CanonicalSingleArmLayoutTransformFn raw width {raw.shape[-1]} is "
f"too small for gripper index {grip} from source_keys={source_keys!r}"
)
arm_keep = min(int(self.raw_arm_dof), 7)
if raw.shape[-1] < arm_keep:
raise ValueError(
f"CanonicalSingleArmLayoutTransformFn raw width {raw.shape[-1]} is "
f"too small for arm slice [:{arm_keep}] from source_keys={source_keys!r}"
)
out_shape = raw.shape[:-1] + (8,)
out = raw.new_zeros(out_shape)
out[..., :arm_keep] = raw[..., :arm_keep]
out[..., 7] = raw[..., grip]
data[target_key] = out
def hydrate(self, ctx: "HydrateContext") -> "CanonicalSingleArmLayoutTransformFn":
schema = ctx.schema
layout = getattr(schema, "arm_layout", None)
try:
from src.schema.arm_layout import ArmCount
is_single_arm_layout = (
getattr(layout, "arm_count", None) == ArmCount.SINGLE
)
except Exception:
is_single_arm_layout = False
grip_idx = getattr(layout, "gripper_index_in_raw", None)
arm_dof = getattr(layout, "arm_dof", None)
enabled = bool(
layout is not None
and is_single_arm_layout
and getattr(schema, "source_state_keys", ())
and getattr(schema, "source_action_keys", ())
and grip_idx is not None
and arm_dof is not None
)
t = replace(
self,
enabled=enabled,
raw_arm_dof=int(arm_dof or 0),
raw_gripper_index_in_raw=int(grip_idx if grip_idx is not None else -1),
state_source_keys=tuple(
getattr(schema, "source_state_keys", ()) or ()
),
action_source_keys=tuple(
getattr(schema, "source_action_keys", ()) or ()
),
)
logging.info(
f"Hydrated {t.__class__.__name__} enabled={t.enabled} "
f"raw_arm_dof={t.raw_arm_dof} raw_grip_idx={t.raw_gripper_index_in_raw} "
f"state_sources={t.state_source_keys} action_sources={t.action_source_keys} "
f"({schema.schema_id})"
)
return t
@DataTransformFn.register_subclass("pad_state_and_action")
@dataclass
class PadStateAndActionTransformFn(DataTransformFn):
max_state_dim: int = 32
max_action_dim: int = 32
def __call__(self, data: DataDict) -> DataDict:
data[OBS_STATE] = self._pad_vector(data[OBS_STATE], self.max_state_dim)
data[ACTION] = self._pad_vector(data[ACTION], self.max_action_dim)
return data
def _pad_vector(self, vector: torch.Tensor, new_dim: int):
# Fail loud when the incoming vector already exceeds the
# padded target size. Silently returning as-is would
# let downstream models receive state/action wider than max_*_dim
# and either crash at a later layer or drop dims. Treat this as a
# schema mismatch, not a condition to paper over.
cur = vector.shape[-1]
if cur > new_dim:
raise ValueError(
f"Vector dim {cur} > pad target {new_dim}. Schema declares "
f"fewer dims than the runtime tensor β bump max_state_dim / "
f"max_action_dim or prune schema keys."
)
if cur == new_dim:
return vector
return F.pad(vector, (0, new_dim - cur))
@DataTransformFn.register_subclass("snap_gripper_to_endpoints")
@dataclass
class SnapGripperToEndpointsFn(DataTransformFn):
"""Threshold gripper width signal to bimodal endpoints {0, max_width}.
Use case: align LabUtopia continuous gripper width [0, 0.04m] to a binary
{0=closed, 0.04m=open} representation that matches the OXE pretrain
canonical convention (which is binary {0, 1} after gripper_canonicalize).
Pipeline ordering: AFTER ComposeFieldsTransform (which builds canonical
OBS_STATE / ACTION arrays) and BEFORE NormalizeTransformFn (so the snap
to {0, max_width} happens on raw width values, then normalize maps to
{-1, +1} cleanly when stats q01=0, q99=max_width).
Stats override: caller MUST ensure stats[OBS_STATE][q01][gripper_dim]=0
and stats[OBS_STATE][q99][gripper_dim]=max_width (same for ACTION) so
that q01/q99 normalize maps {0, max_width} β {-1, +1}. Otherwise the
snap-to-binary values get mapped through the wrong stats range and
deploy gets garbage gripper output.
Default Franka Panda config: gripper_dim=7, max_width=0.04m, threshold=0.5.
"""
gripper_dim: int = 7
max_width: float = 0.04
threshold_ratio: float = 0.5 # > threshold_ratio*max_width β max_width, else 0
state_keys: tuple[str, ...] = ()
action_keys: tuple[str, ...] = ()
state_dims: tuple[int, ...] = ()
action_dims: tuple[int, ...] = ()
def __call__(self, data: DataDict) -> DataDict:
threshold = self.threshold_ratio * self.max_width
applied = False
for key in (OBS_STATE, ACTION):
if key not in data:
continue
applied = self._snap_key(data, key, self.gripper_dim, threshold) or applied
if applied:
return data
self._snap_split_fields(
data,
keys=self.state_keys,
dims=self.state_dims,
threshold=threshold,
)
self._snap_split_fields(
data,
keys=self.action_keys,
dims=self.action_dims,
threshold=threshold,
)
return data
def _snap_split_fields(
self,
data: DataDict,
*,
keys: tuple[str, ...],
dims: tuple[int, ...],
threshold: float,
) -> bool:
offset = 0
for key, width in zip(keys, dims):
width = int(width)
if offset <= self.gripper_dim < offset + width:
return self._snap_key(data, key, self.gripper_dim - offset, threshold)
offset += width
return False
def _snap_key(
self,
data: DataDict,
key: str,
local_dim: int,
threshold: float,
) -> bool:
if key not in data:
return False
v = data[key]
if isinstance(v, torch.Tensor):
# Avoid in-place on shared tensor
v = v.clone()
if v.ndim == 0:
if int(local_dim) != 0:
return False
grip = v
v = torch.where(
grip > threshold,
torch.full_like(grip, self.max_width),
torch.zeros_like(grip),
)
else:
grip = v[..., local_dim]
v[..., local_dim] = torch.where(
grip > threshold,
torch.full_like(grip, self.max_width),
torch.zeros_like(grip),
)
elif isinstance(v, np.ndarray):
v = v.copy()
if v.ndim == 0:
if int(local_dim) != 0:
return False
v = np.asarray(self.max_width if bool(v > threshold) else 0.0, dtype=v.dtype)
else:
v[..., local_dim] = np.where(
v[..., local_dim] > threshold,
self.max_width,
0.0,
).astype(v.dtype)
else:
return False
data[key] = v
return True
def hydrate(self, ctx: "HydrateContext") -> "SnapGripperToEndpointsFn":
# Reads ctx.stats AFTER any GripperSemanticCanonicalizeFn rewrite
# earlier in the chain.
schema = ctx.schema
stats = ctx.stats
t = replace(
self,
state_keys=tuple(schema.state_keys),
action_keys=tuple(schema.action_keys),
state_dims=tuple(schema.state_dims),
action_dims=tuple(schema.action_dims),
)
logging.info(
f"Hydrated {t.__class__.__name__} gripper_dim={t.gripper_dim} "
f"({schema.schema_id})"
)
# snap_gripper_to_binary snaps raw gripper width to {0, max_width}
# BEFORE NormalizeTransformFn. For the intended exact {-1,+1}
# mapping, the gripper-dim stats MUST already be q01=0, q99=max_width.
# This is NOT done automatically β it requires running
# data_process/labutopia_canonicalize_stats.py on stats.json first.
# Verify and fail loud so a forgotten canonicalize step doesn't
# silently train a mis-scaled gripper. Bypass: LABVLA_ALLOW_UNPATCHED_SNAP_STATS=1.
_grip = int(getattr(t, "gripper_dim", -1))
_maxw = float(getattr(t, "max_width", 0.0))
_ack = "action_abs" if ctx.action_mode == "abs" else "action"
_tol = max(1e-4, abs(_maxw) * 0.05)
_bad_snap = []
for _can in ("observation.state", _ack):
_b = stats.get(_can)
if not isinstance(_b, dict):
continue
_q01 = _b.get("q01")
_q99 = _b.get("q99")
# snap-to-binary REQUIRES q01/q99 (it maps {0,max_width}->{-1,+1}
# via q01/q99 normalization). If they are absent/short,
# NormalizeTransformFn falls back to mean_std (core.py
# _resolve_with_fallback) β the snapped values are mis-scaled and
# the guard must NOT silently pass. Treat absent/short/unparseable
# q01/q99 as a failure too.
if _q01 is None or _q99 is None:
_bad_snap.append((_can, "q01/q99 absent (snap needs quantile stats)"))
continue
try:
if 0 <= _grip < len(_q01) and 0 <= _grip < len(_q99):
if abs(float(_q01[_grip])) > _tol or abs(float(_q99[_grip]) - _maxw) > _tol:
_bad_snap.append((_can, float(_q01[_grip]), float(_q99[_grip])))
else:
_bad_snap.append((_can, f"q01/q99 too short for gripper dim {_grip}"))
except (TypeError, ValueError, IndexError):
_bad_snap.append((_can, "q01/q99 unparseable"))
if _bad_snap:
_snap_msg = (
f"[hydrate_all] snap_gripper_to_binary=True but gripper-dim ({_grip}) "
f"stats are NOT canonicalized to q01=0 / q99={_maxw} (found "
f"(canonical, q01, q99) = {_bad_snap}). The snapped {{0, max_width}} "
f"values would be normalized through the dataset's raw q01/q99 instead "
f"of to exactly {{-1, +1}}, mis-scaling the gripper at train AND deploy. "
f"Run data_process/labutopia_canonicalize_stats.py on this dataset's "
f"stats.json before training."
)
if _env_flags.get("LABVLA_ALLOW_UNPATCHED_SNAP_STATS") == "1":
logging.error(_snap_msg + " [BYPASSED via LABVLA_ALLOW_UNPATCHED_SNAP_STATS=1]")
else:
raise ValueError(
_snap_msg + " Set LABVLA_ALLOW_UNPATCHED_SNAP_STATS=1 to bypass intentionally."
)
return t
@DataTransformFn.register_subclass("totensor")
@dataclass
class ToTensorTransformFn(DataTransformFn):
def __post_init__(self):
self.img2tensor_fn = torchvision.transforms.ToTensor()
def __call__(self, data: DataDict) -> DataDict:
for key in data.keys():
# Tight image-key guard: a loose ``"image" in key`` substring
# match would misfire on unrelated keys like
# ``action_image_mask`` or ``task.image_target``. Match only explicit
# image slots: OBS_IMAGES prefix or the singleton OBS_IMAGE key,
# matching the tighter guards used by ResizeImagesWithPadFn and
# ResizeShortestCenterCropFn.
if key.startswith(OBS_IMAGES) or key == OBS_IMAGE:
data[key] = self.img2tensor_fn(data[key])
elif isinstance(data[key], list):
data[key] = torch.tensor(data[key])
elif isinstance(data[key], np.ndarray):
data[key] = torch.from_numpy(data[key])
return data
# Sentinel/companion suffixes carried alongside image tensors. Keys ending in
# these MUST NOT enter the resize path even though they include "image" as a
# substring (e.g. observation.images.primary_invalid is a bool, _is_pad is a
# bool tensor, _mask is an int per-vision-token vector). A loose
# `or "image" in k` would match these and crash at resize_with_pad.
_NON_IMAGE_KEY_SUFFIXES = ("_invalid", "_is_pad", "_mask")
def _is_image_tensor_key(k: str) -> bool:
if any(k.endswith(suffix) for suffix in _NON_IMAGE_KEY_SUFFIXES):
return False
return k.startswith(OBS_IMAGES) or k == OBS_IMAGE or "image" in k
@DataTransformFn.register_subclass("resize_with_pad")
@dataclass
class ResizeImagesWithPadFn(DataTransformFn):
height: int
width: int
mode: str = "bilinear"
def __call__(self, data: DataDict) -> DataDict:
for k, v in data.items():
if _is_image_tensor_key(k):
data[k] = resize_with_pad(v, self.height, self.width, self.mode)
return data
@DataTransformFn.register_subclass("resize_center_crop")
@dataclass
class ResizeShortestCenterCropFn(DataTransformFn):
height: int
width: int
mode: str = "bilinear"
def __call__(self, data: DataDict) -> DataDict:
for k, v in data.items():
if _is_image_tensor_key(k):
data[k] = resize_center_crop(v, self.height, self.width, self.mode)
return data
@DataTransformFn.register_subclass("compose_fields")
@dataclass
class ComposeFieldsTransform(DataTransformFn):
"""
Merge multiple keys' values into a single new key.
Example:
mapping = {
"observation.state": [
"observation.states.joint.position",
"observation.states.effector.position",
]
"action": [
"actions.joint.position",
"actions.effector.position",
]
}
"""
mapping: dict[str, list[str]] = field(default_factory=dict)
def __call__(self, data: DataDict) -> DataDict:
# Two-pass: first read + compute all merges, then write + pop. This
# handles the case where the SAME src_key appears in multiple mappings.
# No CURRENT registered schema overlaps state_keys with action_keys
# (oxe-auge uses observation.state vs observation.joints), and
# schema/validate.py REJECTS such overlap outright because the Delta
# transform would overwrite the shared key with the delta chunk. The
# two-phase order is kept as cheap defense in depth.
merged_by_new_key: dict[str, torch.Tensor] = {}
keys_to_pop: set[str] = set()
for new_key, src_keys in self.mapping.items():
if len(src_keys) == 1 and src_keys[0] == new_key:
continue
merge_list = self._align_for_cat([data[k] for k in src_keys])
merged_by_new_key[new_key] = torch.cat(merge_list, dim=-1)
for k in src_keys:
if k != new_key:
keys_to_pop.add(k)
# Write merged values. new_keys override src keys; don't pop any
# src_key that is itself a new_key (we just wrote it).
for new_key, merged in merged_by_new_key.items():
data[new_key] = merged
for k in keys_to_pop:
if k not in merged_by_new_key:
data.pop(k, None)
return data
def _align_for_cat(self, tensors: list[torch.Tensor], dim=-1) -> list[torch.Tensor]:
max_ndim = max((t.ndim for t in tensors))
out = []
for t in tensors:
t = t if t.ndim == max_ndim else t.unsqueeze(dim)
out.append(t)
return out
def hydrate(self, ctx: "HydrateContext") -> "ComposeFieldsTransform":
t = replace(self, mapping=ctx.feature_map)
logging.info(f"Hydrated {t.__class__.__name__} ({ctx.schema.schema_id})")
return t
@DataTransformFn.register_subclass("remap_image_key")
@dataclass
class RemapImageKeyTransformFn(DataTransformFn):
"""Remap image keys to unified `observation.images.imageN` slots.
Downstream `Qwen3_VLProcessorTransformFn` iterates over exactly
``num_image_slots`` slots. If the schema declares fewer cameras, we fill
the missing slots with zero-tensors AND `_mask=False` so normalization
doesn't corrupt real data with artificial zero-centered padding signal.
The slot count is explicit and mismatches fail loud at hydrate_all time,
rather than silently dropping cameras when the schema declares more or
fabricating dummy slots when it declares fewer.
Example::
mapping = {
"images.rgb.head": f"{OBS_IMAGES}.image0",
"images.rgb.hand_left": f"{OBS_IMAGES}.image1",
"images.rgb.hand_right": f"{OBS_IMAGES}.image2",
}
"""
mapping: dict[str, str] = field(default_factory=dict)
# Hard upper bound on image slots consumed by Qwen3_VLProcessorTransformFn.
# Authored in lockstep with the processor's `for i in range(3):` loop; if
# the processor is extended to N slots, bump this (+ hydrate_all guard).
num_image_slots: int = NUM_IMAGE_SLOTS
def __call__(self, data: DataDict) -> DataDict:
if len(self.mapping) > self.num_image_slots:
raise ValueError(
f"RemapImageKeyTransformFn: schema declares "
f"{len(self.mapping)} cameras but only {self.num_image_slots} "
f"slots are supported by the VLM processor. Drop the extras "
f"from the schema's image_mapping or extend the processor."
)
for old_key, new_key in self.mapping.items():
data[new_key] = data.pop(old_key)
# Respect the adapter's `<old_key>_invalid` sentinel.
# When _read_video_frame fell back to _zero_frame() (missing
# mp4, decode error, index overshoot), the upstream adapter
# tags the camera as invalid. Mark the slot mask=False so the
# VLM processor treats it as padded β never as a real frame.
invalid_key = f"{old_key}_invalid"
is_valid = not bool(data.pop(invalid_key, False))
data[f"{new_key}_mask"] = torch.tensor(is_valid, dtype=torch.bool)
# Pad missing slots with zero tensors + False masks. Zero-images (not
# white) keep normalization statistics benign: after standardization
# a zero tensor maps to -mean/std, which the model quickly learns to
# ignore via the mask=False signal.
for i in range(len(self.mapping), self.num_image_slots):
slot = f"{OBS_IMAGES}.image{i}"
if slot not in data:
# Safe reference image: first populated slot (image0 if the
# schema declared β₯1 camera β guaranteed by upstream validation
# that rejects zero-camera schemas in hydrate_all).
ref_slot = f"{OBS_IMAGES}.image0"
data[slot] = torch.zeros_like(data[ref_slot])
data[f"{slot}_mask"] = torch.tensor(False, dtype=torch.bool)
return data
def hydrate(self, ctx: "HydrateContext") -> "RemapImageKeyTransformFn":
schema = ctx.schema
t = replace(self, mapping=dict(schema.image_mapping))
logging.info(
f"Hydrated {t.__class__.__name__} with {len(schema.image_mapping)} "
f"cameras ({schema.schema_id})"
)
return t
def _resolve_mode(key: str, default_mode: str, mode_overrides: dict[str, str]) -> str:
"""Pick normalization mode for a key. If any override-substring matches the key,
that mode wins; otherwise returns default_mode.
Checks for an exact key match first (O(1) dict lookup) β the hot path is
callers pre-computing mode_overrides keyed by full-key, and the substring
scan only exists for legacy "gripper"-style partial overrides. Exact match
wins over substring; among substrings, first match (insertion order) wins.
"""
# O(1) fast path: exact key match.
exact = mode_overrides.get(key)
if exact is not None:
return exact
# Legacy substring fallback (e.g. {"gripper": "q01_q99"}).
for substr, override_mode in mode_overrides.items():
if substr in key:
return override_mode
return default_mode
def _stats_has_q01_q99(stats: dict) -> bool:
return "q01" in stats and "q99" in stats
# Deduplication is handled by utils.logging_utils.DedupeFilter installed at
# process startup (no module-level warned-set needed here).
def _warn_quantile_fallback_once(key: str, fallback: str, default_mode: str) -> None:
from src.utils.logging_utils import warn_once
warn_once(
logging.getLogger(__name__),
("quantile_fallback", key, fallback),
"[NormalizeTransformFn] Key %r mode=q01_q99 but stats lack q01/q99 β "
"falling back to %r. Run `python -m data_process stats` to regenerate stats "
"with quantiles. (further occurrences for this key are suppressed)",
key, fallback,
)
def expand_canonical_stats_per_key(
stats,
state_keys,
state_dims,
action_keys,
action_dims,
action_canonical_key,
schema_id="",
):
"""Slice canonical concatenated stats ('observation.state' and the chosen
action canonical key) into per-schema-key sub-dicts.
Shared by training-time ``NormalizeTransformFn.hydrate`` and deploy-time
``serve_labvla`` so both build the SAME per-key view from the canonical
norm_stats (``data_process stats`` writes only the concatenated entries).
Returns a NEW dict: canonical entries preserved, per-schema-key entries added
by slicing each canonical array along the key's dim offset.
"""
expanded = dict(stats) # preserve the canonical entries for compat
for canonical, keys, dims in [
("observation.state", list(state_keys), list(state_dims)),
(action_canonical_key, list(action_keys), list(action_dims)),
]:
base = stats.get(canonical)
if base is None or not isinstance(base, dict):
continue
total_expected = sum(dims)
offset = 0
for k, d in zip(keys, dims):
sub = {}
for sk, v in base.items():
# Slice array-like fields; copy metadata (count, β¦) as-is.
if sk == "count":
sub[sk] = v
continue
try:
_len = len(v)
except TypeError:
sub[sk] = v
continue
if _len == 0:
logging.warning(
"[expand_canonical_stats_per_key] empty stats array: "
"stats[%r][%r] is len=0; schema_id=%s, canonical=%r, "
"per-schema-key=%r, expected_dim=%d, slice_offset=%d, "
"total_expected=%d. Passing through as-is.",
canonical, sk, schema_id, canonical, k, d, offset,
total_expected,
)
sub[sk] = v
elif offset + d > _len:
# Fail loud: normalizing a d-dim key against _len-dim stats
# would be wrong numerics with NO error.
raise ValueError(
f"[expand_canonical_stats_per_key] stats[{canonical!r}]"
f"[{sk!r}] has length {_len}, but schema {schema_id!r} "
f"requires slice [{offset}:{offset + d}] for key {k!r}. "
f"Expected total = sum({canonical}_dims) = {total_expected}. "
f"Regenerate stats.json (`python -m data_process stats ...`) "
f"or check that the schema matches this dataset's layout."
)
else:
sub[sk] = v[offset:offset + d]
expanded[k] = sub
offset += d
return expanded
@DataTransformFn.register_subclass("normalize")
@dataclass
class NormalizeTransformFn(DataTransformFn):
"""
Normalize specified keys in a DataDict using precomputed statistics.
Args:
selected_keys: list of keys to normalize (e.g. ["observation.state", "actions"]).
If None, will normalize all keys that exist in norm_stats.
mode: default normalization mode β "mean_std" | "min_max" | "q01_q99".
mode_overrides: per-key substring β mode override. E.g. {"gripper": "q01_q99"}
makes any key containing "gripper" use [q01, q99] percentile normalization
(robust to outliers; recommended for near-binary signals like gripper).
If q01_q99 is requested but stats lack q01/q99, falls back to default mode.
norm_stats: dictionary containing normalization parameters.
Example:
norm_stats = {
"observation.state": {"mean": ..., "std": ..., "min": ..., "max": ...},
"action": {"mean": ..., "std": ..., "q01": ..., "q99": ...},
}
"""
selected_keys: Optional[list[str]] = None
mode: str = "mean_std" # "mean_std" | "min_max" | "q01_q99"
mode_overrides: dict[str, str] = field(default_factory=dict)
# Per-dim mode override within a single key. Format:
# dim_overrides = {key: {dim_idx: mode}}
# Example: {"action": {7: "q01_q99"}, "observation.state": {7: "q01_q99"}}
# β for the 8-dim "action" / "observation.state" key, dims 0..6 use the
# default `mode` (typically mean_std for arm joints, preserves natural
# z-score range), dim 7 uses q01_q99 (gripper width β bimodal/binary
# distribution, q01/q99 matches better than mean_std). Empty dict
# (default) preserves the whole-key fast path bit-identically.
# `hydrate_all` auto-injects these from `schema.gripper_action_dims` so
# callers normally don't set this by hand.
dim_overrides: dict[str, dict[int, str]] = field(default_factory=dict)
norm_stats: dict[str, dict[str, Any]] = field(default_factory=dict)
# Config-tunable for stats with tiny std values (near-constant channels);
# default preserves the previous hardcoded 1e-6.
eps: float = 1e-6
def _resolve_with_fallback(self, mode: str, stats: dict, key_label: str) -> Optional[str]:
"""Resolve q01_q99 β mean_std β min_max fallback chain.
``q01_q99_strict`` is intentionally separate from the ordinary
``q01_q99`` fallback path. It is used for Ο0.5 state discretization:
those bins are uniformly spaced over q01/q99-normalized [-1, 1], so
falling back to z-score normalization silently reintroduces the
boundary-bin saturation bug.
"""
# Explicit "noop" mode passes through unchanged β used by hydrate_all
# when --normalize_arm_joints=false or --normalize_gripper=false to
# disable normalization for those dims without rewriting the chain.
if mode == "noop":
return "noop"
if mode == "q01_q99_strict":
if not _stats_has_q01_q99(stats):
raise ValueError(
f"[NormalizeTransformFn] Key '{key_label}' requires "
"q01/q99 stats for state discretization, but they are "
"missing. Regenerate stats.json with q01/q99 fields."
)
return "q01_q99"
if mode == "q01_q99" and not _stats_has_q01_q99(stats):
if os.environ.get("LABVLA_ALLOW_Q0199_FALLBACK") != "1":
# For gripper dims the q01/q99 bounds ARE the cross-dataset
# open/close alignment contract (width 0.04 vs open_fraction
# 1.0); a silent mean_std fallback breaks it while training
# continues.
raise ValueError(
f"[NormalizeTransformFn] Key '{key_label}' requests "
"q01_q99 normalization but the stats lack q01/q99. "
"Regenerate the stats file with quantiles, or set "
"LABVLA_ALLOW_Q0199_FALLBACK=1 for the legacy "
"warn-and-fall-back behavior."
)
if "mean" in stats and "std" in stats:
fallback = "mean_std"
elif "min" in stats and "max" in stats:
fallback = "min_max"
else:
return None
_warn_quantile_fallback_once(key_label, fallback, self.mode)
return fallback
return mode
def _apply_mode(self, x: torch.Tensor, stats: dict, mode: str, key_label: str) -> torch.Tensor:
"""Apply one of the three normalization formulas to ``x`` using
``stats``. ``stats`` 1-D array fields must already be sliced to
match ``x.shape[-1]`` β the caller (per-dim path) handles slicing
via ``_slice_stats``; the whole-key path passes the original stats
whose 1-D arrays line up with the full key dim already.
"""
if mode == "mean_std":
mean = torch.from_numpy(np.asarray(stats["mean"])).to(x)
std = torch.from_numpy(np.asarray(stats["std"])).to(x)
# Reject non-finite stats up-front β they silently
# propagate NaN through every downstream op and surface as
# "why did the loss become NaN at step 0" much later.
if not torch.isfinite(mean).all() or not torch.isfinite(std).all():
raise ValueError(
f"[NormalizeTransformFn] non-finite mean/std for key "
f"'{key_label}'. Stats file is corrupted; re-run data_process stats."
)
return (x - mean) / (std + self.eps)
if mode == "min_max":
min_v = torch.from_numpy(np.asarray(stats["min"])).to(x)
max_v = torch.from_numpy(np.asarray(stats["max"])).to(x)
if not torch.isfinite(min_v).all() or not torch.isfinite(max_v).all():
raise ValueError(
f"[NormalizeTransformFn] non-finite min/max for key '{key_label}'."
)
return (x - min_v) / (max_v - min_v + self.eps)
if mode == "q01_q99":
q01 = torch.from_numpy(np.asarray(stats["q01"])).to(x)
q99 = torch.from_numpy(np.asarray(stats["q99"])).to(x)
if not torch.isfinite(q01).all() or not torch.isfinite(q99).all():
raise ValueError(
f"[NormalizeTransformFn] non-finite q01/q99 for key '{key_label}'."
)
# pi0/OpenVLA-style: rescale q01..q99 β [-1, 1] without clipping
return 2 * (x - q01) / (q99 - q01 + self.eps) - 1
if mode == "noop":
# Identity for ablation toggles. dim_overrides assigns this when
# --normalize_arm_joints=false or --normalize_gripper=false.
return x
raise ValueError(f"Unknown normalization mode: {mode}")
@staticmethod
def _slice_stats(stats: dict, dims: list[int]) -> dict:
"""Slice every 1-D array in ``stats`` by the given dim indices. Scalars
(e.g. ``count``) and short arrays pass through unchanged so finite-
check in ``_apply_mode`` can still error cleanly on bad input.
"""
idx = np.asarray(dims, dtype=np.int64)
out = {}
for k, v in stats.items():
try:
n = len(v)
except TypeError:
out[k] = v
continue
if n > int(idx.max()):
out[k] = np.asarray(v)[idx]
else:
out[k] = v
return out
def __call__(self, data: DataDict) -> DataDict:
keys = self.selected_keys if self.selected_keys is not None else list(self.norm_stats.keys())
for key in keys:
if key not in data:
logging.warning(
f"[NormalizeTransformFn] Key '{key}' not found in data β skipping normalization."
)
continue
if key not in self.norm_stats:
logging.warning(
f"[NormalizeTransformFn] No normalization stats found for key '{key}' β skipping."
)
continue
x = data[key]
stats = self.norm_stats[key]
default_mode = _resolve_mode(key, self.mode, self.mode_overrides)
per_dim = self.dim_overrides.get(key) or {}
if not per_dim:
# Fast path: whole-key normalization. Bit-identical to the
# legacy single-mode body when `dim_overrides` is empty β
# this preserves backward compat for every existing caller.
resolved = self._resolve_with_fallback(default_mode, stats, key)
if resolved is None:
logging.warning(
f"[NormalizeTransformFn] Key '{key}' mode=q01_q99 but stats lack "
f"q01/q99 AND mean/std AND min/max β skipping."
)
continue
data[key] = self._apply_mode(x, stats, resolved, key)
continue
# Per-dim path: group dims by effective mode (one tensor op per
# mode rather than per-dim), normalize each group with sliced
# stats, write back into a clone.
#
# 0-dim (scalar) handling: some adapters store single-feature keys
# (e.g. robointer_droid `*_gripper_position` with state_dims=(1,))
# as a 0-dim tensor β torch.tensor(0.04) rather than
# torch.tensor([0.04]). `x.shape[-1]` on a 0-dim tensor raises
# IndexError. Upgrade to 1-dim before per-dim slicing and restore the
# original rank after so callers see the same shape they handed in.
was_scalar = (x.ndim == 0)
if was_scalar:
x = x.unsqueeze(0)
D = x.shape[-1]
mode_to_dims: dict[str, list[int]] = {}
for d in range(D):
m = per_dim.get(d, default_mode)
mode_to_dims.setdefault(m, []).append(d)
out = x.clone()
for raw_mode, dims in mode_to_dims.items():
resolved = self._resolve_with_fallback(raw_mode, stats, key)
if resolved is None:
logging.warning(
f"[NormalizeTransformFn] Key '{key}' dims={dims} mode={raw_mode} "
f"but stats lack required fields β skipping these dims."
)
continue
sub_stats = self._slice_stats(stats, dims)
idx_t = torch.tensor(dims, dtype=torch.long, device=x.device)
sub_x = x.index_select(-1, idx_t)
sub_out = self._apply_mode(sub_x, sub_stats, resolved, f"{key}[dims={dims}]")
out.index_copy_(-1, idx_t, sub_out.to(out.dtype))
if was_scalar:
out = out.squeeze(0)
data[key] = out
return data
def hydrate(self, ctx: "HydrateContext") -> "NormalizeTransformFn":
# Reads ctx.stats AFTER any GripperSemanticCanonicalizeFn rewrite.
schema = ctx.schema
stats = ctx.stats
selected_keys = ctx.selected_keys
action_mode = ctx.action_mode
# Structural guard: `stats is None` AND empty-dict both count
# as "stats missing". Without this second check a dataset whose
# meta/stats.json is missing would pass an empty dict, then
# NormalizeTransformFn would log a per-key warning and skip β i.e.
# silently train on un-normalized data.
if not stats:
raise FileNotFoundError(
"NormalizeTransformFn requires non-empty stats but got "
f"{stats!r}. Run: python -m data_process stats --dataset "
"<dataset_root> --schema <schema_name>"
)
# `data_process stats` emits canonical-concatenated entries keyed as
# `observation.state` and `action` β the state/action vectors AFTER
# ComposeFieldsTransform would concatenate schema.state_keys /
# action_keys. But NormalizeTransformFn runs BEFORE Compose in the
# chain, so it sees per-schema-key tensors and looks stats up by the
# schema's own key names. Build a per-schema-key view by slicing the
# canonical stats along each key's dim offset.
# action_mode chooses which canonical action stats to slice per-key.
action_canonical_key = ctx.action_canonical_key
if action_mode == "abs" and "action_abs" not in stats:
raise KeyError(
f"hydrate_all: action_mode='abs' requires stats['action_abs'] "
f"but it is missing in stats.json. Re-run `data_process stats` "
f"(the newer version writes both 'action' and 'action_abs'). "
f"stats keys present: {list(stats.keys())}"
)
# Per-schema-key slicing lives in the shared module function
# expand_canonical_stats_per_key so deploy (serve_labvla) builds the
# IDENTICAL per-key view from canonical norm_stats.
expanded = expand_canonical_stats_per_key(
stats,
schema.state_keys, schema.state_dims,
schema.action_keys, schema.action_dims,
action_canonical_key,
schema_id=schema.schema_id,
)
# Auto-build per-dim mode overrides.
#
# VLM-pretrain discretization exception:
# Ο0.5 bins are uniform over q01/q99-normalized [-1, 1]. If dims
# stay on mean_std, roughly 32% of near-Gaussian joint values fall
# outside [-1, 1] and collapse into boundary bins. Therefore,
# whenever DiscretizeStateTransformFn is present in the input
# chain, every state AND action dim is forced to q01/q99 strictly.
# Posttrain / Ο0-style knowledge-isolation runs do not include
# DiscretizeStateTransformFn, so they keep the mixed policy below:
# arm joints mean_std, gripper q01/q99.
#
# Action/gripper policy:
# The canonical gripper dim (post-Compose, in the concatenated
# action vector) is declared by schema.gripper_action_dims (and
# arm_layout.gripper_indices_canonical which mirrors it). Map each
# canonical gripper dim back to the *local* dim within whichever
# schema-key contains it, then request q01_q99 on those local dims.
# Default mode (mean_std, set at config-time) applies to remaining
# action arm dims.
#
# Why this is the right shape for action data:
# - arm joints span a wide near-Gaussian range β mean_std
# preserves natural z-score precision; q01_q99 would
# compress the working region by ~3Γ and erode spatial
# accuracy.
# - gripper width is bimodal (mostly fully open + brief
# closure events) β q01_q99 [-1, 1] matches the bimodal
# boundaries; mean_std gives huge mean-distance tails on
# the closed mode.
#
# Concretely for LabUtopia single-arm Franka:
# schema.state_keys = ('observation.state',), state_dims = (8,)
# schema.action_keys = ('action',), action_dims = (8,)
# schema.gripper_action_dims = (7,)
# β dim_overrides = {
# "observation.state": {7: "q01_q99"},
# "action": {7: "q01_q99"},
# }
# For multi-key schemas (robointer_droid: separate
# joint/gripper keys), the gripper canonical dim falls inside
# the gripper-specific key at local dim 0, so the override
# fires on the whole 1-dim gripper key β semantically equivalent
# to a whole-key `mode_overrides={"gripper": "q99"}`.
gripper_canonical = tuple(getattr(schema, "gripper_action_dims", ()) or ())
dim_overrides: dict[str, dict[int, str]] = {}
has_state_discretize = ctx.has_state_discretize
# Per-segment normalization toggle. By default both arm joints and
# gripper are normalized. Setting one to False makes those dims pass
# through identity ("noop" mode), letting the model see raw values
# for that segment.
for canonical, keys, dims in [
("observation.state", list(schema.state_keys), list(schema.state_dims)),
(action_canonical_key, list(schema.action_keys), list(schema.action_dims)),
]:
base = stats.get(canonical)
if base is None or not isinstance(base, dict):
continue
offset = 0
for k, d in zip(keys, dims):
for local in range(d):
global_dim = offset + local
is_gripper = global_dim in gripper_canonical
if has_state_discretize:
mode = "q01_q99_strict"
dim_overrides.setdefault(k, {})[local] = mode
elif is_gripper:
mode = ctx.gripper_norm_mode
# only emit override when it differs from default ("mean_std")
if mode != "mean_std":
dim_overrides.setdefault(k, {})[local] = mode
else:
if not ctx.normalize_arm_joints:
dim_overrides.setdefault(k, {})[local] = "noop"
# else: keep default mean_std via fast-path (no override)
offset += d
# Fail LOUD at hydrate time when a normalized key
# has no stats, instead of NormalizeTransformFn silently warning +
# skipping at runtime. A missing canonical entry
# ('observation.state' / action_canonical_key) leaves the per-key
# slices absent from `expanded`, so those keys would be trained
# UN-normalized while other keys are normalized β a silent
# train-correctness break that is very hard to notice in a long run.
# Escape hatch for intentional edge configs (VQA-only / fully-noop
# ablations / partial stats): LABVLA_ALLOW_MISSING_NORM_STATS=1.
_missing_norm_keys = [
k for k in selected_keys if not isinstance(expanded.get(k), dict)
]
if _missing_norm_keys:
_nm_msg = (
f"[hydrate_all] NormalizeTransformFn has NO stats for key(s) "
f"{_missing_norm_keys} (schema {schema.schema_id!r}). The canonical "
f"stats entry ('observation.state' / {action_canonical_key!r}) was "
f"absent or malformed, so these keys would be SILENTLY left "
f"un-normalized at runtime while other keys ARE normalized β a "
f"train-correctness break. Regenerate stats.json via "
f"`python -m data_process stats ...`. Present stats keys: "
f"{list(stats.keys())}."
)
if _env_flags.get("LABVLA_ALLOW_MISSING_NORM_STATS") == "1":
logging.error(_nm_msg + " [BYPASSED via LABVLA_ALLOW_MISSING_NORM_STATS=1]")
else:
raise ValueError(
_nm_msg + " Set LABVLA_ALLOW_MISSING_NORM_STATS=1 to bypass intentionally."
)
t = replace(
self,
norm_stats=expanded,
selected_keys=selected_keys,
dim_overrides=dim_overrides,
)
logging.info(
f"Hydrated {t.__class__.__name__} with {len(selected_keys)} keys "
f"({schema.schema_id}); action_mode={action_mode}; sliced from "
f"[state='observation.state' ({'observation.state' in stats}), "
f"action={action_canonical_key!r} ({action_canonical_key in stats})]; "
f"dim_overrides={dim_overrides if dim_overrides else 'none'}"
)
return t
@DataTransformFn.register_subclass("unnormalize")
@dataclass
class UnNormalizeTransformFn(DataTransformFn):
"""
Unnormalize specified keys in a DataDict using precomputed statistics.
Args:
selected_keys: list of keys to unnormalize (e.g. ["observation.state", "actions"]).
If None, will unnormalize all keys that exist in norm_stats.
mode: default unnormalization mode β "mean_std" | "min_max" | "q01_q99".
mode_overrides: per-key substring β mode override. Mirrors
NormalizeTransformFn's semantics; pair the two transforms identically.
norm_stats: dictionary containing unnormalization parameters.
Example:
norm_stats = {
"observation.state": {"mean": ..., "std": ..., "min": ..., "max": ...},
"action": {"mean": ..., "std": ..., "q01": ..., "q99": ...},
}
"""
selected_keys: Optional[list[str]] = None
# Default q01_q99 to mirror NormalizeTransformFn (Ο0.5 / openpi convention).
mode: str = "q01_q99"
mode_overrides: dict[str, str] = field(default_factory=dict)
# Symmetric per-dim mode override, mirroring
# NormalizeTransformFn.dim_overrides. ``hydrate_all`` injects the SAME
# dim_overrides into both transforms, so a key normalized with mixed modes
# (e.g. arm joints mean_std / gripper q01_q99 / noop ablation) is
# un-normalized as the exact per-dim inverse. Empty dict (default)
# preserves the legacy whole-key fast path bit-identically.
dim_overrides: dict[str, dict[int, str]] = field(default_factory=dict)
norm_stats: dict[str, dict[str, Any]] = field(default_factory=dict)
# Mirror NormalizeTransformFn.eps. Keep these two in sync when tuning: a
# forward/inverse eps mismatch produces slow drift that is hard to isolate.
eps: float = 1e-6
# Reuse NormalizeTransformFn's mode resolution so the forward/inverse pair
# resolve every per-dim mode identically (q01_q99βmean_std fallback, noop
# passthrough, q01_q99_strict). This is the same per-dim mode resolution
# the forward transform uses, applied to the inverse formulas below.
_resolve_with_fallback = NormalizeTransformFn._resolve_with_fallback
def _apply_inverse_mode(
self, x: torch.Tensor, stats: dict, mode: str, key_label: str
) -> torch.Tensor:
"""Inverse of ``NormalizeTransformFn._apply_mode`` for one mode.
``stats`` 1-D array fields must already be sliced to match
``x.shape[-1]`` β the per-dim path slices via
``NormalizeTransformFn._slice_stats``; the whole-key path passes
original stats whose 1-D arrays line up with the full key dim.
"""
if mode == "mean_std":
mean = torch.from_numpy(np.asarray(stats["mean"])).to(x)
std = torch.from_numpy(np.asarray(stats["std"])).to(x)
return x * (std + self.eps) + mean
if mode == "min_max":
min_v = torch.from_numpy(np.asarray(stats["min"])).to(x)
max_v = torch.from_numpy(np.asarray(stats["max"])).to(x)
return x * (max_v - min_v + self.eps) + min_v
if mode == "q01_q99":
q01 = torch.from_numpy(np.asarray(stats["q01"])).to(x)
q99 = torch.from_numpy(np.asarray(stats["q99"])).to(x)
# inverse of (2*(x-q01)/(q99-q01)-1) β x = (x+1)/2 * (q99-q01) + q01
return (x + 1) / 2 * (q99 - q01 + self.eps) + q01
if mode == "noop":
return x
raise ValueError(f"Unknown unnormalization mode: {mode}")
def __call__(self, data: DataDict) -> DataDict:
keys = self.selected_keys if self.selected_keys else list(self.norm_stats.keys())
for key in keys:
if key not in data:
logging.warning(
f"[UnNormalizeTransformFn] Key '{key}' not found in data β skipping unnormalization."
)
continue
if key not in self.norm_stats:
logging.warning(
f"[UnNormalizeTransformFn] No stats found for key '{key}' β skipping unnormalization."
)
continue
x = data[key]
stats = self.norm_stats[key]
default_mode = _resolve_mode(key, self.mode, self.mode_overrides)
per_dim = self.dim_overrides.get(key) or {}
if not per_dim:
# Fast path: whole-key un-normalization. Bit-identical to the
# legacy single-mode body when `dim_overrides` is empty β
# preserves backward compat for every existing caller.
resolved = self._resolve_with_fallback(default_mode, stats, key)
if resolved is None:
logging.warning(
f"[UnNormalizeTransformFn] Key '{key}' mode=q01_q99 but stats lack "
f"q01/q99 AND mean/std AND min/max β skipping."
)
continue
data[key] = self._apply_inverse_mode(x, stats, resolved, key)
continue
# Per-dim path: group dims by effective mode (one tensor op per
# mode), un-normalize each group with sliced stats, write back into
# a clone. Mirrors NormalizeTransformFn.__call__ exactly so the
# inverse is the exact per-dim inverse of the forward transform.
was_scalar = (x.ndim == 0)
if was_scalar:
x = x.unsqueeze(0)
D = x.shape[-1]
mode_to_dims: dict[str, list[int]] = {}
for d in range(D):
m = per_dim.get(d, default_mode)
mode_to_dims.setdefault(m, []).append(d)
out = x.clone()
for raw_mode, dims in mode_to_dims.items():
resolved = self._resolve_with_fallback(raw_mode, stats, key)
if resolved is None:
logging.warning(
f"[UnNormalizeTransformFn] Key '{key}' dims={dims} mode={raw_mode} "
f"but stats lack required fields β skipping these dims."
)
continue
sub_stats = NormalizeTransformFn._slice_stats(stats, dims)
idx_t = torch.tensor(dims, dtype=torch.long, device=x.device)
sub_x = x.index_select(-1, idx_t)
sub_out = self._apply_inverse_mode(sub_x, sub_stats, resolved, f"{key}[dims={dims}]")
out.index_copy_(-1, idx_t, sub_out.to(out.dtype))
if was_scalar:
out = out.squeeze(0)
data[key] = out
return data
@DataTransformFn.register_subclass("delta_action")
@dataclass
class DeltaActionTransformFn(DataTransformFn):
# Every construction site leaves `mask=None` at ctor time and relies on
# ``hydrate_all()`` to inject ``schema.to_bool_mask()`` before the transform
# runs. The `mask is None` raise below is the authoritative guard. A new
# construction site MUST either route through hydrate_all or pass
# mask=schema.to_bool_mask() explicitly β never silently broadcast a default.
mask: Optional[list[bool]] = None
mapping: dict[str, list[str]] = field(default_factory=dict)
def __call__(self, data: DataDict) -> DataDict:
"""Convert absolute action chunks β delta-from-state chunks.
Contract:
1. ``data[state_keys[i]]`` is a single-frame state. If it has a
leading time axis (e.g. when state_keys overlaps action_keys
and delta_timestamps prepends a chunk), the FIRST row
``state[0]`` is taken as the t=0 reference. This requires the
caller to guarantee ``delta_timestamps[<state-key>][0] == 0.0``
(LabVLADatasetConfig builds delta_timestamps with
``[i / fps for i in range(chunk_size)]`` so index 0 is t=0).
2. ``data[action_keys[i]]`` is a (chunk, action_dim) tensor whose
chunk[0] aligns with the t=0 state.
3. ``self.mask`` MUST be hydrated by ``hydrate_all`` to the
schema-derived bool tensor; no silent fallback.
"""
state_keys = self.mapping[OBS_STATE]
state_list, _ = self._align_for_cat([data[k] for k in state_keys])
state = torch.cat(state_list, dim=-1)
action_keys = self.mapping[ACTION]
action_list, size = self._align_for_cat([data[k] for k in action_keys])
action = torch.cat(action_list, dim=-1)
# Defensive: if a state key carries a chunked (T, D) tensor (the
# adapter's delta_timestamps expansion can write a future window onto
# a key that doubles as an action source), recover the t=0 state from
# row 0 (delta_timestamps starts at 0.0).
# schema/validate.py rejects stateβ©action overlap outright β were a key
# shared, after this transform it would hold the DELTA chunk (row0 β‘ 0)
# and downstream Compose would feed a destroyed state to the model. The
# [0]-slice below stays as defense in depth only.
if state.ndim > 1:
state = state[0]
# Guard against true dimension bugs (3+ axes, 0-dim, etc.).
assert state.ndim == 1, (
f"DeltaActionTransformFn expects 1-D state per sample "
f"(state_dim,), got shape {tuple(state.shape)}. If you need "
f"chunked state, update this transform's broadcasting logic."
)
# No silent fallback: a `mask = [True] * state.shape[-1]` default would
# accept a missing mask and could mis-align when state_dim != action_dim.
# hydrate_all always injects schema.to_bool_mask() whose length equals
# sum(action_dims). Require it explicitly β no silent fallback.
if self.mask is None:
raise ValueError(
"DeltaActionTransformFn.mask is None. hydrate_all must inject "
"schema.to_bool_mask() before this transform runs. If constructing "
"manually (tests, ablations), pass mask=schema.to_bool_mask()."
)
if self.mask.shape[-1] != action.shape[-1]:
raise ValueError(
f"DeltaActionTransformFn.mask length {self.mask.shape[-1]} does "
f"not equal concatenated action width {action.shape[-1]}. This "
f"usually means schema.delta_mask length disagrees with "
f"sum(schema.action_dims); re-check the schema author-side."
)
# When state and action widths differ (e.g. robocoin's state has
# 32 dims but action has 32 dims β equal β vs a future robot with
# state_dim=14 but action_dim=32), subtraction of (state_dim,) from
# (..., action_dim) requires explicit padding. Pad state with zeros
# on the right so only the first `state_dim` mask bits effect a delta.
if state.shape[-1] != action.shape[-1]:
if state.shape[-1] > action.shape[-1]:
raise ValueError(
f"DeltaActionTransformFn: state width {state.shape[-1]} "
f"exceeds action width {action.shape[-1]}; cannot align."
)
pad = action.shape[-1] - state.shape[-1]
state = torch.cat([state, state.new_zeros(pad)], dim=-1)
action -= torch.where(self.mask, state, 0)[None]
sid, eid = 0, 0
for i, key in enumerate(action_keys):
eid += size[i]
data[key] = action[..., sid:eid]
sid = eid
return data
def _align_for_cat(self, tensors: list[torch.Tensor], dim=-1) -> list[torch.Tensor]:
max_ndim = max((t.ndim for t in tensors))
out, size = [], []
for t in tensors:
t = t if t.ndim == max_ndim else t.unsqueeze(dim)
out.append(t)
size.append(t.shape[-1])
return out, size
def hydrate(self, ctx: "HydrateContext") -> "DeltaActionTransformFn":
t = replace(self, mapping=ctx.feature_map, mask=ctx.bool_mask)
logging.info(
f"Hydrated {t.__class__.__name__} with mask ({ctx.schema.schema_id})"
)
return t
# ============== Unified hydrate interface using DatasetSchema ==============
@dataclass
class HydrateContext:
"""Bag of schema-derived values a transform's ``hydrate`` may read.
Built once per chain by ``hydrate_all``. MUTABLE by design:
``GripperSemanticCanonicalizeFn.hydrate`` rewrites ``stats`` (gripper-dim
affine into the open_fraction domain) and later hydrations
(``NormalizeTransformFn`` / ``SnapGripperToEndpointsFn``) must observe the
rewritten values.
"""
schema: Any
stats: dict | None
action_mode: str # "delta" | "abs"
gripper_norm_mode: str # "q01_q99" | "mean_std" | "noop"
normalize_arm_joints: bool
has_state_discretize: bool # DiscretizeStateTransformFn in the input chain
feature_map: dict # {OBS_STATE: state_keys, ACTION: action_keys}
bool_mask: Any # torch.BoolTensor from schema.to_bool_mask()
selected_keys: list
@property
def action_canonical_key(self) -> str:
return "action_abs" if self.action_mode == "abs" else "action"
def hydrate_all(
transforms: list[DataTransformFn],
schema,
stats: dict | None = None,
action_mode: str = "delta",
normalize_arm_joints: bool = True,
normalize_gripper: bool | str = True,
) -> list[DataTransformFn]:
"""Hydrate all transforms using a DatasetSchema.
action_mode:
"delta" β action normalization uses stats["action"] (delta-transformed,
what compute_stats wrote when delta_mask-True dims were
subtracted from state).
"abs" β action normalization uses stats["action_abs"] (raw absolute
action values, no delta). Requires compute_stats to have
produced the "action_abs" key; otherwise a clear error is
raised.
Usage: hydrate_all(transforms, schema=dataset_schema, stats=stats).
"""
if action_mode not in ("delta", "abs"):
raise ValueError(f"action_mode must be 'delta' or 'abs', got {action_mode!r}")
if isinstance(normalize_gripper, str):
gripper_norm_mode = normalize_gripper.strip().lower().replace("-", "_")
bool_aliases = {
"true": "q01_q99",
"1": "q01_q99",
"yes": "q01_q99",
"false": "noop",
"0": "noop",
"no": "noop",
"none": "noop",
}
gripper_norm_mode = bool_aliases.get(gripper_norm_mode, gripper_norm_mode)
else:
gripper_norm_mode = "q01_q99" if normalize_gripper else "noop"
if gripper_norm_mode not in ("q01_q99", "mean_std", "noop"):
raise ValueError(
"normalize_gripper must be bool or one of "
f"'q01_q99', 'mean_std', 'noop'; got {normalize_gripper!r}"
)
if schema is None:
raise ValueError(
"hydrate_all: `schema` is required. Pass "
"adapter.meta.schema (populated by schema.discover_schema)."
)
# Qwen3_VLProcessorTransformFn iterates image0..imageN slots.
# If schema declares more, silently dropping them is a correctness hazard;
# raise so the user partitions cameras or drops a few in the manifest.
_MAX_CAMERAS = NUM_IMAGE_SLOTS
if len(schema.image_mapping) > _MAX_CAMERAS:
raise ValueError(
f"hydrate_all: schema {schema.schema_id!r} has "
f"{len(schema.image_mapping)} cameras but the current VLM "
f"processor only consumes {_MAX_CAMERAS} (image0..image{_MAX_CAMERAS-1}). "
f"Trim the images dict in the manifest to the {_MAX_CAMERAS} most "
f"informative cameras, or extend Qwen3_VLProcessorTransformFn "
f"+ UnifyLabVLAInputsTransformFn to consume more slots."
)
# Core loop: every transform pulls what it needs from the schema.
feature_map = {
OBS_STATE: list(schema.state_keys),
ACTION: list(schema.action_keys),
}
bool_mask = schema.to_bool_mask()
selected_keys = list(schema.state_keys) + list(schema.action_keys)
stats = _canonicalize_dual_arm_stats(stats, schema)
# Single-arm equivalent: remap raw 11/12-dim UR/festo/rizon4 stats to
# the canonical 8-dim layout that CanonicalSingleArmLayoutTransformFn
# builds at training time. No-op for canonical 8-dim Franka schemas
# (those have empty source_*_keys).
stats = _canonicalize_single_arm_stats(stats, schema)
# DiscretizeStateTransformFn's presence forces q01_q99_strict on every
# state/action dim during Normalize hydration (Ο0.5 binning contract).
# Precomputed from the INPUT list, exactly like the old in-branch scan.
# Import stays lazy β state_discretize imports transforms.core at module
# top, so a top-level import here would be circular at module-init time.
from src.transforms.state_discretize import DiscretizeStateTransformFn
has_state_discretize = any(
isinstance(t, DiscretizeStateTransformFn) for t in transforms
)
ctx = HydrateContext(
schema=schema,
stats=stats,
action_mode=action_mode,
gripper_norm_mode=gripper_norm_mode,
normalize_arm_joints=bool(normalize_arm_joints),
has_state_discretize=has_state_discretize,
feature_map=feature_map,
bool_mask=bool_mask,
selected_keys=selected_keys,
)
# Order-faithful single pass via per-class hydrate() methods; transforms
# with no schema needs inherit the base no-op. GripperSemanticCanonicalizeFn
# .hydrate may REWRITE ctx.stats (gripper-dim affine) for downstream
# Normalize/Snap hydration.
return [t.hydrate(ctx) for t in transforms]
|