mj-rlwrld's picture Joocjun's picture
Publish the DexBench GA eval conditions
efc9aa1
|
Raw History Blame Contribute Delete
17.8 kB

H3 evaluation inputs

Portable inputs for 120 Train-object and 120 GA-OS policy evaluation episodes. Each cell has four trios, three target types and ten starts per trio/target. Training demonstrations are separate; this directory does not publish them.

GA-OS couples held-out utensil combinations with unseen, geometry-dependent grasps from the same generation procedure as training. It is not an independent spatial extrapolation test. Layout ranges overlap; see layout_audit.json.

Runtime JSON bundles (dexbench_h3_runtime_v2, package dexbench_h3_inputs_v2) pin the exact original saved states and commanded joint targets. Their utensils are not in this dataset: they are asset packages of DexBench (assets/object_library/cutlery_set_*, butter_knife_d00052, fork_d00313, kitchenware_*), and each bundle names the package's main USD relative to assets/ and the SHA-256 of every package file, so a different utensil is refused. The 240 selected starts passed native 10-second retention checks with the collection actuator profile, same-side robot collision exclusions and isotropic-max utensil inertia. These are retention qualifications, not policy extraction-success scores.

qualification_reports.json preserves the native reports. preparation.json and *_reserved.jsonl preserve the immutable geometry and cross-platform admission provenance. Their historical notes about a pending independent angle split are superseded by the coupled GA-OS definition above; the state and physics evidence is unchanged. Physical states may be reused across different targets.

Download with DexBench scripts/download_datasets.sh --eval-conditions at this dataset commit. The checked-in H3 selection and frozen manifest validate all inputs and reproduce all 240 row hashes before installation. Outputs are installed under outputs/db_h3_inputs and outputs/db_conditions.

Quaternions are XYZW. No operator target cues are included in policy cameras.

2026-11-09: utensils moved to the asset pool

Made from the earlier package with a migration tool (package, then rows on the 24 Train/GA-OS and 20 sensitivity row files). asset_payload/ (138 files) and asset_payload_manifest.json are removed; their utensils are byte-identical, apart from the wrapper layers' reference path and layer metadata, in DexBench assets/object_library/. Saved states (141 files), qualification reports, preparation evidence, reserved rows and layout audit are unchanged. In every row file only task_object.fit changes: runtime_bundle_sha256 and the three utensil digests take the v2 values and asset_payload_sha256 is removed; starts, states, reset contracts, draws and recorded outcomes are unchanged. eval/db_conditions/db_h3_sensitivity_frozen.json is the re-pinned sweep set (DexBench benchmark/frozen/db_h3_sensitivity.json).

File Old SHA-256 New SHA-256
eval/db_conditions/db_h3_sensitivity_frozen.json d00e3488edab0cb365c3d866b1d3f9e384b5d228e6d58b3c28608dfcef9caff0 cef6aaf34019e4d46d347cbb22221d51961457ae066079ef3c9ac094db6aae59
eval/db_conditions/db_h3_ga_os/ga_os_01_fork.jsonl 1258f04b989616ed548172ef49ef5ce553f1f2bad54a2df788d54ce5ea35844b 73a687a1074f64789f7a77139034231827cf23fa00bc591eeaefe3f6eda60174
eval/db_conditions/db_h3_ga_os/ga_os_01_knife.jsonl e9aa31f63ca189b33c4e15cfafda01a52d0a3911bb1672d4c522232855a17980 dee9e500b47d421f6ba7a3a28d6ac7a5ac766dc9aeb8664fbba5545bdb7def74
eval/db_conditions/db_h3_ga_os/ga_os_01_spoon.jsonl 281490b7c0704c3878861e01c5b4e6c3a0f326888f6c2b89aa499de7bd60e45d c9873bb928687a03471b0613196e78fd6e765364c59679254382f959681aeb0f
eval/db_conditions/db_h3_ga_os/ga_os_02_fork.jsonl f704c51d98f564b43b6cdcee27d39e696dde9d1585abc6235cce33064a453327 bef0f5d019dcd4e82deb237c938a1b9c3cd515d388a7c84cb2f8d50cf990cd2e
eval/db_conditions/db_h3_ga_os/ga_os_02_knife.jsonl 46ded5b619a7402a5a31e4d74c7d94acce0c2c75926effa4dda0ee3e3da3bc15 89ed1e37728b83d3c689683a535e435f8752837c3650837fce5c6e41ec4803fd
eval/db_conditions/db_h3_ga_os/ga_os_02_spoon.jsonl 90f272e32896edd51b4b59633e30eb6e5eeb316330019475e36b8d59d998802a 43c2ac284ba0a25b05003a5e4480c4059d5db42b89be5d8148fcca67e538dc9b
eval/db_conditions/db_h3_ga_os/ga_os_03_fork.jsonl c4fe5a13404b72282d6d9c773b97c767bafb9c4bc19c6a1406f3157eca2c0f96 e0d9b0535b66c30c4eb29ca1d831917cd7c884c05b917b3715a5c5b8b3623da6
eval/db_conditions/db_h3_ga_os/ga_os_03_knife.jsonl f3a6659ed69f5d12371c557604601c57faba792bcc0f8497e57d3479985f708b aac83c292d44dc825d202aede7b931b106b2f499c0685adcf039d4d62bcd7ea9
eval/db_conditions/db_h3_ga_os/ga_os_03_spoon.jsonl 60f69658ac8cb78b5b2910d190d7dfac3f78a34c8a297fbf5472da44262c1fc2 d47626f1c2f62fd96ffc120d1aafa74da7706fb4e7a28f449f36b4afaeceadc7
eval/db_conditions/db_h3_ga_os/ga_os_04_fork.jsonl 194372d30fbdd6f6adee00c3f173b2d4f44d5cae0a981e9671e122b9ee5d665b b12c72454f75e37f35b7255b6120f8fe64a6469e57fa3a1ee32d2fd9f91b7bc7
eval/db_conditions/db_h3_ga_os/ga_os_04_knife.jsonl e85819937b53ccb0075277add8bcbe912bc3a4637a4c094f1d3e83f401fedaf3 2f2023a2190de03252c7113d47d1ee8059cc862d27d2fba13f61463331a1d2d4
eval/db_conditions/db_h3_ga_os/ga_os_04_spoon.jsonl 90661b39e655c4e617614b5f39b19d0cb686a035731cf29c8c0be6e312b45dc5 88d3218487854f54df8d414a35d178ad296b20ce4fc1b04c80d60ff96b21015a
eval/db_conditions/db_h3_sensitivity/train_01_spoon_rebuild0.jsonl b28ba8f26b5f38f8d56bb87c32d8b2737ff6715ed1827969af2b969e6418a28c a29588f1c1449699a5f3baa1082083368f96208d6082a2065eaafaf40abee95b
eval/db_conditions/db_h3_sensitivity/train_01_spoon_rebuild1.jsonl 9f681e729284d5c0375996500e4d06f1fce9c9f2e5269491d10b679556c8f111 121bce56ccd8e1de63cbb669a79042250e12a78650f0abd10c2aa4d403cfa1db
eval/db_conditions/db_h3_sensitivity/train_01_spoon_rebuild10.jsonl 3023741283d49f502951d6945e6743e1ca26a8b1ca9226a9815d8a023f482f27 3becd985f3b5e9ceaff8f4dc9068e18436ba7cebe9964927bf21871b341d9daf
eval/db_conditions/db_h3_sensitivity/train_01_spoon_rebuild11.jsonl 8ffecce6d7023791e248448b208156ad5827eff277e911ddb2b98d9f80ec6cb4 49fe01a0610bf38a4a83df3344f291e9b0bc8426766c0a0e9ee007248e377c9d
eval/db_conditions/db_h3_sensitivity/train_01_spoon_rebuild12.jsonl af51a0e783e35f4cd931462e525359019cd6cc92c4556301cae756a1301300fe 808c96247ad82d00de51bf4589dfd1330ccf1a51753cdf5d165a5541292ef9ab
eval/db_conditions/db_h3_sensitivity/train_01_spoon_rebuild13.jsonl 22270808a8f6c621ff3bde08ffc7a860d0be83b4ca294078c9cdd7757156ef34 48833cc22790f62e6633346d9fcd8762c983ed5b5ac7046e1073302eae474762
eval/db_conditions/db_h3_sensitivity/train_01_spoon_rebuild14.jsonl 154371365d8f660e453aee0b65ffeb5422be52d73605b1b0081dce597e6a297d 65021dd8c6c995d87323e0750d875dc48edeca4b99598cd2a1e628806a716924
eval/db_conditions/db_h3_sensitivity/train_01_spoon_rebuild15.jsonl db014eda7138e733039ddc6fafada5727510760ac7146153b40f54c3137d8745 28d9741a8deec2ecb702591dbc705b9cb8bf8e979ad843235bb1ce8a14491cdf
eval/db_conditions/db_h3_sensitivity/train_01_spoon_rebuild16.jsonl 70522a7a5a78d6458d2d489369fb29ecea7ff31128ef909737d375314e6ce456 30ad8ca8b852efdb3e6a7104fda84443ffdc7e8934e7ebe2bb9b1d00393d1ef7
eval/db_conditions/db_h3_sensitivity/train_01_spoon_rebuild17.jsonl c47d696bd35300218b7072f0d92269f0de572f64913e4f7b58e0fd75cfed07ca 7795e9c0add5d8665344d58dc8686a4cd617244f2ecc521cb00fd3fc4d92a2bd
eval/db_conditions/db_h3_sensitivity/train_01_spoon_rebuild18.jsonl 386963e998bf8fbac2917b4f363e2d077bd040b318f5fa291fbe9a539a3316f0 47ee2870743455cea712d89352d0be2fb32ff0abad1a184a046270ad02317285
eval/db_conditions/db_h3_sensitivity/train_01_spoon_rebuild19.jsonl ae13d9a741a70478b1a6bd2b342844d7d1bee295aafff8136c27c5da61e94b4d 1ef14041514c0dd4c49eab7ad7cf40bde96549645f4fbe975ea7069753a5d623
eval/db_conditions/db_h3_sensitivity/train_01_spoon_rebuild2.jsonl f8852b9637ffb5e6bb39e1a30bf16db3bbe6c6db9f280f643d614d6b78f881b8 38807a0a7d860378a33e9f9675e1e1376cf93c708efd7376df0ffc2969d3f3c7
eval/db_conditions/db_h3_sensitivity/train_01_spoon_rebuild3.jsonl ba055b6cf9734b79f159d4b9271dc85960a0a5d1e7bb7a486a5af6e9924b23a2 9d1053dbc4661e73679f097d5c00a7a0e1c647bc5f2b9c8e3a666ded5915afa6
eval/db_conditions/db_h3_sensitivity/train_01_spoon_rebuild4.jsonl 3e8a478cbe5c68f6efcc7edb669af991ee77ff1257650ed81277428c38c66868 ee7c489adc31ca7122cfab301db8acca129c8780385d2b5961e420ef7b1a44f1
eval/db_conditions/db_h3_sensitivity/train_01_spoon_rebuild5.jsonl b1a3debbbf7ef1cf35da5760f55deed24d9053b20a9c02acf013ef30303fcf5b d942b2f5912133ad7f05918f153b730bbd1a9a9198c4a9004b68c90c38a940b8
eval/db_conditions/db_h3_sensitivity/train_01_spoon_rebuild6.jsonl b597544f4fd5229ad5e7bd4e7e42276b39c321d3dd500db4c8135488c76d52e7 73c6ba2ff30388361bb556788f2de4efb5910b6f9c055485a1ba4d10b924380c
eval/db_conditions/db_h3_sensitivity/train_01_spoon_rebuild7.jsonl 09842ad6780f39accbf79ef5f29d682540019fd6f5d180b23127625c8c571c45 942fa30e294a65df2ffd8b70bd4ae152c5ede0d3ee41ca9d60800034cf897f41
eval/db_conditions/db_h3_sensitivity/train_01_spoon_rebuild8.jsonl c296c080e28c7203ddbd7dadb2112e5ff9f5f9fd50a2150ed1c81fadd356edaf 5937e5de71f235a752fa1af0a39f3e8eae07185e3396d0f64ee4f43053badee4
eval/db_conditions/db_h3_sensitivity/train_01_spoon_rebuild9.jsonl 2506f6d6e8403a032d08da01f1fe160e5bb22b52367043896f4f523ea2e22b2d 16b26e7cc53cab6733ba6b03f9a525ba528fd15b3ee9007b3562938dad171a76
eval/db_conditions/db_h3_train/train_01_fork.jsonl c275e24d5b79298df67c61164daf3ac572fa28bdd82d66237b2f3885f1f488c1 18f7070f875a33799a2e12f0da78e7fcb133fddb7d2deca0989c0dbd86589552
eval/db_conditions/db_h3_train/train_01_knife.jsonl 3df59fb4e53b5eaf50e5d83caec0a13b1ace4261123a78ae32c123ce995393c4 33f0df0a829049680740aad058c3a333a6c0e500918138a07d7fd2aaa8ec6773
eval/db_conditions/db_h3_train/train_01_spoon.jsonl b2e82bf74fcb73a240d0e037635191b3c748a5ba0729814eef91cef4c472819f 44635cb2b5fdcfdc8773ce7ca3d9688605cf9030e933a6a18154657eda4b5988
eval/db_conditions/db_h3_train/train_02_fork.jsonl da309cae763559a20ca1c5a6a8647bf81fa31960459935c99c17e6ef85587ad7 3a38140525c8dc76d76b75d535d194afbc20fdfe1d412eb9fecdb0f7cd50014f
eval/db_conditions/db_h3_train/train_02_knife.jsonl 18065ab6721f25ed3032a065af63ca023139d3810d469bdb6a00c492ba9961cb 56d3a88e5f493312f81b9da1387644a3baeb9f48255e0b7e6d9d37b6015a69a5
eval/db_conditions/db_h3_train/train_02_spoon.jsonl 493a4e748cd34a6e936dc5effcfff2513161bc8453c85421b214f830cb7781e0 c59b9c773b7c043448b3ac6aebf83b39f6b20f0112e521c338250f1cc6926983
eval/db_conditions/db_h3_train/train_03_fork.jsonl 36afff27cafc463ab92fff9ccf23960bc15d594f0e1dd0e33ca7ae0ef2679643 71f906cfdd0fd63e79064b8439f03977248d9e1456e604704d963f410e9f679d
eval/db_conditions/db_h3_train/train_03_knife.jsonl 4d6465144a0272283bf253baaf6d605efffbb3cf7d579009394b5b962b7e8000 98f6010d1e44cf7cdfbc61bb6cd8ef91ba9f402e60b951df5088c0353994a306
eval/db_conditions/db_h3_train/train_03_spoon.jsonl bd15ffdb07581bad317763033d41fcbffbe4b4f135b2c3acca5ed3e523eebe71 595eb2ae2f399dd9baab3bca37451aeb441862e8adf855bbb4cba4a9b998454f
eval/db_conditions/db_h3_train/train_04_fork.jsonl 155ece9b70125ad4744e0489f6cd446f8ae931a561c10ca97a51665dc8287ea2 1372a58044d2cd5aa55f340d8da0baa829db3270b54ea36a42d3de87f609faaa
eval/db_conditions/db_h3_train/train_04_knife.jsonl f1d89b559ae87eb222c4dbc8b4ced44c1210856347cfb503f4ddded88ee9e86c d0002c9ed9fa50a8f8e0195044578835ff11a52d914300ff1409621a6ea361c6
eval/db_conditions/db_h3_train/train_04_spoon.jsonl b3e38af4376d969de35db7e6c2a03db9bf254cd859bf74f1404ebead225a3dd4 2e89db7aac2a71f6942a980456a89d6b00480dcc31008354c02b7332f8701a06
eval/db_h3_inputs/manifest.json 0ddaf90b1e4b8cc02ef5af24bb37c331fafc3b1abb3fa2a3619d4089c95d783f fb26522605941a85c8b01552906f699c4f853856fa1abdb71b01e67f7aa607eb
eval/db_h3_inputs/runtime/ga_os_01_fork.json 01c22713e6589bab415abb551efb938e18bee1db8d9fc6b4b1d80ba9facace55 2fde82923c733ead1fe8ac42e098cb22f15e631ef3cd7cf2a173f77d2a8509f1
eval/db_h3_inputs/runtime/ga_os_01_knife.json e5e26f604da07e9a5ce6dc73915c4c63ed72efd8e7ae070033fb97506f471e23 fcecd425166ca2440b8615b6d20dc4f7669abf39b0d0004305e3bf3a24890994
eval/db_h3_inputs/runtime/ga_os_01_spoon.json 1081b26aef7ec7162afc4aa2228fc6b08bf70de396349d9b35b9db4ff6640424 efdfb3f64108b86a1f833f505bde635dc1b457cb61f8e1646f1a5ee8dcc4fe3b
eval/db_h3_inputs/runtime/ga_os_02_fork.json 42f5ead8ba71f1180895506deb68c31004a0e71836a411694bf7c3a92b5ee9fb 24b2f4a93e8acfb2baebdd45ceb967fb3ab8aa24b416266038f9fa7848bbdf57
eval/db_h3_inputs/runtime/ga_os_02_knife.json 02444386154c56eae27534c8766b642eb604d4eede82662f1446eac5acfc1fcd 91c639e7bf5cc376119d401960cddd12f5ca7bd7b5cf57a557daa357b5059ae0
eval/db_h3_inputs/runtime/ga_os_02_spoon.json f2fe924c7474323d252a633e3183b821546c788bab20e0927f23d4fb65ebc3f7 b060b690b1df234e91c187cf0330708eb790595e8b48b4fbd92c808fa19400f3
eval/db_h3_inputs/runtime/ga_os_03_fork.json 7e5f8082308f87961904603d6c2de153cf781f5f5a243d999ea73bc54958f3e3 2fb8aa0aff734298c260af3b8e409e4fe58454470d27d55dadbc3b933d68ec31
eval/db_h3_inputs/runtime/ga_os_03_knife.json e1dc63e8fb4323c7907dd6e8b780750281ca11fd13565ff71f2d04b3be489ad4 f25d1ab2ef347b0e18ec0735167781a7d3305c509f5285aa359dab910bfdd8dd
eval/db_h3_inputs/runtime/ga_os_03_spoon.json 157c01de37b3cee9c936bcf71db53b2618910f2b85019d92631ad2ad9d729142 519bc1a8e65dd3e4b4c42b489d39b4104ecb53cd2eda51d5f2bc95040be541c9
eval/db_h3_inputs/runtime/ga_os_04_fork.json 720c4fb914216ad9e3def99dabd3424b42de166bc4245b5861951a7dfbd7c92e c51310138a11ef9760354c997b7974bdd477958aff8cece3f3520d888b752eaa
eval/db_h3_inputs/runtime/ga_os_04_knife.json c8d4f9d03aff68bad3e0f3a1020b6b7f51569b293f849b4b7a7a0b7519359174 43d0fa68b9da59706ec419d44a343c2bd8b40655b76e6fc6261ddd369cf3e865
eval/db_h3_inputs/runtime/ga_os_04_spoon.json bb4018cb9fc0e73e6fd2b6c5e47a13c187289974c936e6f769c325f2f828aa76 7753f1258c01abab04d01f2f0b4d61ae66b1e1e065a382d27434d4b77dd1dce2
eval/db_h3_inputs/runtime/train_01_fork.json d433ed147350d9e2e39ab32c8510927c504739891d6aba62483567001e1301aa 455f37ac8bb22f9790d896f03de2a8f63e7bd9784b7e2f2f0d0b99e4451eb644
eval/db_h3_inputs/runtime/train_01_knife.json a66fdc99ca60694bf4511fdbe2501be6850b95f9d31b92e63ea3f908d8d3b3e3 47bccd026059b6604d23f36f17f2f5af6419ab420ece735dd6197f7ee2504fa1
eval/db_h3_inputs/runtime/train_01_spoon.json 11659d3285d76ce0144582b12b074313cf1d7adcc94c372eb9fc1be26b3c2f51 197a33de4921eb4188cbf3acb9cc6b5ff9ab70d20754bd3969241717a8ef6b4a
eval/db_h3_inputs/runtime/train_02_fork.json f709648113ca956e9ddad5436da4b9d0c86cdc664c5233bd007f6042f1ad2e84 4104a02ccff032d5894bfb4163a86430c53621d0abf2367c2831a1d45bca3e70
eval/db_h3_inputs/runtime/train_02_knife.json f5d3068fed5261f4497db677ae92ed29cf05735d964f84b0fad8f981485645b3 aa6e5550528cce02ef6f4a2132a8c5cc737d5a8a9ef39149b5d2fdd793048a2b
eval/db_h3_inputs/runtime/train_02_spoon.json d66d48ddecebbbcd5e90955fce3ddfd17dcea0c5b826d377e86dd8a7c7eceb5c b9ed868cc7edcb1548cbf5211b2137bbd471f3aa38b0c6dec751275ab5ccd0fa
eval/db_h3_inputs/runtime/train_03_fork.json 0e02376cdced2dfa21dacd641d1c5c36ea73efe982684b0e2dbfb763132ddc60 7a2d1d0e2865b94dde1dd0f99f97b6700e68a24c90dba3c3175d987224f77372
eval/db_h3_inputs/runtime/train_03_knife.json e03b30a81aa9d83ce8a59ad64a7b828fbfb686775949d0c553ac264324e78dc4 d7061b0e06836e4dec6e6d25cadc55f9ed8a241e9c40d54520cb1e136b5f17c3
eval/db_h3_inputs/runtime/train_03_spoon.json 47fa7a42327e031bb21a67c88b33e54b22813a3f7d3e0f9662f5707234a57bc1 0cc84a1ad42813abca19cb90b53e13c4aa5aad6358ec1459be31157b6bfa43e9
eval/db_h3_inputs/runtime/train_04_fork.json 8f18f595966e34fa907500a8437797cedc4581a50ff9d43a09b9992d8be6631c ff4c148c539280060325c6f03a5f00e31891d2942900502ce52e902bf497e896
eval/db_h3_inputs/runtime/train_04_knife.json 45882344ef950efb1bf2da53d5b9a10ee8f2ce1d0ead8cf324fa03556d5df6bd dff55acb418a20e00d71ed51ed154e43a17eb71abf686c2449a05f3e382efc6e
eval/db_h3_inputs/runtime/train_04_spoon.json 652fc43c57bbae8f9ad5b7478c3c5899b09e69d342139b212c5fed3b92996235 91d4ca2f6fc435c7860020f5387503c8d4c411af54bda4a6dd927caf22ee17ac

2026-11-10: machine-local provenance paths left out

preparation.json, train_reserved.jsonl and ga_os_reserved.jsonl recorded each preparation input by an absolute path on the machine that prepared it. These 2,654 paths are now null; the SHA-256 beside each path is unchanged. preparation.json records the new SHA-256s of the two reserved files. States, bundles, qualification reports, the layout audit, manifest.json and every row are unchanged, and the inputs reproduce all 240 frozen rows with the new preparation.json SHA-256.

File Old SHA-256 New SHA-256
preparation.json 17bf94f66cc77ef9ac8b646dd649ab4bd76f7230c516f3025477a8183c5fdbd3 66a67163a26c9e28d47b2b45441b930b6471e714274625ae1fb85f36181fdf35
train_reserved.jsonl 3b3598fde59722bec9be4e7a7beaaa4a09a6087e1382a9c42e0d09710e593563 832accdab6950e0c1f5104467727292dd800f036cb5871c490eb144ab7393832
ga_os_reserved.jsonl acdbdea8755ed6e41b0c8e12bdb42bf3a55b47c1e1bb28451051cc0a125cc413 97ee90f919f285047a0e6d14a434058518d0c3f4c27dd342be62ba587b65cb05