README.md

metadata

license: cc-by-4.0
pretty_name: DexBench GA eval conditions
task_categories:
  - robotics
tags:
  - dexbench
  - benchmark
  - robotics
  - simulation
  - dexterous-manipulation

DexBench GA · eval conditions

The frozen start states every DexBench GA benchmark cell replays, as pinned by DexBench. Each row is one episode's initial scene: object choices, poses and the robot's start, sampled once and frozen so every model is scored on the same episodes.

Contents

eval/db_conditions/ holds the row files (JSONL, one start state per line) and, as <set>_frozen.json, copies of DexBench's benchmark/frozen/ pin files, which list every row file with its SHA-256 and row count.

Task	Headline cells (120 rows each)	Sensitivity rows
DB-R1	Train, GA-O	200
DB-R2	Train, GA-O, GA-S, GA-OS	200
DB-R3	Train, GA-O, GA-S, GA-OS	200
DB-L1	Train, GA-O, GA-S, GA-OS	200
DB-L2	Train, GA-O, GA-S, GA-OS	200
DB-L3	Train, GA-S	200
DB-L4	Train, GA-S, GA-P, GA-SP	200
DB-M1	Train, GA-S	200
DB-M2	Train, GA-S	200
DB-M3	Train, GA-S	200
DB-M4	Train, GA-S	200
DB-M5	Train, GA-O, GA-S, GA-OS	200
DB-H1	Train, GA-O, GA-S, GA-OS	200
DB-H2	Train, GA-O, GA-S, GA-OS	200
DB-H3	Train, GA-OS	200

46 headline cells (5,520 rows) and 3,000 sensitivity rows, in 963 row files.

Beside the rows, eval/db_conditions/ holds the start-state sets some cells read at every reset, which the rows bind by digest:

Set	Files	Read by	Bound in each row as
db_r1_layouts/eval/	240	DB-R1's cells and sweep (r1_layouts): the 30 settled evaluation layouts of each of the 8 fruit kinds (db_r1_layouts/batch_1/, 320 more, are the collection layouts, which no cell reads)	task_object.fit.layouts and layouts_sha256, the kind's layout count and digest
db_m5_piles/	617	DB-M5's cells and sweep (parts_bin_clutter_dir): the settled parts-bin piles of its 15 parts; its own README.md describes them	parts_bin_clutter_set_sha256
db_h3_starts/	166	Not read yet: the saved in-hand starts of eval/db_h3_inputs/, byte for byte, as a declared start-state set (dexbench_h3_starts_v1). --eval-conditions installs it unused.	task_object.fit.runtime_bundle_sha256 and each row's state_sha256

A set's digest is the SHA-256 of one <path relative to the set> <file SHA-256> line per file, sorted by path (DexBench's benchmark.frozen.tree_digest). DexBench pins the digests of db_r1_layouts/eval/ and db_m5_piles/ in benchmark/frozen/inputs.json.

eval/db_h3_inputs/ holds DB-H3's runtime inputs: the saved start states and runtime bundles its rows are drawn from (172 files; its own README.md describes them). DB-H3's utensils are not included; they are packages of RLWRLD/DexBench-assets, which each bundle names with the SHA-256 of every package file.

2,511 files, 212 MB.

Use

DexBench fetches and checks them for you. From a DexBench checkout:

./scripts/download_datasets.sh --eval-conditions

That downloads eval/db_conditions/ and eval/db_h3_inputs/ at the revision of this repository that the checkout's reproducibility/datasets.lock.json pins. Before it installs anything, it checks every pinned row file's SHA-256 and row count and the start-state sets' digests against the checkout's benchmark/frozen/ pins, and it checks that DB-H3's inputs reproduce all 240 of its frozen rows. Then it places them in outputs/db_conditions/ and outputs/db_h3_inputs/, where the runner and the scorer read them. The scorer refuses a result that ran on any other rows.

Changes

2026-11-10: machine-local provenance removed, history replaced

This revision replaces the history of this repository: the earlier revisions are no longer available. A DexBench checkout whose reproducibility/datasets.lock.json pins an earlier revision cannot download the rows. Update the checkout.

• eval/db_h3_inputs/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.
• eval/db_conditions/db_m5_piles/README.md says how DexBench installs the piles, and names no other download location. The set's digest changes with it.
• eval/db_conditions/db_h3_starts/manifest.json names this repository as the source of its inputs and records the new SHA-256s of the three DB-H3 files above.
• This dataset card.

No row file, start state, runtime bundle, qualification report or pile changed. With the new preparation.json SHA-256 in DexBench's DB-H3 selection, the inputs reproduce all 240 frozen DB-H3 rows.

Path	Old SHA-256	New SHA-256
eval/db_h3_inputs/preparation.json	17bf94f66cc77ef9ac8b646dd649ab4bd76f7230c516f3025477a8183c5fdbd3	66a67163a26c9e28d47b2b45441b930b6471e714274625ae1fb85f36181fdf35
eval/db_h3_inputs/train_reserved.jsonl	3b3598fde59722bec9be4e7a7beaaa4a09a6087e1382a9c42e0d09710e593563	832accdab6950e0c1f5104467727292dd800f036cb5871c490eb144ab7393832
eval/db_h3_inputs/ga_os_reserved.jsonl	acdbdea8755ed6e41b0c8e12bdb42bf3a55b47c1e1bb28451051cc0a125cc413	97ee90f919f285047a0e6d14a434058518d0c3f4c27dd342be62ba587b65cb05
eval/db_conditions/db_m5_piles/README.md	0a3fe122a69eee0aad4d9d38841f6930c6597f6188ecf4e300ac9bafad4b258c	450b92757b1ec7139402a8ad6105d0920adc7d2a82e065247ed2cca2b8869ba9
eval/db_conditions/db_m5_piles/ (617 files, set digest)	e161628a33faa8f2c34101d855e283915f13ccaf6fd672faa756ba9316798a46	a4790dcdbda50f0f7ff388fdfacbc3654caa745c6afb623559a7db2768cf161d
eval/db_conditions/db_h3_starts/manifest.json	c8b8a891c8027ecd455724b321a59a9d9adaa6afc37be00b6181ad81370d0977	ad10c5ac4377c94c75a475e8209c272969c9dbef8a75abb0a04fcc496b5d74dd
eval/db_conditions/db_h3_starts/ (166 files, set digest)	fb63a621049ba01c7b1055842d2dd4a80aad8d866f5aaee45d3ed397a329073c	252f01c4e0d1eb17cf228c7527e4dd808d57543f624868139a00a518fd2695e4

License

Creative Commons Attribution 4.0 International (CC BY 4.0); see LICENSE.md.