feat(memory): select versioned LIBERO corpora by model by ye1457 · Pull Request #202 · RLinf/RPent

ye1457 · 2026-09-17T12:08:32Z

LIBERO selects the GPT-5.5 or Astra memory corpus independently of the running model's effort. CLI, Dashboard, prompts, tools and Flash use the same selected root.

Changes

Keep version selection, manifest/hash verification, cache management and replay lookup under robots/libero/.
Add optional RobotSpec argument-validation and memory-preparation hooks. Shared startup checks errors before services start; other robots retain their default synchronization path.
Register --memory-version with LIBERO. Move the standalone download command to python -m robots.libero.memory sync; the shared rpent-memory commands remain merge, validate and build-index.
Reuse a cache only when its exact file set and hashes match the receipt. Rebuild invalid caches online and reject them offline, without changing versions.
Align the bilingual layer names with feat(memory): standardize task layers and enable RoboCasa global memory #190 and HF #13. The frozen reproduce/memory branch retains its historical layout and is not migrated.
Rebase onto current main eecf2069. Data Restore LIBERO-10 recipes and integrate memory #12 must land before doc: add init doc structure #13; coordinate feat(memory): standardize task layers and enable RoboCasa global memory #190's shared layer naming with this reader.

Validation

628 unit tests passed; 3 optional-dependency skips. 69 focused checks, pre-commit and strict English/Chinese docs passed.
A real fixed-revision HF download passed the extra-global-file regression: offline rejection, online clean rebuild, matching manifest/receipt hashes, then valid offline reuse. No network mocks.
Both LIBERO corpora passed actual GPT-5.6-sol/xhigh → MCP reads; all returned files matched source bytes. No robot actions.
Combined feat(memory): standardize task layers and enable RoboCasa global memory #190/fix(codex): enforce turn budgets for reasoning and tool-only responses #201/feat(memory): select versioned LIBERO corpora by model #202/docs(leaderboard): synchronize scores and evaluation notes #215: 697 unit tests passed, 3 skipped; pre-commit and strict English/Chinese builds passed.

read-the-docs-community · 2026-09-17T12:10:35Z

Documentation build overview

📚 RPent-cn | 🛠️ Build #34764153 | 📁 Comparing af7908d against latest (d6876ca)

🔍 Preview build

18 files changed · + 5 added · ± 13 modified

+ Added

± Modified

read-the-docs-community · 2026-09-17T12:10:48Z

Documentation build overview

📚 RPent | 🛠️ Build #34764152 | 📁 Comparing af7908d against latest (d6876ca)

🔍 Preview build

18 files changed · + 5 added · ± 13 modified

+ Added

± Modified

ye1457 · 2026-09-22T17:51:41Z

Synced latest main (eb269c8) in aff285e. No manual conflict resolution was needed; a local integration merge with #190 also completed cleanly. Validation: 620 passed / 3 skipped on this PR; 673 passed / 3 skipped with #190. Pre-commit and strict EN/ZH documentation builds passed on both. No force-push or PR merge.

ye1457 · 2026-09-23T09:49:21Z

Synced main #164 and checked the new evaluate/explore routes in 10c6523. The selected root already reaches prompts, toolkit and Flash correctly; the added checks cover that path and Astra’s unavailable replay assets.

623 tests pass here; the #190/#201/#202 integration has 693 passes, with 3 optional-dependency skips in each. Both actual model/MCP checks read the selected LIBERO corpus successfully. Pre-commit and both docs builds also pass.

qurakchin

@Sonorous281 please review this pr.

qurakchin · 2026-09-24T03:01:18Z

conflict in memory.rst:
existing: suite/+task_only/，but the newly added is task-family/+task-specific/

Updated both guides to use task-family/ and task-specific/ consistently, retaining the full layer descriptions from #190. The frozen reproduction branch is documented separately because it keeps its historical layout. See 003552b and the follow-up documentation commits.

qurakchin · 2026-09-24T03:06:30Z

+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""Model selection and complete, revision-isolated LIBERO memory downloads."""


why to write LIBERO-specific code here?

Moved the LIBERO selection, cache and replay code under robots/libero/. RobotSpec now provides optional validation and memory-preparation hooks; the shared loader keeps the default path for other robots. CLI and Dashboard tests cover validation before service startup and the selected root. See 003552b and the follow-up documentation commits.

qurakchin · 2026-09-24T03:07:29Z

+    sync = subparsers.add_parser(
+        "sync", help="Download one model-specific LIBERO memory corpus."
+    )
+    sync.add_argument("--robot", choices=["libero"], default="libero")


Is this file only for libero now?

The merge, validate and build-index commands remain shared. The LIBERO download command is now python -m robots.libero.memory sync. Updated both guides and the companion HF README examples. See 003552b and the follow-up documentation commits.

ye1457 · 2026-09-25T17:27:54Z

Updated the English and Chinese Flash and memory guides for the complete GPT-5.5 corpus: 78 plan/anchor pairs under GPT_5.5_xhigh/flash/. The examples use that directory and identify the two unavailable tasks. Historical evaluation figures and runtime behavior are unchanged.

The source archive tags were not created. Historical download examples now use the original commit IDs, and the source revisions and hashes remain in the data manifest.

HF #12 and #13 have now been rebased and pushed (b14360f, bddc229). The final candidate passed the 120 related regressions, all 78 Flash loader checks, and RoboCasa memory selection and file-read checks for all 50 tasks at seeds 1 and 5. Pre-commit and fresh strict English/Chinese documentation builds also passed.

Code documentation commits: 112078b, e2ac7d3, af7908d. Merge the data PRs in order (#12, then #13) and coordinate with #190/#202. The four planned model-to-MCP checks against merged HF main remain pending; no simulator evaluation was rerun.

ye1457 mentioned this pull request Sep 22, 2026

feat(memory): standardize task layers and enable RoboCasa global memory #190

Open

ye1457 marked this pull request as ready for review September 22, 2026 03:21

qurakchin requested a review from Sonorous281 September 24, 2026 02:57

qurakchin reviewed Sep 24, 2026

View reviewed changes

Sonorous281 reviewed Sep 25, 2026

View reviewed changes

Comment thread robots/libero/memory.py

ye1457 added 4 commits September 25, 2026 09:52

feat(libero): select versioned memory corpora by model

c339283

refactor(libero): own memory preparation and verify exact cache contents

003552b

docs(memory): align layers and migration guidance across updates

adc6c85

docs(memory): distinguish published and local RoboCasa global files

4a0c218

ye1457 force-pushed the feat/libero-memory-versions branch from 10c6523 to 4a0c218 Compare September 25, 2026 10:09

docs: align Flash plans with the complete GPT-5.5 corpus

112078b

ye1457 added 2 commits September 25, 2026 17:30

docs: update memory guide for published Flash assets

e2ac7d3

docs: use original memory revisions for historical downloads

af7908d

Sonorous281 approved these changes Sep 26, 2026

View reviewed changes

Navigation Menu

feat(memory): select versioned LIBERO corpora by model - #202

feat(memory): select versioned LIBERO corpora by model#202
ye1457 wants to merge 7 commits into
RLinf:mainfrom
ye1457:feat/libero-memory-versions

ye1457 commented Sep 17, 2026 •

edited

Loading

Uh oh!

read-the-docs-community Bot commented Sep 17, 2026 •

edited

Loading

Uh oh!

read-the-docs-community Bot commented Sep 17, 2026 •

edited

Loading

Uh oh!

ye1457 commented Sep 22, 2026

Uh oh!

ye1457 commented Sep 23, 2026

Uh oh!

qurakchin left a comment

Uh oh!

qurakchin Sep 24, 2026

Uh oh!

ye1457 Sep 25, 2026

Uh oh!

qurakchin Sep 24, 2026

Uh oh!

ye1457 Sep 25, 2026

Uh oh!

qurakchin Sep 24, 2026

Uh oh!

ye1457 Sep 25, 2026

Uh oh!

Uh oh!

ye1457 commented Sep 25, 2026 •

edited

Loading

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

Conversation

ye1457 commented Sep 17, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Changes

Validation

Uh oh!

read-the-docs-community Bot commented Sep 17, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Documentation build overview

Uh oh!

read-the-docs-community Bot commented Sep 17, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Documentation build overview

Uh oh!

ye1457 commented Sep 22, 2026

Uh oh!

ye1457 commented Sep 23, 2026

Uh oh!

qurakchin left a comment

Choose a reason for hiding this comment

Uh oh!

qurakchin Sep 24, 2026

Choose a reason for hiding this comment

Uh oh!

ye1457 Sep 25, 2026

Choose a reason for hiding this comment

Uh oh!

qurakchin Sep 24, 2026

Choose a reason for hiding this comment

Uh oh!

ye1457 Sep 25, 2026

Choose a reason for hiding this comment

Uh oh!

qurakchin Sep 24, 2026

Choose a reason for hiding this comment

Uh oh!

ye1457 Sep 25, 2026

Choose a reason for hiding this comment

Uh oh!

Uh oh!

ye1457 commented Sep 25, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

ye1457 commented Sep 17, 2026 •

edited

Loading

read-the-docs-community Bot commented Sep 17, 2026 •

edited

Loading

read-the-docs-community Bot commented Sep 17, 2026 •

edited

Loading

ye1457 commented Sep 25, 2026 •

edited

Loading