Skip to content

feat(flash): introduce evaluation-only Flash Mode with selected memory - #195

Merged
wilburx813 merged 2 commits into
RLinf:mainfrom
Elessar123:refactor/flash-mode
Sep 17, 2026
Merged

wilburx813 merged 2 commits into
RLinf:mainfrom
Elessar123:refactor/flash-mode

Conversation

@Elessar123

@Elessar123 Elessar123 commented Sep 16, 2026

Copy link
Copy Markdown
Contributor

Description

Expose recorded-plan execution as Flash Mode with --planner flash, and restrict it to evaluation. Previously, replay could be selected with --explore and read the default corpus even when --memory-dir selected a different one. Flash Mode now rejects exploration before services start and reads plans from the current toolkit's memory root.

  • Rename the planner, LIBERO modules, RobotSpec hook, CLI help, Dashboard display, and paired documentation with minimal terminology/path updates and necessary evaluation/memory usage notes. Remove the old CLI/module names without compatibility aliases.
  • Synchronize only <robot>/flash/** for Hugging Face-backed Flash evaluation in both CLI and Dashboard. Local evaluation performs no memory download and requires matching plan/anchors files in the selected corpus.
  • Use memory/libero/flash/ and the existing flat plan/anchors JSON schema. Preserve replay actions and grounding behavior. The standalone generator retains its audit/recipe inputs; no new memory-generation workflow is introduced.

Migration requires the new --planner flash command, robots.libero.flash module paths, and libero/flash/ data directory. Companion resource PRs target their respective main branches:

  • Hugging Face dataset PR #9: all 78 plans (156 JSON files), replacing the previous directory.
  • misc PR #63: renamed/redrawn figures with unchanged comparison values, plus CSV inputs and a reproduction script.

Merge the resource PRs before this change so the default download and documentation image URLs resolve.

Testing

  • pytest tests/unit_tests -q: 581 passed, 3 skipped on Python 3.11 with CPU PyTorch.
  • pre-commit run --all-files: passed.
  • sphinx-build -W --keep-going docs/source-en /tmp/rpent-flash-docs/en: passed.
  • sphinx-build -W --keep-going docs/source-zh /tmp/rpent-flash-docs/zh: passed.
  • Loaded all 78 migrated plans and verified unchanged runtime contents; verified all 156 uploaded JSON files against local Git blob hashes.
  • The previous replay test module is removed; additional CLI regression tests are excluded from this PR. The remaining unit suite passes.

Manual verification

CLI and generator help commands were checked without starting robot services. No simulator, GPU, model-service, or real-robot evaluation was rerun; existing primitive execution and grounding algorithms are unchanged. This change does not report new benchmark results.

Checklist

  • My code follows the code style of this project.
  • I have updated related documentation when needed.
  • All new and existing tests passed (3 dependency-gated skips).

Signed-off-by: Elessar123 <571360229@qq.com>
@read-the-docs-community

read-the-docs-community Bot commented Sep 16, 2026

Copy link
Copy Markdown

Documentation build overview

📚 RPent | 🛠️ Build #34594721 | 📁 Comparing e624db3 against latest (902ac6b)

  🔍 Preview build  

5 files changed · + 1 added · ± 3 modified · - 1 deleted

+ Added

± Modified

- Deleted

@read-the-docs-community

read-the-docs-community Bot commented Sep 16, 2026

Copy link
Copy Markdown

Documentation build overview

📚 RPent-cn | 🛠️ Build #34594722 | 📁 Comparing e624db3 against latest (902ac6b)

  🔍 Preview build  

5 files changed · + 1 added · ± 3 modified · - 1 deleted

+ Added

± Modified

- Deleted

@wilburx813
wilburx813 merged commit de921c5 into RLinf:main Sep 17, 2026
8 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

3 participants