any4lerobot
Guides Any4LeRobot dataset conversions, LeRobot-to-RLDS export, and LeRobot v1.6-v3.0 format migrations with schema validation, version-aware environments, and safe publication boundaries.
- License 许可
- MIT
- 最近更新
- 2026-09-03
- Stars
- ★ 351
- Forks
- 27
01 前言 Front Matter
- name
- any4lerobot
- license
- MIT
- metadata
- { "disco-role": "operating" }
- description
- Guides Any4LeRobot dataset conversions, LeRobot-to-RLDS export, and LeRobot v1.6-v3.0 format migrations with schema validation, version-aware environments, and safe publication boundaries.
- disable-model-invocation
- true
02 资源文件 5
- 参考文档references/repo-provenance.md2.6 KB
- 参考文档references/repo-routing-metadata.json314 B
- 参考文档references/troubleshooting.md4.7 KB
- 参考文档references/workflow-selection.md3.6 KB
- 脚本scripts/check_environment.py4.7 KB
03 SKILL.md 正文
Any4LeRobot
Any4LeRobot is a script-oriented toolbox for moving robotics datasets into and between LeRobot formats. Use this skill to plan a conversion from a documented source layout, select the matching route, validate schemas before writing, and keep Ray, Beam, simulator, video, and Hub side effects explicit.
Start here
- Identify the source format, target format/version, local source root, output root, desired modalities, and whether publication is requested.
- Read workflow selection and choose one focused route below. Load troubleshooting before changing dependencies or retrying a failed run.
- Check repository provenance when working from a checkout or deciding whether the guidance needs refreshing.
- Run the safe environment probe in
scripts/check_environment.pybefore executing any converter. It reports core and optional import/API status but never downloads data or writes a dataset. - Build a no-write preflight: verify input layout, feature shapes/dtypes, version/API compatibility, distinct output path, backup/rollback plan, and resource budget. Start with local/debug execution; defer Hub publication.
This skill contains distilled guidance, not a dependency on the original
checkout. Do not tell a future agent to run a source-repository script or
checkout-relative convert.sh; use the route references and their commands as
recipes, then verify the equivalent entry point in the selected environment.
Route map
| User request or input | Read this route |
|---|---|
| Design an adapter, use local/Ray DataTrove execution, aggregate task outputs, resume, or publish | generic-conversion |
| Open X-Embodiment/RLDS/TFDS → LeRobot | openx-conversion |
| AgiBotWorld raw tree → LeRobot | agibot-conversion |
| RoboMIND benchmark/embodiment HDF5 → LeRobot | robomind-conversion |
| LIBERO HDF5 → LeRobot or LIBERO rerender planning | libero-conversion |
| RoboCasa HDF5, subset, depth/segmentation, or rerender planning | robocasa-conversion |
| LeRobot → RLDS/TFDS | rlds-export |
| LeRobot v1.6, v2.0, v2.1, or v3.0 migration | version-migration |
Do not conflate the two RLDS directions: openx-conversion consumes an RLDS
builder, while rlds-export produces one. Do not use version-migration to
convert raw HDF5 or to repair missing modalities.
Environment contract
Any4LeRobot has no pyproject.toml, setup.py, or declared distribution. The
usual baseline is a private Python 3.10/3.11 environment with a LeRobot writer
compatible with the selected source script, NumPy, h5py, pyarrow/pandas,
Pillow/OpenCV, tqdm, and the video codec stack. Add only what the route needs:
datatrovefor the shared pipeline; Ray/DataTrove Ray extras only for distributed execution.tensorflowandtensorflow-datasetsfor either RLDS direction; Apache Beam is optional and should be disabled for small/lossless exports.datasets,jsonlines,safetensors, and the exact historical LeRobot revision for version migration. Keep v2.x and v3.x environments separate.- RoboCasa/robosuite or LIBERO plus their assets only for simulator rerendering; these are not required for ordinary HDF5 conversion and are not verified by this skill.
Verify the installed API rather than trusting a package import alone:
python -c "from lerobot.datasets.lerobot_dataset import LeRobotDataset, LeRobotDatasetMetadata; print('LeRobot API OK')"
python -c "import h5py, pyarrow, torch, torchcodec; print('core dependencies OK')"
The source snapshot mixes historical and current LeRobot import locations. In
particular, a current release may not re-export LeRobotDataset from
lerobot.datasets or provide lerobot.datasets.dataset_writer.py. Treat that
as a compatibility stop gate; do not patch imports or metadata behavior by
guesswork. Read the route-specific troubleshooting before selecting another
revision.
Safety gates
- Never use the raw source tree as the output root. Inventory existing output and temporary paths first; several converters delete an existing destination.
- Never enable
--push-to-hub, Hub deletion/tagging, Git-LFS moves, or public upload during the first local pass. Use a new staging root or test branch. - Do not run Ray, Beam, simulator rendering, or large video/data conversions as an import or help check. Record external datasets, credentials, assets, and hardware as explicit prerequisites.
- Preserve source version, feature schemas, task text, and episode counts. A successful file open is not proof that all episodes or modalities are convertible.
Shared references
- Read workflow selection for route triggers, directionality, and the preflight record.
- Read troubleshooting for install/import, optional backend, schema, output, and publication failures.
- Run the environment probe for a no-write dependency/API status summary.
- Read provenance before refresh or when the repository commit differs from this generated baseline.
04 同仓库更多 Skills
passnet-triton-opt
Make a single PassNet Triton kernel fast AND numerically faithful: performance model (when a replacement can win at all), block/grid/warp tuning, autotune policy, launch overhead, and per-op numeric recipes to pass the dtype baseline tolerances. Use when a pass matches and is correct but speedup below expected, or when correctness fails by small numeric margins.
passnet-solve
END-TO-END playbook for solving one PassNet sample: analyze the computation graphs, decide the optimization strategy (what to fuse, what to replace, what to leave alone), drive the iteration loop, and maximize the sample score. This is the ENTRY skill — invoke it first for any PassNet optimization task; it tells you when to use passnet-pattern-fusion, passnet-triton-opt and passnet-feedback.
passnet-skill
PassNet GPU kernel optimization via compiler passes. Design and implement Triton-based optimization passes, create pass files under ./pass_dir/, self-evaluate with pass_evaluator, and iterate to maximize GPU speedup.
passnet-pattern-fusion
Author PassNet pass files: write patterns that actually MATCH the FX graph, pick fusion regions (which ops to absorb into one Triton kernel), and structure multi-pass submissions with the mandatory shared-dispatch architecture. Use when creating or fixing pattern()/replacement_args()/replacement_func() files, when a pass "failed to match", or when deciding how to fuse multiple kernels/ops.
passnet-orchestrate
Round/batch planning and decision support for PassNet graph optimization. Use for multi-sample triage, worker allocation, eval-budget policy, and post-eval keep/revert decisions. For solving one concrete sample end to end, use passnet-solve as the entry skill; consult this skill only when you need broader planning or a specific decision gate.
passnet-feedback
Fast, deterministic feedback for PassNet work: pre-flight pattern/match verification WITHOUT burning a GPU evaluation, per-node bottleneck analysis of a sample's graphs, and eval-log parsing into per-variant status + estimated score + failure classification. Use BEFORE every GPU evaluation (check_pattern), at the START of a sample (analyze_graph), and AFTER every evaluation (parse_eval_log).