Run your first local audit.

Choose a folder, run the checks in your terminal, and read the evidence before training. No account or data upload required.

1. Install the local toolkit

Use Python 3.9–3.12. This installs openbot-data with its optional LeRobot support: [lerobot] adds pyarrow for reading Parquet files. Your data stays on your machine.

terminal

pip install 'openbot-data[lerobot]'

Commands below use macOS/Linux shell quoting. For Windows setup and troubleshooting, read the toolkit guide.

2. Choose the folder and checks

Already have data? Enter its root folder below. For LeRobot, this is the folder containing meta/info.json, not an individual episode or video.

Need a dataset to try?

Download and extract this small synthetic LeRobot v3 example. It contains one episode, two state/action frames and no videos. It is for learning the reports, not training. Missing optional statistics is an expected warning.

After extracting it, use ./lerobot-v3-smoke as the local folder and confirm v3.0 below.

Download example ZIPChoose another dataset

Local dataset folder

Replace ./lerobot_dataset with your own folder. This path is used only to compose commands in your browser.

Local format

Integrity scope

Version in your local meta/info.json

Zero errors only describes the checks that ran. The report’s coverage shows what was checked, skipped or unavailable.

Planned coverage · Full

Structure: Adapter-supported checks. The report lists missing dependencies and skipped checks.
Video evidence: Full decode pass for discovered videos. Large datasets can take time.
Core readiness: Confirm v3.0 in meta/info.json. LeRobot v2.1 supports audit only.

Preview only. No files have been checked; task performance still needs separate validation.

audit → ./audit.json

openbot-data audit \
  --format lerobot \
  --integrity full \
  --out ./audit.json \
  --fail-on error \
  -- './lerobot_dataset'

A core readiness command requires confirmed LeRobot v3 and full integrity. For other formats or scopes, review the audit findings and coverage.

3. Read the evidence and decide what to do next

Reports are written to your terminal’s current folder. Open audit.json for findings and coverage; open readiness.json for the profile decision. An audit with error findings exits with code 2 after saving its report.

Example audit excerpt — not your result

{
  "schema_version": "openbot.dataset_audit.v1",
  "summary": { "error": 0, "warning": 1, "info": 0, "videos": 0 }
}

This published fixture has one warning about missing optional statistics. Read the actual findings and coverage before deciding; a zero count alone does not establish training suitability.

Error findings or BLOCKED: Review each finding and its evidence. Correct the source or create a repair plan, then rerun the checks before training.
Warnings or PARTIAL coverage: Review the missing evidence and skipped checks. Use full integrity when required; do not treat unchecked capabilities as passed.
READY for lerobot-core: The LeRobot v3 structural profile passed within the reported scope. Run ACT or SmolVLA readiness for your intended policy, and verify task performance separately.

Read complete example reports Policy profiles and repair workflow