feat(datasets): add dataset diagnostics CLIs (calibration drift + episode quality)

[maeste]

Summary / Motivation

This PR adds two complementary, read-only dataset diagnostic CLIs that audit recorded teleoperation datasets before GPU hours are spent training on them:

1. lerobot-check-calibration — detects leader/follower calibration drift. In teleoperation, action is the leader position and observation.state is the follower position; a calibration mismatch between the arms bakes a systematic offset into every frame. This is invisible during training (the policy learns the biased mapping, loss looks healthy) but breaks deployment, since inference feeds follower observations to a policy trained on leader-shifted targets. The tool compares the two signals on "stable" frames (robot not moving, so the follower controller has converged) and reports per-motor mean/std with a verdict. Real-world motivation: on an SO-101 setup, a ~17° offset on one joint (≈6 cm Cartesian error at the gripper) went completely unnoticed through recording and training — this finds it in seconds.

2. lerobot-dataset-quality — flags outlier episodes. Once a dataset grows past a few dozen episodes, reviewing each one in lerobot-dataset-viz is impractical. The tool computes deterministic per-episode metrics from the actions (duration, median/p95 jerk, peak velocity, static fraction, end-pose consistency), flags statistical outliers via the IQR rule, and prints a ranked worst-episodes list with a ready-to-edit lerobot-edit-dataset --operation.type delete_episodes command — turning "watch 200 episodes" into "watch these 8".

Both tools load any dataset via the standard LeRobotDataset API, read only the action/state columns (no image/video decoding, runs in seconds), require no new dependencies, and never modify the dataset.

Submitted as one PR since they share the same shape (read-only diagnostics following the lerobot-dataset-viz CLI conventions) and documentation page; supersedes #3759 and #3761.

Related issues

• Closes: [feat] lerobot-check-calibration: detect leader/follower calibration drift in recorded datasets #3758
• Closes: [feat] lerobot-dataset-quality: per-episode quality metrics and outlier flagging for recorded datasets #3760

What changed

• src/lerobot/scripts/lerobot_check_calibration.py: calibration drift analysis. Core logic in pure, importable functions (compute_episode_deltas, summarize_calibration, check_calibration) with a thin argparse main(). Flags: --vel-threshold (stability cutoff), --ok-threshold/--warn-threshold (verdict boundaries — configurable since action units are dataset-dependent), --arm-length-cm (optional Cartesian impact estimate), --output-format table|json.
• src/lerobot/scripts/lerobot_dataset_quality.py: per-episode quality metrics + IQR outlier flagging (compute_episode_metrics, detect_outliers, evaluate_dataset_quality). Flags: --k-iqr (outlier strictness), --top-bad, --output-format table|json.
• pyproject.toml: two new entry points, lerobot-check-calibration and lerobot-dataset-quality.
• tests/scripts/test_check_calibration.py + tests/scripts/test_dataset_quality.py: 22 unit tests on synthetic trajectories with known properties (known offsets, spikes, holds, divergent end poses), covering verdicts, outlier detection, episode grouping/sorting, and end-to-end reports including error paths.
• docs/source/using_dataset_tools.mdx: user-facing documentation sections for both tools.

No breaking changes; purely additive and read-only.

How was this tested (or how to run locally)

• Tests added: uv run pytest tests/scripts/test_check_calibration.py tests/scripts/test_dataset_quality.py -v (22 passed)
• pre-commit run passes on all touched files (ruff, mypy, bandit, typos, prettier).
• Manually verified on real SO-101 teleop datasets:
  • calibration check correctly flags a known miscalibrated dataset (17° offset on shoulder_lift) and reports ok on a freshly recalibrated one;
  • quality flags matched episodes already identified as bad by manual review (over-long struggled attempts, wrong end poses).
• Quick reviewer repro: lerobot-dataset-quality --repo-id lerobot/pusht (any dataset with an action feature), lerobot-check-calibration --repo-id <any teleop dataset with action + observation.state>.

Checklist (required before merge)

• Linting/formatting run (pre-commit run -a)
• All tests pass locally (pytest)
• Documentation updated
• CI is green
• Community Review: I have reviewed another contributor's open PR and linked it here: # (pending)

Reviewer notes

• The stability filter is the heart of the calibration method: on moving frames action - state is dominated by controller lag, so only converged frames are informative (compute_episode_deltas).
• The IQR rule was chosen over z-scores because episode metrics are typically skewed (a few very long episodes); it also flags nothing on a perfectly uniform dataset. median_jerk vs p95_jerk is deliberate: the median is robust to isolated corrective spikes, the p95 catches exactly those spikes.
• The two commits are kept separate (one per tool) for easier review; happy to squash, rename the commands, or split back into two PRs if maintainers prefer.

🤖 Generated with Claude Code