Your loss curve looks perfect but the robot pecks at the table. The bug might be in your data, not your model.
lerobot-lint is a CLI (lelint) and Python package (lerobot_lint) that checks
LeRobot robot-training datasets for
behavioral/kinematic data bugs: dead joints, encoder wraparound, frozen telemetry,
leader/follower calibration mismatch, frame drops, before you burn GPU-hours training
on bad data.
Deterministic, rule-based checks. No ML, no training, no GPU required to run it.
20 checks across 5 groups are implemented and tested (130 passing tests). The CLI runs
end-to-end against real Hugging Face Hub datasets: loading, checking, console/JSON
reports, exit codes, a guard() context-manager API for use inside training scripts.
Known limitation: camera/video checks (frozen camera, degenerate image, video length mismatch) are implemented and unit-tested against synthetic frame data, but real video decoding isn't wired up yet, so they don't see real camera frames from an actual dataset run today. Everything else (joint kinematics, action/state consistency, timing, dataset hygiene) runs against real data now.
pip install lerobot-lintlelint check lerobot/some-dataset --no-videoAuto-detects a robot profile (so101/koch joint-naming convention) from the dataset's
own metadata when --profile isn't passed, and always tells you which profile it
used and why, never silently — the disclosure appears in the report itself, not just
before the run. --fail-on info|warning|error controls what makes the exit code
nonzero (default: error, exit 0/1/2). --json report.json writes a machine-readable
report alongside the console one.
LeRobot datasets don't declare their joint-state units, and the Hub mixes radians,
degrees, and normalized values. lelint infers the units from the observed state
extent (disclosed as a UNITS_INFERRED finding), converts degrees to radians before
running kinematic checks, and skips velocity-threshold checks — with one loud
finding instead of a false-positive storm — when the scale looks like raw encoder
counts. Override the inference with --units radians|degrees|normalized.
import lerobot_lint
with lerobot_lint.guard("your/dataset"):
... # raises before you spend GPU-hours on data with real errorsRunning it against lerobot/svla_so101_pickplace
(a real SO-101 pick-place dataset) surfaces genuine recorder dropouts undisclosed in
the dataset card: 5 of 50 episodes have stretches where the robot's telemetry is
bit-identical for up to 2+ seconds while actively recording (verified by hand against
the raw data, not just this tool's own output):
$ lelint check lerobot/svla_so101_pickplace --no-video
Using profile: koch (auto-detected from joint names, override with --profile)
...
Errors (...)
[FROZEN_STATE] episode 0: State frozen (identical for 67 consecutive frames) from
frame 27 to 93 -- likely a recorder hiccup or serial dropout
An imitation-learning policy trained on episode 0 as-is would learn "do nothing" for over a quarter of that episode's frames.
- All 20 checks across 5 groups (kinematic signal, action/state consistency, timing
integrity, camera/video, dataset hygiene), see
lerobot_lint/checks/. - The dataset loader, verified against real public LeRobot datasets (including real SO-101 hardware, not just simulated tasks).
- The check engine: two-pass streaming, per-check crash isolation, episode- and dataset-scoped checks.
- The
lelintCLI:check(with--profile,--episodes,--no-video,--json,--verbose,--fail-on),profiles,version. lerobot_lint.guard(), a context-manager API for use inside training scripts.- Console and JSON reports.
- Robot profile auto-detection from joint-naming convention.
The A-F scorecard formula, the bug-card renderer, real video-decode wiring, CI, and the full field study (calibrating thresholds against 30-50 real datasets).
Apache-2.0, see LICENSE.