adduct

Adduct finds the defects in your robot demonstration data that will break your policy — before you spend four hours training.

$ adduct scan lerobot/pusht

adduct  ·  lerobot/pusht
lerobot_v3 · 206 episodes · unknown · 10 Hz · action_dim 2
╭──────────────────╮
│ 3 MEDIUM   3 LOW │
╰──────────────────╯
by family: smoothness 3  dynamics 1  stats 1  temporal 1

MEDIUM ▸ The dataset's distribution shifts partway through collection
         → Check for a recalibration, tool change, or operator change midway; consider
treating the segments as separate datasets.
           206 eps  [stats.distribution_drift]

LOW    ▸ State evolves inconsistently with the actions in 2.1% of transitions (206
episodes)
         → Inspect the flagged segments; drop episodes containing resets or frame drops.
           206 eps  [dynamics.forward_residual]

MEDIUM ▸ Same state, different next action in 74 episode(s)
         → Prefer action chunking (ACT) or Diffusion Policy over plain BC.
           74 eps  [temporal.non_markovian_pause]

MEDIUM ▸ Shaky teleoperation in 9 episode(s) (up to 6.3× median jerk)
         → Re-record or smooth the flagged episodes; consider a low-pass filter on
teleop input.
           9 eps  [smoothness.jerk_outlier]

LOW    ▸ 4 episode(s) wander: up to 18.6× the direct path
         → Review the flagged episodes; re-record the ones where the operator was
searching.
           4 eps  [smoothness.path_efficiency]

… 1 more finding(s) — --html or --json for all.
Next: adduct scan lerobot/pusht --html report.html --open

That is a real, unedited run against a real public dataset — about two seconds on a laptop with a warm cache, no GPU, and no video decoded. Findings are ordered by severity × blast radius, so a dataset-wide LOW can outrank a narrow MEDIUM: the top of the list is what to look at first, not simply what is loudest.

Install

pip install adduct
adduct scan lerobot/pusht

Python 3.10–3.13. Point it at a local directory or a Hugging Face Hub owner/name.

Add @revision to pin a Hub dataset to one commit, so a result stays reproducible even if the dataset is re-uploaded under the same id:

adduct scan lerobot/pusht@7628202a2180972f291ba1bc6723834921e72c19

What it checks

Adduct ships 48 detectors across 12 families — 45 run by default, 3 are held back pending recalibration (see below) and reachable with --all. What makes a finding useful is not that something is statistically unusual — it is the mechanism: why this specific defect degrades a trained policy. Every finding carries one, plus the measured value, the threshold it crossed, the affected episodes, and a concrete fix.

Run adduct list-detectors for the full set, or adduct explain <id> for the mechanism behind any one of them.

What it does not do

Being clear about this matters more than any feature claim:

• No simulator, no training, no GPU. It reads your data and does statistics on it.
• No network, except the explicit Hugging Face fetch when you pass a owner/name repo id. Local scans make zero network calls.
• No telemetry. Nothing about your data, your findings, or your usage is transmitted anywhere. Ever. There is no opt-out because there is nothing to opt out of.
• It never decodes video by default. It reads Parquet columns, which is why it is fast.
• It does not catch everything. Some failure modes only show up in a rollout — a policy that is subtly bad at a contact-rich sub-task will look fine to every static check here. Adduct rules out a class of data problems; it does not certify that a dataset is good.
• The findings are not calibrated against training outcomes yet. They are grounded in documented failure mechanisms, not in a corpus of runs that measures how much each defect actually costs. Treat severity as a triage ordering, not a prediction.
• Three detectors are excluded from a default scan. Each was measured over-firing on a 20-dataset sweep of curated public LeRobot data, and gating beats deleting. smoothness.discontinuity_jump and integrity.declared_mismatch reported HIGH on 70% and 60% of that corpus. dynamics.inverse_residual fired on 100% of it and reported HIGH on 50%; excluding it drops the corpus from 23 HIGH findings to 13, and the datasets carrying at least one HIGH from 17 of 20 to 11 of 20. Pass --all to run them anyway; nothing is deleted, just held back pending recalibration. Please report them as false positives if they fire on data you trust. The evidence and the method are in benchmarks/.

Measured evidence

The unit suite plants known defects in synthetic fixtures and checks they are found — that measures recall, and it cannot measure precision on healthy data. benchmarks/ holds the runs that measure the other half, on real published datasets.

The latest, A precision audit of robot data-quality detectors (20 datasets, 4,342 episodes, 545,964 frames): all 20 parsed without error in 22.5 s on a laptop CPU, producing 190 findings.

Its most consequential result is not about adduct. 18 of the 20 datasets declare robot_type: "unknown", so any method conditioning on embodiment via Hub metadata collapses to a single bucket on 90% of the corpus — including adduct's own conformal gate.

It is also unflattering about the tool: six detectors fire on 70% or more of curated public data, a median of 9.5 findings per dataset is too many to act on, and one hypothesis for the worst offender was tested and rejected rather than quietly dropped. It is explicit about what it does not show: a fire rate is a rate of complaint, not a rate of correctness, and no policy was trained, so the premise that these defects degrade policies is untested there.

Supported formats

Using something else? Tell us which format — what gets built next is decided by what people actually have.

Found a false positive?

Please report it. This is the most valuable thing you can send us.

Adduct's only real asset is that its findings are trustworthy, and a detector that cries wolf is worse than no detector. We cannot find those on synthetic data — we need yours.

👉 Report a false positive

You do not need to share your dataset. The detector id, the numbers adduct printed, and why it is wrong is enough to act on.

Using it in CI

By default adduct scan exits 0 whether or not it finds anything, so it never breaks a pipeline you did not ask it to gate. Opt in explicitly:

adduct scan ./data --ci --fail-on HIGH    # exit 1 only when a HIGH finding is present

Machine-readable output: --json (stable, versioned schema_version), --sarif (SARIF 2.1.0 for GitHub code scanning), --html (a self-contained report). Findings go to stdout; errors, notices, and progress go to stderr, so adduct scan ./data --json - | jq works cleanly.

Python API

import adduct

report = adduct.scan("./my_lerobot_dataset")
for cluster in report.clusters:
    print(cluster.severity, cluster.title)

report.to_json("report.json")

Extending it

Adapters and detectors are plugins discovered through entry points — the same mechanism the built-ins use, with no privileged path for first-party code:

[project.entry-points."adduct.detectors"]
"myteam.my_check" = "my_pkg.checks:MyCheck"

Contributing

See CONTRIBUTING.md. Commits need a DCO sign-off (git commit -s); there is no CLA, and you keep the copyright to what you write.

Security issues: please report privately per SECURITY.md, never as a public issue.

License

Apache-2.0. Copyright 2026 Adduct.