robovet 0.1.0

Vet your robot datasets: diagnose, repair, and quality-score LeRobot-format episode data before you waste a training run.

robovet

Vet your robot data. Diagnose, repair, and quality-score LeRobot-format datasets — before you waste a training run on broken episodes.

$ robovet doctor ./my_dataset

  FAIL DATA-104   1 episode where metadata 'length' disagrees with the parquet
                  row count — the classic signature of a corrupted episode map.
  FAIL STATS-302  1 stat block disagrees with the actual data — every training
                  run normalizes with these numbers.
  WARN TIME-202   Loading this dataset requires tolerance_s ≥ 7.7e-03
                  (77× the default). Worst: episode 2, 7.29 ms off the grid.

  3 fail · 4 warn · 22 pass
  UNSAFE TO TRAIN — fix the FAILs first.        (exit code 1 — CI-gate it)

Why this exists

Robot learning's bottleneck moved from models to data, and the data is quietly broken. An April 2026 audit of 10 popular open-source robot datasets found floating-point drift that breaks video decoding after ~45 episodes, a v2.1→v3.0 conversion bug that silently corrupts episode↔frame mapping (your run "works" — the policy just learns from jumbled sequences), datasets that only load with tolerance_s set to 100× the default, and no quality metrics anywhere. Hugging Face's own community-dataset cleaning run tells the same story: 111 of 240 datasets failed validation — and that pipeline is internal, not something you can run on yours. Meanwhile the 2026 consensus is that a well-curated 500-demo fine-tune beats a poorly-curated one at 10× the scale — curation tooling is the gap, not model size.

Every check in robovet maps to a documented, real-world failure. The receipts — issue numbers, audit findings, papers — live in PAIN.md.

Try it in 30 seconds (no robot required)

pip install robovet[video]

robovet demo ./demo          # synthetic SO-100-style dataset, 10 real-world
                             # defect classes injected (each tagged with the
                             # GitHub issue it reproduces)
robovet demo ./demo3 --v3    # same idea in the v3.0 shared-file layout
robovet doctor ./demo        # catches all of them; exit 1
robovet fix    ./demo --apply  # repairs the metadata class; .bak backups
robovet doctor ./demo        # metadata FAILs gone
robovet report ./demo -o report.html   # one self-contained, shareable page

robovet demo ./demo --clean builds the same dataset with zero defects, so you can see what all-green looks like.

What it checks

Group	Catches	Maps to
STRUCT-0xx	missing/invalid metadata, dangling episodes, orphan files	lerobot#761 (no validator for hand-rolled conversions)
DATA-1xx	episode↔frame mapping corruption, schema drift, NaN/Inf, dead dims	lerobot#2401 (silent v2.1→v3.0 corruption)
TIME-2xx	off-grid timestamps with the exact tolerance_s you'd need, non-monotonic time, cumulative FP drift	lerobot#933, lerobot#3177
STATS-3xx	stored normalization stats that disagree with the data ("normalization poison"), broken quantile stats (q01/q99)	HF docs warning; phospho repair post; lerobot#2189
VIDEO-4xx	video/parquet frame-count desync — including per-episode windows inside shared v3 files, codec-aware compatibility tiers (h264 ✓ / AV1 info — it's lerobot's own default / mpeg4-hevc warn), fps mismatch	Correll-lab postmortem; phospho notes

robovet doctor exits 1 on any FAIL and takes --json, so it drops straight into CI: gate dataset merges the way Codecov gates coverage.

Quality scoring (triage, not truth)

robovet score ./my_dataset --csv scores.csv

Per-episode signals, all computed in one pass: jerk smoothness, idle ratio, gripper chatter, duration outliers, action saturation, exact duplicates. This is deliberately the cheap first pass — the smoothness-first approach the 2026 curation literature (rinse, Demo-SCORE, QoQ) argues should precede expensive policy-rollout or influence-function filtering. Scores put the worst episodes in front of a human in seconds; review before you delete. Statistical flags carry practical-significance guards, so homogeneous datasets don't self-flag.

Repair contract

robovet fix is dry-run by default. With --apply it rewrites only metadata (episode lengths, normalization stats, info.json counters), backs up every touched file as .bak, never modifies parquet or video payloads, and preserves everything it doesn't understand: quantile keys (q01/q99 — the v3 QUANTILES-normalization era), image-stat blocks, and unknown episode fields such as tags. A repair tool must never be the thing that deletes your data; the test suite enforces these guarantees. Frame surgery (tail-trimming desynced episodes, timestamp re-gridding) is on the roadmap under the same contract.

Scope, honestly

• LeRobot v2.0 / v2.1 and v3.x are both first-class for diagnosis — each has its own synthetic fixture and test suite, and v3 gets per-episode video alignment inside shared files (VIDEO-405) plus per-episode stats checks parsed from the v3 metadata. fix currently rewrites v2.x episode metadata and global stats; v3 per-episode stats regeneration is on the roadmap.
• robovet does not merge/split/delete episodes — lerobot ships that natively now. We do what the official stack doesn't: deep validation, metadata repair, and quality triage.
• Local-first by design. Your data never leaves your disk — deployment-specific data is a competitive asset; treat it like one.

Library use

from robovet import load_dataset, run_doctor, score_dataset

ds = load_dataset("./my_dataset")
rep = run_doctor(ds)          # rep.exit_code, rep.results, rep.counts
sc  = score_dataset(ds, scan=rep.scan)   # reuses the same single IO pass

Apache-2.0. Issues and broken-dataset war stories very welcome — if your dataset breaks in a way robovet doesn't catch, that's a bug report we want.