LamportLab
Data audit

Know what is in a dataset before it trains anything.

Paste a LeRobot dataset. Every episode is scored from its joints and looked at through its cameras. You get a report that says which episodes to keep, which to drop, and why, plus a clean copy.

Public LeRobot v3 datasets up to 500 episodes, two a day per visitor, free. The dataset is downloaded, audited and deleted; the report keeps eight key frames per episode. Larger or private sets: support@lamportrobotics.com.

What is checked

From the joints

  • Smoothness: log dimensionless jerk and spectral arc length
  • Stalls mid-episode, idle head and tail
  • Time near a joint limit, time outside the calibrated range
  • Command tracking error, reversals per second, command jumps
  • Duration against the task median, dropped frames

From the cameras

  • Key frames from each camera plus the final frames at full size
  • The model lists what happens frame by frame, then judges
  • Does it attempt the instructed task; does it finish
  • Drop, collision, more than one attempt
  • Dark, blown-out or static cameras

What you get

  • report.html: every episode with its frames, verdict and reasons
  • report.json: keep list, reject list with reasons, rejection rate
  • episodes.json: every score and the model's note, per episode
  • A clean LeRobot v3 copy without the rejects, on request
  • Thresholds are plain numbers, listed with the tool
Sample reports · 400 SO-101 teleop demos, 8 tasks

Source data from the ArmnetBench reference sets. Audited copies with the audit files inside are on the Hub. On these sets, dropping the rejected episodes did not reliably lower held-out loss, and on one task it raised it: with forty demonstrations, less data costs more than a bad demonstration does. Keeping the rejects at a fifth of the weight did better than both on a first run of each task. The report tells you which episodes those are; the weights file is part of it.

Plans

Free

$0

Up to 500 episodes per dataset, two a day. Report and JSON. Public datasets only.

Monthly

from $3,000

Private datasets, no episode cap, clean copies and format conversion (v2.1, HDF5, RLDS to LeRobot v3), API access, your own thresholds.

Signed audit

$5,000 to $20,000

We run it, review every reject by hand, and sign the report with the rejection rate. For a dataset you are about to deliver or buy.

Write to support@lamportrobotics.com. Bangkok, UTC+7.