Lamport Sieve audits robot teleop data: deleting bad demos made training worse
R4nsack · x · 2026-11-08
Lamport Robotics launched Lamport Sieve, a data audit tool for robot training: it scores every LeRobot episode from joint signals (smoothness, stalls, command tracking) and camera frames (model judges task completion, collisions, camera quality), producing HTML/JSON reports plus a cleaned dataset copy.
Key findings from auditing 400 SO-101 teleop demos: 10 dropped for drops/collisions, 73 flagged, and an unnoticed wrist servo mis-calibration spanning 3 tasks. Counterintuitively, deleting rejected episodes did not reliably lower held-out loss and even raised it on one task — with only 40 demos, less data costs more than a bad demo. Keeping rejects at 1/5 weight beat both approaches. Free tier: public LeRobot v3 datasets up to 500 episodes, two audits per day.
More from Embodied
- GenrobotAI's DAS Dex 2.0 captures 21-DoF hand motion plus 980 tactile points for robot learning — CyberRobooo · 2026-11-08
- Gelato open-sources an Apache-2.0 protocol linking AI models to phones, glasses, and cameras — Milo_Bindu · 2026-11-08
- Force sensor + agentic harness one-shots blind precision insertion at industrial reliability — ihorbeaver · 2026-11-08
- Wheeled humanoid Theo restocks 400 items/hour in a Zurich supermarket — Distinct-Question-16 · 2026-11-08
- WSJ review: the $400 ProLED XDR MacBook Pro upgrade is worth it — whurley · 2026-11-08
- RoboCap adds tactile sensing via Matrix Gloves for dexterous robot learning — micoolcho · 2026-11-08