Show HN: Automate Robot Data Quality Improvement

A new open-source toolkit just dropped to grade robot training videos like a fussy teacher, and the early vibe is cautiously hyped. It watches for blurry footage, jerky moves, and crash-y moments, then gives each episode a 0–1 score. There’s even an optional check using Google’s Gemini (an image-and-text AI) to judge if the robot actually did the task. One early commenter called it a “pragmatic” fix for messy data, and the crowd’s collective eyebrow is up: great idea, but does it actually make robots smarter?

That’s the headline drama: builders love the promise of auto-cleaning bad demos; skeptics want receipts. Expect demands for hard before/after charts showing training gains—baseline vs filtered. Also brewing: side-eye over depending on Gemini’s API limits and costs, and whether filtering out “weird” episodes might erase valuable edge cases. The vibe is very “please, no more vibe-based data”—score it, sort it, ship it.

Humor isn’t far behind either. With features reading like a robot report card—“visual clarity,” “smoothness,” “collision spikes”—folks joked about robotic Gordon Ramsay yelling “BLURRY!” at shaky clips. The repo is here for brave souls to try it: score_lerobot_episodes. Bottom line: the community’s ready to be convinced—show the gains, or this grade gets a redo.