I built an open-source tool that finds bad robot demonstrations before training

I've been building Calibra, an open-source dataset observability and coreset optimization tool for robot learning.

It checks things like timestamps, duplicate/frozen frames, blurry cameras, jerky motion, action-state divergence, quality, and behavioral coverage.

The interesting part is the optimization side: across several public robotics datasets, we've found that quality-aware selection can preserve model performance while using substantially less training data than the full dataset or random selection.

I recently had someone try it on their real robotics dataset, inspect the episodes Calibra flagged, and confirm that the demos were actually bad. That was probably the most valuable validation I've gotten so far.

I'm now looking for robotics teams and researchers who are willing to try it on their own datasets and tell me thoughts.

https://www.producthunt.com/products/calibra-train-with-less-data