Hi everyone! I'm Omer, the founder of Calibra. 👋 I built Calibra after realizing that robotics teams spend huge amounts of GPU time training on datasets without knowing whether the data is trustworthy, diverse, or full of redundant demonstrations. Calibra helps answer these four questions before training starts: > Can I trust this dataset? (Integrity) > Is it clean? (Quality) > Is it diverse enough? (Coverage) > Can I train with less data? (Quality-aware coreset selection) In our PushT benchmark, Calibra retained just 25% of the data while achieving 99.5% of full-data prediction performance, suggesting that many robotics datasets contain significant redundancy. The project is open source and already supports LeRobot, Isaac Lab, robomimic, RLDS, MCAP, and HDF5 datasets. I'd genuinely love your feedback. Thanks for checking it out!

Calibra: Train with Less Data

Train robot policies with up to 75% less data

Train robot policies with up to 75% less data