The author introduces Calibra, an open-source toolkit designed to analyze and ensure the integrity of robot-learning datasets. The tool aims to prevent wasted compute by identifying problematic data before training begins.
- duplicate or redundant demos
- frozen camera frames
- jittery robot motion
- calibration drift
- corrupted or inconsistent data
Calibra provides dataset observability and coreset selection capabilities for robotics imitation learning, with a Hugging Face Space available for health checks.