Calibra: Train with Less Data
Train robot policies with up to 75% less data
Start new thread

Calibra Dataset Intelligence - Dataset-aware intelligence for robot training data

by•
Calibra v0.11 makes robotics dataset analysis more accurate and reliable. This release adds dataset profiles, so Calibra can adapt its analysis and pruning logic to how a dataset was actually collected instead of applying one global threshold everywhere. It also fixes and verifies key workflows across LeRobot, Isaac Lab, GR00T, HDF5, dataset export, calibration baselines, duplicate-frame detection, and deterministic pruning. The result: cleaner dataset decisions and training-ready exports.

Add a comment

Replies

Best
Maker
📌
Hey Product Hunt 👋 Calibra v0.11 is focused on one thing: making dataset decisions more trustworthy. The biggest addition is dataset profiles. Different robotics datasets can have very different action spaces, collection rates, teleoperation styles, and noise characteristics. Applying the same thresholds everywhere can make clean data look corrupted. v0.11 lets Calibra adapt to those differences while keeping global defaults unchanged. We also re-ran the main workflows end to end and fixed several issues across LeRobot exports, Isaac Lab → GR00T, HDF5 pruning, calibration baselines, duplicate-frame detection, and deterministic pruning. For example, PushT was previously being over-pruned because its normal motion characteristics crossed a global noise threshold. With the new profile, Calibra correctly recommends 175/206 episodes instead of incorrectly dropping 125 as corrupted. Would love feedback from anyone working with robotics datasets, imitation learning, or robot foundation models.