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

Calibra: Cut Robot Training Costs - Find bad data. Train on less. Save GPU.

by•
Calibra helps robotics teams audit dataset integrity, measure quality and coverage, and build smaller training sets. v0.8 adds measured training experiments: record real GPU-hours, wall-clock time, energy, and policy performance, then compare Full vs Random vs Calibra across retention levels.

Add a comment

Replies

Best
Maker
📌
Hello everyone! I'm Omer, creator of Calibra. With the The biggest change in v0.8 is that we're moving from predictions to measurement. Previously, Calibra could estimate what might happen if you reduced a dataset. Now teams can record their actual training results — GPU-hours, wall-clock time, energy and policy performance — and compare: Full dataset vs Random subset vs Calibra subset across 10%, 25%, 50%, 75% and 100% retention. This is the workflow we're using with our early design partners to answer a much more important question: Does Calibra actually save compute while preserving policy performance? We're still validating this across different datasets and robotics setups, so we'd love feedback from researchers and engineers working with robot-learning data.