[ Data scoring for physical AI ]
Train on a fraction of the data.
Match the performance.
Score every demonstration by its real impact on your policy, so you know exactly what makes your model better.