Memorize, Adapt, Ignore: Diagnosing Robot Learning Mechanisms under Training Data Variation
The study investigates how varying training data affects a robot's learning policy by using representation kernels to diagnose internal representations. By systematically changing factors like colors or textures, researchers found metrics that distinguish between memorizing specific situations, adapting to new variations, and ignoring irrelevant noise. This helps understand when a model generalizes well versus when it overfits.