Multi-Dimensional Featurizers Yield Fine-Grained Steering and Data Shaping in Robot Foundation Models
The research investigates Block-Sparse Featurizers (BSFs) to uncover hidden, multi-dimensional features within vision-language-action models. BSFs successfully recover task-relevant concepts like perception, language, and control variables. These recovered features allow for precise steering of robot behavior and enable targeted data masking to remove specific behaviors from training data.