Same Scene, Different Task: Skill Alignment for Compositional Generalization in VLAs
CRAFT is a counterfactual supervision method that helps Vision-Language-Action (VLA) models perform skill combinations they haven't seen during training. It uses existing demonstrations of individual skills to train the model on new, unseen sequences by creating 'counterfactual pairs.' This allows VLAs to generalize better to complex tasks without needing new data for every possible combination.