APT: Action Expert Pretraining Improves Instruction Generalization of Vision-Language-Action Policies
The paper proposes Action Expert Pretraining (APT), a two-stage training method to improve how Vision-Language-Action (VLA) models follow instructions in new situations. It decouples the visual-action relationship from language conditioning by first training an action expert on vision and action data alone, then fine-tuning it with language tokens. This results in better generalization to unseen tasks.