ROI-Driven Foveated Attention for Unified Egocentric Representations in Vision-Language-Action Systems