💡 This paper proposes HiVLA, a visual-grounded-centric hierarchical embodied manipulation framework that decouples high-level VLM semantic planning from low-level flow-matching DiT action execution, add