Cue the Flow: Steering Flow-Matching Policies for Open-World Delivery Manipulation
Steers a pretrained flow-matching VLA with a spatial cue from a grounding model: a small adapter predicts an affine transform of the action chunk toward the cued target. Up to ~2x task success for open-world delivery.用视觉定位模型给出的空间提示引导预训练流匹配 VLA：轻量适配器预测作用在动作块上的仿射变换，使动作对准目标。开放世界配送任务成功率最高提升近 2 倍。
BibTeX
@misc{wang2026cue,
title = {{Cue the Flow: Steering Flow-Matching Policies for Open-World Delivery Manipulation}},
author = {Haoxuan Wang and Griffin Galimi and Junhua Huang and Selina Song and Wayne Wu and Yan Yan and Bolei Zhou},
year = {2026},
eprint = {2609.38989},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2609.38989},
}