DEFLECT：通过流匹配似然估计反事实调优实现延迟鲁棒的VLA策略执行

Yixiang Zhu, Yonghao Chen, Rui Meng, Jingyu Guo, Jiaxiang Zou, Zijie Yang, Taowen Wang, Xinyu Chen

2026

摘要

本文提出DEFLECT方法，通过离线后训练将推理延迟转化为无标签偏好信号，解决视觉-语言-动作（VLA）策略在异步部署中因预测与执行错位导致的性能崩溃问题。该方法在高速延迟场景下将成功率提升6.4%，并成功迁移至真实尺度VLA系统。

VLA policyasynchronous inferencedelay robustnesscounterfactual tuningflow matching

LEARNING

开放获取📊 1 引用

Keyi Shen, Glen Chou

2026

LEARNING

📊 0 引用

Xiwei Wu, Wei Wu, Qiqi Chen 等 9 位作者

Robotics and Computer-Integrated Manufacturing · 2026

LEARNING

📊 0 引用

Hedi Boukamcha, Anas Neumann, Monia Rekik 等 6 位作者

Robotics and Computer-Integrated Manufacturing · 2026

LEARNING

📊 0 引用

Jiaxin Liu, Xiaofeng Zhou, Suyang Yu 等 8 位作者

Robotics and Computer-Integrated Manufacturing · 2026