CarryOn: A Dual-Loop Harness for Personalized and Reliable Robot Assistance
Wraps a frozen VLA in two loops — one keeps user preferences and scene memory for planning, one blocks hazardous actions and calls recovery skills; reports violations down from 52.6% to 6.1%, recovery up from 13% to 80%.给冻结的 VLA 套上双环框架：规划环保存用户偏好与场景记忆以做个性化任务分解，执行环拦截危险接触并调用可复用的恢复技能；报告安全违规率从 52.6% 降到 6.1%，失败恢复成功率从 13% 提升到 80%。
BibTeX
@misc{wang2026carryon,
title = {{CarryOn: A Dual-Loop Harness for Personalized and Reliable Robot Assistance}},
author = {Zixuan Wang and Yuqi Liu and Yuxin Chen and Sengqiao Yang and Jinhui Ye and Pengguang Chen and Changsheng Lu and Shu Liu and Bei Yu and Jiaya Jia},
year = {2026},
eprint = {2610.21453},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2610.21453},
}VLA & Robot Foundation ModelsVLA 与机器人基础模型Manipulation & Dexterous Hands操作与灵巧手agenticharnesspersonalizationsafetyfailure-recoverysimreal