MessyMem: Learning-from-Doing Memory for Mobile Manipulation
Persistent memory for mobile manipulators: a 3D scene graph updated with what the robot learns by acting (locked cabinets, object locations). Reports 80% progress over a 25-task, 3-hour run.移动操作机器人的持久记忆：在三维场景图中记录通过交互学到的知识（如柜子上锁、物体位置）。报告在连续 25 个任务、3 小时的仿真中任务进度达 80%。
BibTeX
@misc{banwasi2026messymem,
title = {{MessyMem: Learning-from-Doing Memory for Mobile Manipulation}},
author = {Anuva Banwasi and William Muckelroy and Priya Sundaresan and Linfeng Zhao and Jeannette Bohg and Cherie Ho},
year = {2026},
eprint = {2609.15976},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2609.15976},
}Navigation & Embodied Agents导航与具身智能体Manipulation & Dexterous Hands操作与灵巧手Robot arm机械臂Mobile robot移动机器人memorymobile-manipulationscene-graphlong-horizonreal