Astra：面向通用移动机器人的层次化多模态学习
机器人学
2025-06-09 v1 人工智能
摘要
现代机器人导航系统在多样化且复杂的室内环境中面临挑战。传统方法依赖多个模块或小模型或基于规则的系统，难以适应新环境。为此，我们开发了Astra，这是一套全面的双模型架构，包括Astra-Global与Astra-Local，用于移动机器人导航。Astra-Global是一种多模态大语言模型（LLM），处理视觉与语言输入，利用混合拓扑语义图作为全局地图，实现自我定位与目标定位，优于传统视觉姿态识别方法。Astra-Local是一种多任务网络，处理本地路径规划与里程计估计。其4D时空编码器通过自监督学习训练，生成用于下游任务的稳健4D特征。规划头采用流匹配与新颖的掩码ESDF损失，以最小化碰撞风险生成本地轨迹；里程计头部通过变压器编码器整合多传感器输入，预测机器人的相对姿态。我们在实际内部移动机器人上部署了Astra，跨越多样化室内环境实现高比例的端到端任务成功率。
引用
@article{arxiv.2506.06205,
title = {Astra: Toward General-Purpose Mobile Robots via Hierarchical Multimodal Learning},
author = {Sheng Chen and Peiyu He and Jiaxin Hu and Ziyang Liu and Yansheng Wang and Tao Xu and Chi Zhang and Chongchong Zhang and Chao An and Shiyu Cai and Duo Cao and Kangping Chen and Shuai Chu and Tianwei Chu and Mingdi Dan and Min Du and Weiwei Fang and Pengyou Fu and Junkai Hu and Xiaowei Jiang and Zhaodi Jiang and Fuxuan Li and Jun Li and Minghui Li and Mingyao Li and Yanchang Li and Zhibin Li and Guangming Liu and Kairui Liu and Lihao Liu and Weizhi Liu and Xiaoshun Liu and Yufei Liu and Yunfei Liu and Qiang Lu and Yuanfei Luo and Xiang Lv and Hongying Ma and Sai Ma and Lingxian Mi and Sha Sa and Hongxiang Shu and Lei Tian and Chengzhi Wang and Jiayu Wang and Kaijie Wang and Qingyi Wang and Renwen Wang and Tao Wang and Wei Wang and Xirui Wang and Chao Wei and Xuguang Wei and Zijun Xia and Zhaohao Xiao and Tingshuai Yan and Liyan Yang and Yifan Yang and Zhikai Yang and Zhong Yin and Li Yuan and Liuchun Yuan and Chi Zhang and Jinyang Zhang and Junhui Zhang and Linge Zhang and Zhenyi Zhang and Zheyu Zhang and Dongjie Zhu and Hang Li and Yangang Zhang},
journal= {arXiv preprint arXiv:2506.06205},
year = {2025}
}
备注
Astra Technical Report