【摘要】 为应对深空探测中月面任务的复杂性挑战，本文围绕空间智能机器人自主导航问题开展综述，重点聚焦非结构化环境下的无地图策略建模与系统实现路径。首先，系统梳理了月球表面导航中的三类核心技术路径，包括环境感知方法、基于地图的传统导航策略以及无地图自主导航方法，揭示各自的技术构成与关键瓶颈。其次，基于部分可观测马尔可夫决策过程(POMDP)建模范式，构建一套覆盖状态表达、观测建模、奖励结构与策略表示的无地图导航系统框架，结合深度强化学习方法，实现“感知—决策—动作”的一体化闭环策略学习机制。第三，为支撑相关方法的开发与验证，设计并开源了astro＿nav＿sim仿真平台，集成典型月球与火星地形模型、Xmobot多驱动机器人配置与ROS接口，构建具备代表性、可复用的测试基准环境。最后，展望了多源感知融合、轻量化策略部署、分层决策建模、系统安全保障与基于大语言模型的导航等关键技术的发展趋势，旨在为构建高自主性、强泛化能力与任务适配性的空间智能机器人导航技术提供方法支持与工程参考。更多还原

【Abstract】 To address the challenges of lunar missions in deep space exploration, autonomous navigation technologies for space intelligent robots were reviewed, with a focus on mapless strategies in unstructured environments. Firstly, three major technical pathways were systematically comparatively analyzed, including environmental perception methods, traditional map-based strategies, and mapless navigation techniques. Their key components and limitations were summarized. Secondly, a mapless navigation framework was proposed based on the partially observable Markov decision process(POMDP), covering state representation, observation modeling, reward structure, and policy formulation. By integrating deep reinforcement learning, a closed-loop learning mechanism for perception-decision-action was established. Thirdly, to support algorithm development and validation, the astro＿nav＿sim simulation platform was designed and open-sourced, incorporating typical lunar and Martian terrain models, Xmobot multi-drive robot configuration, and ROS interfaces to provide a representative and reusable testing environment. Finally, the future trends in key technologies was pointed out, such as multisource perception fusion, lightweight policy deployment, hierarchical decision-making, system safety assurance, and large language model-based navigation. These perspectives aim to support the development of highly autonomous, generalizable, and task-adaptive intelligent navigation systems for space robotics.更多还原