面向未知环境的频谱与激光雷达联合具身建图与导航

摘要：在卫星导航信号不可达的室内、地下等封闭场景中，利用已有无线通信基础设施所产生的信号场进行辅助导航成为重要需求。然而，传统基于频谱信号强度的导航虽然能够提供频谱梯度引导，但无法感知障碍物，在复杂室内环境中碰撞风险高；传统基于激光雷达的几何导航虽能有效避障，但缺乏目标方位的显式引导，易陷入盲目探索与路径冗余。针对上述问题，本文提出一种频谱与激光雷达联合具身建图与导航联合框架。该框架以多通道栅格张量表征环境状态，在感知层通过建图网络将稀疏频谱采样与局部几何信息联合反演为全局频谱信号势场与障碍概率图，打破单一模态导航的限制，实现目标引导与几何约束的有效融合；在决策层引入深度强化学习网络以捕捉历史轨迹与环境特征，将部分可观测状态映射至高维潜在决策空间并实现动作策略优化，缓解策略震荡与重复访问问题。仿真结果表明，在40×40室内栅格环境中经过5 000轮训练后，所提方法的导航成功率达到82%，相较基线方法显著降低了碰撞率与路径冗余度。进一步在多种地图尺度及室外低空参数配置下的泛化测试表明，该框架具备良好的场景适应能力，验证了多模态联合驱动框架的有效性与鲁棒性。本文以室内环境为典型验证场景，所提框架可推广至更大规模的低空通信网络覆盖场景。

关键词：无人机室内导航；频谱地图构建；强化学习；具身智能

Abstract: In enclosed scenarios such as indoor and underground environments where satellite navigation signals are inaccessible, the utilization of signal fields generated by existing wireless communication infrastructure for auxiliary navigation has emerged as a critical requirement. However, conventional spectrum-based navigation, while capable of providing spectrum gradient guidance, fails to perceive obstacles, resulting in a high collision risk in complex indoor settings. Conversely, traditional geometric navigation based on LiDAR can effectively avoid obstacles but lacks explicit guidance on target orientation, rendering it susceptible to blind exploration and path redundancy. To address these challenges, this paper proposes a joint framework for spectrum-LiDAR integrated embodied mapping and navigation. This framework represents environmental states using multi-channel grid tensors. At the perception layer, a mapping network jointly inverts sparse spectrum samples and local geometric information into a global spectrum signal potential field and an obstacle probability map, thereby breaking the limitations of single-modal navigation and achieving effective integration of target guidance and geometric constraints. At the decision-making layer, a deep reinforcement learning network is introduced to capture historical trajectories and environmental features, mapping partially observable states to a high-dimensional latent decision space and optimizing action strategies, which mitigates strategy oscillation and repeated visits. Simulation results indicate that after 5 000 training episodes in a 40×40 indoor grid environment, the proposed method achieves an 82% navigation success rate, significantly reducing the collision rate and path redundancy compared to baseline methods. Further generalization tests under various map scales and outdoor low-altitude parameter configurations demonstrate that the framework exhibits excellent scene adaptability, validating the effectiveness and robustness of the multi-modal joint-driven framework. This paper employs the indoor environment as a typical verification scenario, and the proposed framework can be extended to larger-scale low-altitude communication network coverage scenarios.

Keywords: UAV indoor navigation; map reconstruction; reinforcement learning; embodied intelligence