The emergence of multimodal large language models (MLLMs) has laid the foundation for the vision-language-navigation paradigm, which integrates visual perception, natural language understanding, and navigation control within a unified strategic framework. This paradigm has been rapidly adopted in the UAV domain, attempting to enable UAVs to understand natural language instructions, reason in three-dimensional environments, and make flight decisions. Compared with traditional modular navigation approaches, the end-to-end framework based on MLLMs can simultaneously process linguistic and visual signals, directly mapping perceptual information into control commands. However, a systematic review of UAV-VLN remains scarce. This paper presents a comprehensive review of recent advances in this area: from early modular solutions to reason-centric vision-language-action models. It elucidates how visual, linguistic, and control information are progressively integrated to enhance autonomous navigation capabilities. It further summarizes existing datasets and evaluation protocols, including simulation tasks in both indoor and outdoor complex environments as well as real-world UAV flight trajectories, with evaluation metrics covering success rate, time cost, and semantic comprehension. Finally, it identifies key challenges, including difficulties in cross-modal alignment, insufficient real-time responsiveness in dynamic environments, high annotation costs, and poor decision-making robustness in complex scenarios. This paper outlines new pathways and future research directions for UAV autonomous navigation research. It highlights the potential of MLLMs in enhancing intelligent decision-making and interpretability of UAVs, and serves as a reference for research and practice in achieving safe and efficient autonomous flight of UAVs.

Cross-city transfer learning (CCTL) has emerged as a crucial approach for managing the growing complexity of urban data and addressing the challenges posed by rapid urbanization. This paper provides a comprehensive review of recent advances in CCTL, with a focus on its applications in urban computing tasks, including prediction, detection, and deployment. We examine the role of CCTL in facilitating policy adaptation and influencing behavioral change. Specifically, we provide a systematic overview of widely used datasets, including traffic sensor data, GPS trajectory data, online social network data, and map data. Furthermore, we conduct an in-depth analysis of methods and evaluation metrics employed across different CCTL-based urban computing tasks. Finally, we emphasize the potential of cross-city policy transfer in promoting low-carbon and sustainable urban development. This review aims to serve as a reference for future urban development research and promote the practical implementation of CCTLs.

The Autonomous Modular Bus (AMB) introduces an innovative approach to public transportation by allowing modular buses to dock and undock seamlessly while in motion. This capability effectively alleviates traffic congestion and decreases energy usage through smoother and more efficient vehicle operation. However, achieving autonomous docking for AMBs poses significant challenges, including the need for precise localization in both horizontal and vertical dimensions and the ability to manage dynamic persistent obstacles in close-range scenarios. Existing Light Detection and Ranging (LiDAR)-based Simultaneous Localization and Mapping (SLAM) algorithms, such as LIO-SAM, perform well in static environments but encounter limitations in dynamic scenarios, particularly with occlusions and vertical drift during AMB docking. In this paper, we propose an enhanced LiDAR-Inertial Measurement Unit (IMU) SLAM framework focused on improving localization accuracy and robustness during AMB docking. Key contributions include: (1) A two-stage scan-to-map matching method with ground constraints to reduce z-axis drift; (2) A factor graph optimization strategy integrating IMU roll and pitch constraints and periodic resetting to mitigate long-term drift; (3) A deep learning-based front vehicle detection and point cloud filtering mechanism to reduce occlusion effects. Experimental evaluations on single-vehicle and dual-vehicle datasets demonstrate that our method significantly reduces Absolute Pose Error (APE) and Relative Pose Error (RPE) compared to existing methods. These results highlight the framework's ability to address the unique challenges of AMB docking, therefore helping alleviate traffic congestion and reduce energy consumption.