Task-Related Mapping for Ground Autonomous Platforms: Current Status and Development Trends

Objectives Driven by the escalating demands in autonomous driving, autonomous combat, and other cutting-edge fields, task-related mapping has become a linchpin for enhancing the decision-making precision and operational efficiency of ground autonomous platforms. The aim is to conduct a systematic review of the current research status, technical challenges, and development trends of task-related mapping, with a particular focus on four core components: perception data filtering, prior information filtering, map model design, and modeling methods. By establishing a standardized mapping index system, efforts are made to address the deficiencies of existing task-related mapping approaches and provide a theoretical foundation for mapping tailored to autonomous platforms. Methods Task-related mapping is systematically summarized and evaluated from four perspectives. For perception information relevance filtering, an in-depth exploration of the technical details of the mathematical rule-based Information Bottleneck theory and the neural network-based cross-attention mechanism is carried out, emphasizing semantic depth exploration and multimodal data processing to extract task-related perception data. Regarding prior information relevance filtering, the fusion and filtering methods for rule-based, neural parameter-based, and geospatial prior data are analyzed. Knowledge graphs and neural network technologies are leveraged to enhance the integrity and accuracy of prior data integration. In terms of task-related design of map models, strategies for designing map models with embedded user layers, task reasoning layers, and decision-making layers are explored, aiming to strengthen the hierarchical logical structure of map models. Concerning task-related mapping, it examines task-driven mapping approaches, such as pre-training-based mapping and spatial perception engine construction, are examined to tackle the challenges of real-time modeling and multi-source data fusion. Results Task-related filtering of perception data emphasizes calculating the correlation between task information and perception data from the dimensions of semantic depth and multimodal data processing, ensuring the precision of taskrelated perception data extraction. Task-related filtering of prior information focuses on effectively integrating heterogeneous multisource prior data via knowledge graphs or neural network methods, resolving issues of prior data heterogeneity and alignment. Taskrelated model design significantly improves the model's task adaptability and dynamic adjustment capabilities by embedding "task feature modules," providing robust support for task-related mapping. Task-related mapping, with the goals of task-related data filtering and efficient modeling, emphasizes the importance of task-related learning performance and holistic integration of the modeling process. Conclusions Based on a comprehensive review, a task-related mapping index system encompassing perception data, prior data, map models, and modeling methods is proposed, offering standardized metrics for task-related mapping. It also conducts a comparative analysis of existing works, such as HDMap mapping, 3DSG mapping, and Active SLAM mapping. The future development directions of task-related mapping are outlined, including multi-dimensional hierarchical map models, fine-grained multimodal perception algorithms, deep multi-source data fusion, and task-driven real-time modeling engines, which will promote the application of task-related mapping in complex autonomous systems.