A Survey of Map Construction and Management Methods for Single Embodied Agents

Abstract

Embodied intelligent systems need to complete the closed loop of perception–understanding–decision–execution in the real world. One of their core capabilities is constructing, maintaining, and utilizing environmental maps. Maps not only support localization and navigation but also directly influence obstacle avoidance, safe interaction, mobile manipulation, long-term memory, and task planning. With the rapid development of multi-sensor fusion Simultaneous Localization and Mapping (SLAM), semantic understanding, and neural scene representations, robot maps have evolved from traditional two-dimensional occupancy grids to three-dimensional point clouds, voxel/octree, distance fields, mesh models, topological maps, object-level semantic maps, and dynamic scene graphs, further progressing toward continuous implicit maps centered on Neural Radiance Fields (NeRF), 3D Gaussian Splatting (3DGS), and open-vocabulary/foundation models. However, different map representations exhibit significant differences in geometric accuracy, queryability, updatability, computational/storage overhead, and support for downstream tasks. Meanwhile, long-running embodied systems also face engineering and theoretical challenges such as dynamic environments, sensor drift, multi-session fusion, version management, and forgetting/compression. This paper, oriented toward embodied intelligence applications, systematically reviews map representation forms, corresponding construction methods, and map management and update mechanisms. In terms of representation, a hierarchical classification of geometric/metric maps, semantic/structural maps, and neural/foundation model maps is presented, and their adaptability in navigation, manipulation, and language interaction is discussed. In terms of construction, traditional probabilistic mapping, geometric registration (ICP/NDT), graph optimization and factor graphs, semantic fusion and scene graph construction, as well as online optimization methods for neural implicit maps and 3DGS maps are systematically reviewed. In terms of management, key technologies such as submap management, map indexing and compression, dynamic environment updating, change detection, multi-session map fusion, and neural map continual learning and catastrophic forgetting mitigation are summarized. Finally, the paper concludes and envisions future development trends, emphasizing that hierarchical hybrid maps, task-driven cognition, and long-term continuous maintenance will be the key to deploying embodied intelligence in real-world settings.