A Multimodal Agentic AI Framework for Environmental Reconstruction and Semantic Channel Modeling

Integrated sensing and communication (ISAC) and AI-based channel models have emerged as key technologies in the future sixth generation (6G) systems. However, existing AI-based channel models suffer from limited generalization capability, as they often overlook the underlying propagation mechanisms, fail to characterize small-scale fading by ignoring multipath components (MPCs), and rely heavily on accurate three-dimensional (3D) environmental maps when based on raytracing approaches. This work proposes an Agentic AI based modeling methodology that integrates multimodal information, combining semantic attributes from the vision modality(the information obtained from photographs), spatial locations of clusters-of-scattererss (CoSs) derived from channel data, and their geometric and topological relationships characterized by pivoted and hierarchical clusters within a unified framework. This design facilitates a comprehensive representation of multi-domain channel characteristics and offers a more accurate description of the underlying physical propagation mechanisms in complex environments, leading to improved generalization performance of the channel model. An intelligent clustering module groups MPCs by scatterer locations to extract CoS for target identification and environment reconstruction, while autonomously optimizing algorithms and parameters through a reward strategy according to the data characteristics. To validate the proposed method, a monostatic channel measurement campaign was conducted in a typical Urban Microcell (UMi) scenario. Experimental results demonstrate that the scatterers extracted from real channel data show strong correspondence with physical structures such as buildings and trees. The environment can thus be approximately reconstructed and semantically segmented. Finally, a semantic-aware channel model is developed, which accounts for the physical propagation mechanisms influenced by the environmental information of the MPCs, without requiring an environmental map. This work establishes a standard agentic AI framework for future ISAC research, integrating data acquisition, parameter estimation, semantic channel analysis, and channel modeling.