Context-Aware Query for Semantic Maps Transmission in Embodied Agent Systems

Abstract

Semantic maps provide essential structured environmental information for embodied agents, but their transmission over wireless channels faces challenges from signal degradation and bandwidth constraints. Recent advances in semantic communication have improved resilience against channel impairments. However, when applied to semantic map reconstruction for embodied agents, these methods typically process semantic categories independently and lack explicit modeling of spatial-contextual dependencies critical for preserving finegrained details like object boundaries. To address this limitation, we propose a semantic communication framework for highfidelity semantic map reconstruction via context-aware query refinement. Our approach integrates joint source-channel coding at the transmitter with a decoder that does not require a shared knowledge base at the receiver. The decoder employs learnable memory tokens to capture semantic priors, while query tensors progressively refine noisy features through attention mechanisms. This design explicitly models interdependencies among semantic classes, enabling precise reconstruction with preserved structural details. Extensive experiments under diverse channel conditions demonstrate that the proposed framework consistently outperforms baseline methods and achieves robust semantic map reconstruction.