Semantic Inpainting-Driven
Sensing–Communication Trade-off
Distributed Cross-Modal ISAC
SSL-JEPA
The integration of sensing and communication in future wireless systems requires efficient management of limited network resources, particularly when handling large volumes of redundant multimodal sensory data. In this paper, we propose a novel semantic inpainting-based framework for distributed Integrated Sensing and Communication (ISAC) systems that minimizes raw data transmission by reconstructing missing or redundant sensor observations at the semantic embedding level. Inspired by self-supervised architectures such as JEPA and its spatio-temporal extension V-JEPA, our framework leverages semantic inpainting to reduce sensing overhead while preserving task-relevant information. We formulate the joint sensing–communication trade-off as a communication resource allocation problem that seeks to maximize overall network throughput while maintaining bounded semantic distortion for sensing tasks. Due to the non-deterministic nature of inpainting-based sensing, the optimization problem is inherently intractable. We therefore propose a suite of lightweight heuristic strategies that make real-time decisions on whether to transmit raw sensor data or rely on inpainted predictions, dynamically allocating communication resources to maximize efficiency. Extensive simulations and proof-of-concept evaluations on real multimodal datasets demonstrate that our approach can significantly increase communication capacity—achieving up to ??? throughput gain—with minimal loss in sensing performance. Our results highlight the potential of embedding-level semantic reconstruction as a control mechanism for optimizing distributed ISAC systems under bandwidth and power constraints.