An AI framework now handles the fiddly job of turning simulated wireless data into models that work in the real world.
Researchers describe AIMS, an agentic AI system for multi-modal integrated sensing and communication, or ISAC - tech that lets wireless networks sense their surroundings while staying connected. Give it a plain-language deployment request (the task, the conditions, how much real data you can afford to collect) and AIMS works out a sim-to-real configuration and runs it. Two agents split the job: one builds synchronized sensing and wireless records from shared physical scenes, the other picks the right data types and sets up mixture-of-experts learning for zero-shot or few-shot use. On the real-world DeepSense 6G dataset, the system beat baseline simulation and fusion methods at vehicle detection and beam prediction, and a separate benchmark found it planned and replanned deployments more accurately when given structured domain knowledge and feedback.
That matters because ISAC is quietly becoming load-bearing infrastructure for 6G, self-driving cars, and smart infrastructure, and the models behind it are usually starved for annotated real-world data. Simulation was supposed to fix that, but configuring scene, sensing, wireless, and learning settings for each new deployment has mostly been manual, fragile work. Automating that configuration step, rather than just generating more synthetic data, is the actual contribution here.
Still, this is one dataset and one lab's benchmark, not a field deployment - the gap between "beat our baselines" and "works on your rooftop sensor array" is exactly the gap this paper is trying to close.