AIMS coordinates simulation and learning for sensing-aware wireless networks

Training a wireless system to recognize vehicles or select a useful radio beam requires observations that describe the same physical scene. Camera or positional data cannot compensate for radio measurements that correspond to a different environment. Gathering aligned, labeled examples for each deployment is expensive.
In the AIMS preprint, Yijie Bian and colleagues at the Hong Kong University of Science and Technology propose two agents that coordinate simulation with task learning. Their objective is to adapt a simulation pipeline to the deployment conditions and the amount of real-world training data available.

A deployment request becomes an executable configuration

AIMS takes a natural-language request describing the target task, deployment conditions and real-data budget. The system derives a deployment-specific sim-to-real configuration and coordinates its execution. Integrated sensing and communication, or ISAC, combines environmental perception with wireless connectivity; the paper focuses on learning relationships between sensing observations and wireless measurements.
The authors explain why generating more synthetic examples alone is insufficient. Scene geometry, sensor settings and wireless configurations must agree. A learning model trained on mismatched records can inherit those errors even when the dataset is large. A roadside unit communicating with vehicles gives the paper a concrete setting: locating a vehicle in a scene and predicting a beam label require different outputs, while both depend on consistent physical conditions.

Two agents share validated experiment state

A scene construction agent generates synchronized sensing and wireless records from shared physical states. A scene understanding agent chooses task-relevant modalities and configures mixture-of-experts learning. Depending on the real-data budget, the resulting task model supports zero-shot inference or few-shot adaptation.
Structured domain knowledge records available capabilities and their dependencies. Shared experiment state tracks validated outputs and their provenance, so a changed deployment condition need not trigger a complete rebuild. The agents can decide which outputs remain usable and which affected steps require regeneration. Validation feedback supports revision of earlier decisions.
This design keeps simulation and numerical learning in the workflow. The language-driven agents coordinate those operations; the proposal does not replace physical models with a language model's description of a scene. That distinction matters for reproducing the method: an executable plan still needs compatible simulators, records and training configurations.

Task accuracy and orchestration need separate tests

The authors report improved vehicle detection and beam prediction on DeepSense 6G compared with the simulation and fusion baselines they considered. They also evaluate orchestration across 140 normal and challenging natural-language deployment requests, examining task interpretation, dependency handling, reuse of outputs and feedback-driven replanning.
Those evaluations answer different questions. A correct plan does not establish that its trained model will transfer to every roadside environment, and an accuracy gain does not prove the planner handles every changed dependency. The reported results support further testing of coordinated sim-to-real pipelines. They do not establish a production-ready autonomous 6G deployment or remove the need to validate sensing and radio alignment at a new site.