AWDT cuts ray-tracing RSRP prediction MAE from 11.46 dB to 5.25 dB using smartphone measurements

AWDT organizes wireless digital twin construction and calibration into three specialized agents: EnvAgent builds the propagation environment, OpAgent infers BS and sector configurations, and MatAgent calibrates radio material properties. Whereas building a high-fidelity wireless digital twin typically requires substantial manual effort to integrate heterogeneous information and infer unknown propagation-related parameters, this work moves that process to end-to-end autonomous agent execution. At the abstract level, the framework composition and each agent's role are described as a method design; no ablation or per-agent performance data are provided.

The three agents iteratively refine the wireless digital twin using discrepancies between LTE/NR RSRP measurements and ray-tracing predictions. Measured-versus-simulated residuals serve as a closed-loop feedback signal driving twin calibration, rather than one-off manual parameter setting. The abstract states the iterative refinement mechanism and the measurement type used, but gives no iteration counts, convergence criteria, or data scale.

In real-world experiments, AWDT reduces RSRP prediction MAE from 11.46 dB to 5.25 dB. The improvement in ray-tracing fidelity is quantified with a concrete error metric. The abstract reports before-and-after MAE values from real-world experiments, but does not state the number of measurement points, number of scenarios, or statistical uncertainty.

Evaluation with an independent measurement system shows cross-device transferability with lightweight device-specific bias adaptation. This indicates the framework is not tied to the device used during construction and can migrate to other measurement devices through lightweight adaptation. The abstract reports the independent-measurement-system evaluation and lightweight bias adaptation as a conclusion, without transfer error values or the amount of adaptation data required.

The work targets the construction and calibration stages of wireless digital twins and applies to settings where environmental information can be obtained and LTE/NR RSRP measurements can be collected, with measurements sourced from commercial smartphones. Its value lies in enabling scalable twin construction and calibration with limited prior knowledge, and in supporting migration to a different measurement device through lightweight device-specific bias adaptation. For network planning, simulation-based evaluation, and AI-native radio access network R&D staff seeking to reduce the cost of manually integrating heterogeneous information and inferring propagation parameters, this framework offers a reusable process paradigm.

The abstract does not state the number of measurement points, number of scenarios, or geographic extent of the experiments, nor does it give iteration counts, convergence criteria, or computational cost, so the robustness of the MAE improvement remains to be confirmed in the main text. Cross-device transferability is achieved with lightweight device-specific bias adaptation, but the amount of adaptation data required and the error level after transfer are not quantified in the abstract. In addition, the abstract does not discuss whether framework performance is consistent across frequency bands, propagation environments, or base station configurations, which are directions for follow-up attention. Because this assessment is based on the abstract only, figures and experimental details could not be incorporated, and the applicable scope of the above numbers and conclusions should be judged against the main text.