mmRadarTwin: A Measurement-Calibrated Signal-Level Digital Twin Platform for Indoor mmWave Radar

Indoor mmWave radar perception is notoriously difficult to reproduce because radar responses are highly sensitive to multipath propagation, material properties, and specific hardware configurations. Existing digital twin tools often focus on communication-oriented metrics (like channel impulse responses) rather than the complex signal-level products required for radar sensing. This paper asks: how can we build a reusable, signal-level digital twin that allows researchers to compare simulated radar data directly with real-world measurements to diagnose and calibrate sensing models?

To address this, the authors developed mmRadarTwin, a platform that links a real-world radar measurement branch with an Unreal Engine 5 (UE5) scene simulation branch. The core innovation is a shared signal-processing interface: both the real radar data and the simulated data are processed through the same fast-time and spatial FFT chain, terminating in an identical range-angle (RA) representation.

Crucially, the simulator does not just output a final heatmap; it exports per-path contribution records. These records identify the specific actor, material, and propagation event associated with every simulated return. This allows researchers to perform path-attributed residual diagnosis, where discrepancies between the simulation and reality are categorized into actionable labels—such as missing path support, shifted responses, or unmodeled physical mechanisms—rather than simply tuning material parameters until the images look similar.

The authors evaluated mmRadarTwin in an office environment using a commodity monostatic mmWave radar. Across 154 measured poses, the platform successfully recalled 70.8% of the measurement-active, geometry-supported response regions in the central field of view. The system effectively exposed residuals caused by weak path support and missing physical mechanisms, demonstrating that the platform provides a practical, interpretable workflow for diagnosing why a digital twin might deviate from real-world observations.