Hybrid Radar Baseband Simulator

• Hybrid radar baseband simulator is a framework that synthesizes radar signals by integrating physically grounded scene modeling with analytic signal formation.
• It employs techniques such as 3D ray tracing, FMCW/chirp-sequence modeling, and learned mappings to generate range–Doppler–angle data cubes with high fidelity.
• The simulator supports applications in localization, SLAM, and human activity recognition while addressing trade-offs from approximated material properties and hardware nonidealities.

A hybrid radar baseband signal simulator is a simulator that generates synthetic radar baseband data, or closely related products such as range–Doppler–angle cubes, by coupling physically grounded scene or channel modeling with explicit receive-signal formation. In the current literature, this includes frameworks that combine 3D modeling and ray tracing with FMCW or chirp-sequence baseband equations, systems that fuse ray tracing with measured two-way antenna gain and reflector radar cross section, and learned simulators that condition range–azimuth–Doppler generation on waveform-parameterized attributes rather than full hardware descriptions (Liu et al., 2023, Hinderer et al., 19 Sep 2025, Xiao et al., 3 Jun 2025).

1. Definition and uses of the term “hybrid”

In radar simulation, the term hybrid is used in more than one technical sense. In environment-driven simulators, it denotes a division of labor between a scene or propagation model and an analytic signal model. The Blender-based FMCW simulator, for example, uses Blender and the Cycles ray-tracing engine to obtain distance and signal-strength information, then inserts those quantities into beat-signal models of TDM-MIMO FMCW radar (Liu et al., 2023). The indoor localization simulator similarly combines Matlab Antenna Toolbox ray tracing, measured two-way antenna gain, accurate reflector RCS simulation, and a chirp-sequence baseband model (Hinderer et al., 19 Sep 2025).

A second usage appears in data-driven simulation. SA-Radar represents the scene as an environment tensor of reflection points and intensities, but replaces explicit point-spread-function convolution with a learned mapping conditioned on waveform-parameterized attributes $\{\sigma, g, Rs, \lambda\}$ (Xiao et al., 3 Jun 2025). This is hybrid in the sense that the scene representation and conditioning variables remain physically interpretable while the rendering stage is learned.

A third usage appears in joint radar-communications at THz, where hybrid refers to analog/digital beamforming rather than scene synthesis. In the GoSA ultra-massive MIMO setting, the baseband signal is $$\mathbf{x}[m] = \mathbf{F}_\mathrm{RF}\mathbf{F}_\mathrm{BB}[m]\mathbf{s}[m]$$, with a frequency-independent analog precoder and per-subcarrier digital precoders (Elbir et al., 2021). This is a distinct but related strand, because it places waveform, channel, and radar beampattern design directly inside the baseband simulation loop.

2. Architectural patterns

A recurrent architecture is a staged pipeline in which scene construction, propagation, signal synthesis, and task-level processing are separated but coupled. The Blender-based FMCW simulator states four main stages: scenario modeling (3D environment), image rendering & ray-tracing, signal generation (FMCW/TDM-MIMO baseband), and target estimation (range–Doppler–angle) (Liu et al., 2023). In that system, each rendered frame is a 2D image, and each pixel can provide a range estimate, signal strength or intensity, angle of arrival, and time variation across frames for velocity.

The single-channel indoor localization simulator follows a more measurement-calibrated variant. Its inputs include room geometry, reflector positions, radar parameters, robot trajectory, measured $G^2(\theta,\phi)$, and reflector RCS from Feko RL-GO. It then computes multipath rays for room surfaces, direct LOS reflector paths analytically, and assembles a synthetic baseband matrix $\mathbf{X} \in \mathbb{R}^{M \times N}$ together with derived range profiles and range–Doppler maps (Hinderer et al., 19 Sep 2025).

RadaRays uses hardware-accelerated ray tracing as a propagation engine for rotating FMCW radar, including reflection, refraction, absorption, and scattering, but accumulates returned energy into a polar intensity image rather than explicit time-domain I/Q (Mock et al., 2023). RadHARSimulator V2 begins from recorded video, extracts temporally smoothed 3D human poses, and then synthesizes raw complex IF/baseband, range-time maps, Doppler-time maps, and ridge features (Gao, 12 Nov 2025). SimHumalator instead uses IEEE 802.11g waveform generation and marker-based motion capture to produce passive-radar micro-Doppler signatures (Vishwakarma et al., 2021).

Framework	Hybrid components	Output representation
Blender-based FMCW simulator	Blender + Cycles ray tracing + analytic FMCW/TDM-MIMO	Complex beat signals and data cubes
Indoor localization simulator	SBR ray tracing + measured $G^2(\theta,\phi)$ + Feko RL-GO RCS + chirp-sequence model	Real baseband $\mathbf{X}$, range profiles, range–Doppler maps
RadaRays	Hardware-accelerated ray tracing + Snell/Fresnel/BRDF + noise	Polar intensity image
RadHARSimulator V2	Video/CV poses + propagation models + IF synthesis	Raw complex IF, RTM, DTM
SA-Radar	Environment tensor + waveform-parameterized attribute embedding + ICFAR-Net	RAD tensor

3. Signal-level formulations

The analytic core of many hybrid simulators is a waveform-specific baseband model. For FMCW radar, the transmitted chirp is written as

$$s_{\text{tx}}(t) = \exp\left(j2\pi\left(f_c t + \frac{\mu}{2} t^2\right)\right), \quad 0 \le t < T_b,$$

with chirp slope $\mu = B/T_b$. In the Blender-based TDM-MIMO simulator, the beat signal at receive antenna $n$ is

$$\mathbf{h}_{T,n}(n_s; l) = \sum_{m=1}^{M} \sigma_{m,n,l}\, \exp\left(j2\pi\left( \frac{2 \mu R_{m,n}}{c} \frac{n_s - 1}{F_s} - f_D (l-1) T_b \right)\right),$$

and the extended model adds RT-derived range, Doppler, and array-angle phase terms to synthesize $$\mathbf{x}[m] = \mathbf{F}_\mathrm{RF}\mathbf{F}_\mathrm{BB}[m]\mathbf{s}[m]$$0 (Liu et al., 2023).

For chirp-sequence indoor localization with a single real baseband channel, the receive model is expressed as

$$\mathbf{x}[m] = \mathbf{F}_\mathrm{RF}\mathbf{F}_\mathrm{BB}[m]\mathbf{s}[m]$$1

where $$\mathbf{x}[m] = \mathbf{F}_\mathrm{RF}\mathbf{F}_\mathrm{BB}[m]\mathbf{s}[m]$$2, $$\mathbf{x}[m] = \mathbf{F}_\mathrm{RF}\mathbf{F}_\mathrm{BB}[m]\mathbf{s}[m]$$3, $$\mathbf{x}[m] = \mathbf{F}_\mathrm{RF}\mathbf{F}_\mathrm{BB}[m]\mathbf{s}[m]$$4, and $$\mathbf{x}[m] = \mathbf{F}_\mathrm{RF}\mathbf{F}_\mathrm{BB}[m]\mathbf{s}[m]$$5 are obtained either from LOS reflector modeling via the radar equation or from ray-traced multipath (Hinderer et al., 19 Sep 2025).

Passive-radar simulators use the same superposition principle with different illuminators. SimHumalator models the surveillance signal as a coherent sum over body-part scatterers:

$$\mathbf{x}[m] = \mathbf{F}_\mathrm{RF}\mathbf{F}_\mathrm{BB}[m]\mathbf{s}[m]$$6

where $$\mathbf{x}[m] = \mathbf{F}_\mathrm{RF}\mathbf{F}_\mathrm{BB}[m]\mathbf{s}[m]$$7 depends on primitive-body RCS, $$\mathbf{x}[m] = \mathbf{F}_\mathrm{RF}\mathbf{F}_\mathrm{BB}[m]\mathbf{s}[m]$$8 is the IEEE 802.11g transmit waveform, and $$\mathbf{x}[m] = \mathbf{F}_\mathrm{RF}\mathbf{F}_\mathrm{BB}[m]\mathbf{s}[m]$$9 is the body-part Doppler (Vishwakarma et al., 2021).

Sub-Nyquist OFDM radar introduces a different hybridization between hardware constraints and signal processing. The folded signal after sampling at $G^2(\theta,\phi)$0 is

$G^2(\theta,\phi)$1

and per-sub-band unfolding uses

$G^2(\theta,\phi)$2

Stacking $G^2(\theta,\phi)$3 reconstructs a full-band representation $G^2(\theta,\phi)$4, where $G^2(\theta,\phi)$5 is symbol-mismatch noise and $G^2(\theta,\phi)$6 is folded noise (Han et al., 2023).

Learned RAD simulators operate one stage later in the signal chain. SA-Radar models the cube as

$G^2(\theta,\phi)$7

then replaces direct PSF accumulation with a conditioned 3D U-Net that maps the environment tensor and attribute embedding to a simulated RAD tensor (Xiao et al., 3 Jun 2025).

4. Scene, propagation, and target modeling

The scene side of a hybrid simulator determines whether the baseband product is merely kinematically correct or also propagation-aware. In the Blender-based FMCW framework, geometry, material properties, and motion are defined in Blender; Cycles provides range, intensity, and AoA; and each pixel is treated as a single scatterer or path. The paper explicitly notes that occlusion and specular reflections are handled implicitly by ray tracing, while a full multipath superposition model is not explicitly built (Liu et al., 2023).

The indoor localization simulator uses a more explicit path-based decomposition. Matlab Antenna Toolbox’s SBR method provides path loss $G^2(\theta,\phi)$8, propagation distance $G^2(\theta,\phi)$9, AoA $\mathbf{X} \in \mathbb{R}^{M \times N}$0, and phase $\mathbf{X} \in \mathbb{R}^{M \times N}$1. For passive reflectors, the simulator bypasses the generic room ray tracer and instead uses the radar equation with measured two-way gain and RL-GO RCS. Octahedral corner reflectors are modeled as electrically large PEC objects, with scalar RCS formed from $\mathbf{X} \in \mathbb{R}^{M \times N}$2 and $\mathbf{X} \in \mathbb{R}^{M \times N}$3, while cross-polar RCS is treated as negligible (Hinderer et al., 19 Sep 2025).

RadaRays extends propagation realism by incorporating Snell’s law, Fresnel energy splitting, and a BRDF-like reflection model. Each ray carries an energy fraction of the transmitted signal, and material-dependent wave speed enters the effective range-bin calculation through path segments in different media. This is a physically grounded propagation layer, but its direct output remains a range-compressed polar image rather than raw chirp-domain baseband (Mock et al., 2023).

Human-centered simulators introduce yet another target model. RadHARSimulator V2 uses object detection, global nearest neighbor tracking, HRNet 2D pose estimation, nearest-match 3D pose reconstruction, Kalman filtering, interpolation, and Savitzky–Golay smoothing to obtain joint trajectories in meters, then treats joints as moving scatterers in free-space and through-the-wall scenarios (Gao, 12 Nov 2025). SimHumalator derives 3D joint trajectories from a PhaseSpace motion-capture system and attaches primitive scattering shapes—ellipsoids for limbs and torso, a sphere for the head—with angle-dependent RCS and body-part Doppler (Vishwakarma et al., 2021).

A common misconception is that realistic geometry alone implies baseband fidelity. The literature distinguishes these layers sharply. RadaRays explicitly states that it does not explicitly simulate the time-domain baseband I/Q waveform of FMCW chirps, whereas RadHARSimulator V2 explicitly outputs a synthetic radar baseband signal $\mathbf{X} \in \mathbb{R}^{M \times N}$4, and the Blender-based simulator explicitly constructs beat signals from RT outputs (Mock et al., 2023, Gao, 12 Nov 2025, Liu et al., 2023).

5. Processing outputs, validation, and applications

Hybrid baseband simulators typically produce intermediate and final products at several levels of the radar chain. In the Blender-based FMCW system, simulated beat signals are organized into $\mathbf{X} \in \mathbb{R}^{M \times N}$5 data cubes, followed by range FFT, Doppler FFT, and angle processing. In its validation scenario, a human walks on a circular track of radius 3 m for 100 frames at 24 Hz, giving $\mathbf{X} \in \mathbb{R}^{M \times N}$6, and comparison between estimated and ground-truth velocity yields $\mathbf{X} \in \mathbb{R}^{M \times N}$7 (Liu et al., 2023).

The indoor localization simulator outputs synthetic baseband, range profiles, and range–Doppler maps, then uses a particle filter for localization against ceiling-mounted local reference points. In the reported L-shaped room scenario, consecutive range–Doppler maps at 100 ms intervals show high transient consistency, and trajectory reconstruction reaches mean RMSE $\mathbf{X} \in \mathbb{R}^{M \times N}$8 cm (Hinderer et al., 19 Sep 2025).

RadaRays validates its image-level output against real rotating radar data using Mutual Information Score and Structural Similarity Index. The reported average gains are MIS $\mathbf{X} \in \mathbb{R}^{M \times N}$9 higher and SSI $G^2(\theta,\phi)$0 higher than a lidar-like baseline, with failure rate $G^2(\theta,\phi)$1 of $G^2(\theta,\phi)$2 on DCC and $G^2(\theta,\phi)$3 on KAIST, and average RTE around 5–6 cm/m (Mock et al., 2023). These results concern polar intensity images rather than raw baseband, but they define a practical reference point for scene realism.

RadHARSimulator V2 starts from raw complex IF/baseband, generates RTM by FFT over fast time, applies MTI and DnCNN, then computes DTM by STFT and extracts ridge features by a maximum local energy method. Reported PSNR values place RTM often in the 40–45 dB range, while processed RTM and DTM are typically in the 17–32 dB range; in the RWSet evaluation, SPNet is the only network with validation accuracy $G^2(\theta,\phi)$4 (Gao, 12 Nov 2025).

SimHumalator uses cross ambiguity function processing between WiFi reference and surveillance channels to obtain range–Doppler maps and Doppler-time spectrograms. The paper reports simulation-based human activity recognition accuracy of $G^2(\theta,\phi)$5 and measured-data accuracy of $G^2(\theta,\phi)$6, with ResNet18 reaching 92.6% on the real measurement set (Vishwakarma et al., 2021). SA-Radar, although RAD-level rather than raw-baseband, shows that simulated cubes can improve 2D and 3D object detection and radar semantic segmentation when used standalone or in combination with real data, with runtime around 0.036–0.037 s per cube (Xiao et al., 3 Jun 2025).

The application space therefore spans AI/ML dataset generation, localization, SLAM, human activity recognition, detection, tracking, and radar semantic segmentation (Hinderer et al., 19 Sep 2025, Mock et al., 2023, Gao, 12 Nov 2025, Xiao et al., 3 Jun 2025).

6. Limitations, misconceptions, and extensions

The principal limitation of many hybrid simulators is that one layer of fidelity is often substituted for another. The Blender-based simulator uses Blender materials as stand-ins for radar scattering properties, but Blender materials are designed for optics, not RF, so the mapping to radar RCS is approximate. The same work also notes that small-scale fading, hardware impairments, phase noise, quantization, and richer multipath are not added, even though they could be overlaid at the signal stage (Liu et al., 2023).

The indoor localization simulator makes a different set of simplifications. The SBR ray tracer omits refraction, corner diffraction, diffuse scattering, and GPU acceleration in the referenced version, and hardware nonidealities such as phase noise, IQ imbalance, and non-linearities are omitted for simulation simplicity. Diffuse scattering is only emulated later by a $G^2(\theta,\phi)$7 Gaussian blur filter over the range–Doppler map (Hinderer et al., 19 Sep 2025).

A second recurring misconception is that all radar simulators that return range–azimuth structure are baseband simulators. RadaRays explicitly does not simulate explicit time-series baseband I/Q, Doppler and moving targets in the frequency domain, or phase coherency across chirps; its physical layer can be coupled to a more detailed FMCW baseband model, but it is not that model by itself (Mock et al., 2023). SA-Radar has an analogous boundary: it works directly at RAD level, and its coarse attribute set $G^2(\theta,\phi)$8 is acknowledged as insufficient to replace real data fully in fine-grained tasks such as 3D object detection (Xiao et al., 3 Jun 2025).

Human-centered simulators reveal a third limitation: motion realism does not automatically imply sensor realism. RadHARSimulator V2 omits phase noise, IQ imbalance, and RCS fluctuation, while SimHumalator notes that measured spectrograms contain multipath, shadowing, and path-loss effects not explicitly modeled in the simulator (Gao, 12 Nov 2025, Vishwakarma et al., 2021).

A plausible implication is that future hybrid radar baseband simulators will increasingly combine calibrated propagation, waveform-accurate baseband synthesis, and learned rendering. The literature already points in that direction: SA-Radar proposes differentiable extensions toward raw I/Q simulation, and THz joint radar-communications work shows how hybrid precoding, beam split correction, and wideband channel modeling can be integrated directly into the baseband formulation $G^2(\theta,\phi)$9 (Xiao et al., 3 Jun 2025, Elbir et al., 2021). The resulting trajectory is not toward a single canonical simulator, but toward modular systems in which scene modeling, propagation, baseband generation, and downstream task evaluation remain separable and therefore calibratable.