Source-linked AI summary
SAC-Controlled RIS-Assisted Monopulse Radar for Multipath-Robust Vital-Sign Localization
Jafar Norolahi, Alireza Vahid
TL;DR
Multipath can distort radar direction estimates and produce ghost responses, complicating contactless vital-sign localization. This paper uses SAC to control an RIS-assisted monopulse radar from monopulse observations. At the nominal setting, SAC improves separation MAE over no-RIS and direction-only RIS processing, although full calibration performs better at moderate-to-high SNR and under strong ghost reflections.
Problem
Indoor multipath can create ghost targets and distort direction estimates, while adding spatially separated radar nodes increases hardware, synchronization, and data-fusion demands.
Source: §I · p. 1Method
The paper models direct, RIS-assisted, and ghost returns in a monopulse radar and uses SAC to select continuous RIS phases from monopulse observations.
Source: §I · p. 2Results
At 0.3° separation and 5 dB SNR, SAC separation MAE is 0.402°, versus 1.494° without RIS and 0.629° with direction-only RIS; full calibration performs better at moderate-to-high SNR and under strong ghost reflections.
Source: A. Angular-Separation Accuracy and SNR… · p. 5, §V · p. 6Takeaways & Limitations
The results support SAC-controlled RIS phase adaptation for two-target angular localization in the tested multipath settings, without deployment access to true target angles.
Source: §V · p. 6
Abstract
from arXiv · showhide
Contactless radar sensing offers a promising approach to continuous vital-sign monitoring. However, indoor multipath can distort direction estimates and generate ghost responses. This paper presents a reconfigurable intelligent surface (RIS)-assisted monopulse radar whose RIS phase configuration is controlled by a soft actor-critic agent to improve angular localization. The proposed framework combines simultaneous sum and difference beams with a programmable reflected path, enabling a single radar node to adapt the propagation environment without relying on multiple synchronized sensors. A compact multipath model formulates RIS control as a continuous-action reinforcement-learning problem whose reward penalizes bearing errors and misleading multipath reinforcement. Numerical results compare the proposed method with conventional monopulse processing without an RIS, a frequency-modulated continuous-wave range-angle fast Fourier transform baseline, random RIS phase configurations, and deterministic RIS controllers. The evaluation considers different target separations, signal-to-noise ratio levels, and multipath conditions.
I. Introduction
Indoor multipath can create ghost targets and bias radar direction estimates, while additional radar nodes increase hardware and synchronization demands. The paper proposes adapting an RIS with SAC to improve monopulse bearing localization in multipath environments.
I. Introduction
- Multipath reflections can create ghost targets and distort direction estimates, while multiple radar nodes add hardware, synchronization, and data-fusion requirements.Blockage, unfavorable target orientation, and weak propagation paths can also create blind regions.Source: §I · p. 1
- Monopulse radar estimates bearing from simultaneous sum- and difference-channel ratios, but multipath phases can bias that estimate, especially when the direct reflection is weak.This approach avoids mechanically scanning the surveillance region.Source: §I · p. 1
- An RIS adds a controllable reflected path, but its phase space grows with the number of elements and useful settings vary with target position and reflectors.Steering from a nominal target direction may be insufficient in dynamic multipath.Source: §I · p. 1
- The paper develops a compact model incorporating direct, RIS-assisted, and ghost-path returns in the sum and difference channels.Its stated scope is bearing estimation and RIS phase adaptation, not complete physiological-signal reconstruction.Source: §I · p. 2
- SAC determines continuous RIS phase shifts from monopulse observations, with a reward that penalizes bearing error and misleading multipath reinforcement.The policy can map observations directly to RIS settings after training.Source: §I · p. 2
- The controller’s bearing-offset performance is characterized for two closely spaced targets across azimuth separations from 0.1° to 1°.The evaluation quantifies absolute angular error.Source: §I · p. 2
II. System Model
The system model describes a monostatic monopulse radar and passive RIS observing multiple subjects through direct, RIS-assisted, and wall-induced ghost paths. The array snapshot combines these returns with noise.
II. System Model
- The modeled system is a monostatic monopulse radar with an M-element planar array and a passive RIS containing N programmable reflecting elements.It tracks K subjects in an environment with a direct path, an RIS-assisted path, and G wall-induced ghost paths.Source: §II · p. 2
- In the simulations, M = N = 40 × 40 = 1600 and K = 2; range and physiological-frequency estimation are reserved for the extended journal version.The modeled targets are subjects in an indoor environment.Source: §II · p. 2
- The target manifold matrix A stacks the array steering vectors for the K targets, with element responses defined using radar-element positions and wavelength.The wavelength is λ = c/fc, where c is the speed of light and fc is carrier frequency.Source: §II · p. 2
- Each target’s direct-path coefficient combines blocked-path attenuation and target reflectivity with the corresponding propagation phase, treated as constant over one coherent dwell.The target coefficients are stacked into a vector.Source: §II · p. 2
- The post-dwell array snapshot combines target-manifold returns, ghost-path returns, RIS-assisted returns, and noise.The RIS-assisted return is formed through the effective RIS-to-radar channel, the target-to-RIS spatial response, and effective two-way path coefficients.Source: §II · pp. 2–3
A. Monopulse Bearing Estimation
The bearing estimator forms sum and azimuth-difference outputs around a prior azimuth, then applies a locally calibrated correction. The study estimates azimuth only; elevation is outside its scope.
A. Monopulse Bearing Estimation
- The radar maintains a prior azimuth for each target, while elevation is fixed at a known value in the simulations.The normalized sum-channel weight is formed around that prior direction.Source: A. Monopulse Bearing Estimation · p. 3
- The azimuth-difference weight removes the component parallel to the sum steering vector, and the estimator computes sum and difference outputs.A locally calibrated monopulse slope supports the correction for sufficiently small angular offsets.Source: A. Monopulse Bearing Estimation · p. 3
- The resulting bearing estimate applies an azimuth correction to the prior, while RIS phases are selected to strengthen the target and reduce ghost-induced monopulse bias.Only azimuth is estimated; elevation estimation is outside the study’s scope.Source: A. Monopulse Bearing Estimation · p. 3
III. SAC-Based RIS Phase Control
The controller uses a one-step, entropy-regularized SAC policy to map radar observations to RIS phase configurations. Its training reward balances target-bearing and inter-target separation errors, while deployment uses a frozen actor without true target angles or online updates.
III. SAC-Based RIS Phase Control
- One calibration state produces a RIS configuration evaluated over one radar dwell in the implemented one-step, entropy-regularized SAC controller.SAC maximizes reward plus an entropy term.Source: §III · pp. 3–4
- The state combines noisy calibration estimates, bounded monopulse-ratio components, and normalized power, separation, SNR, and ghost-offset features, without true target angles.The calibration estimates include target-enhancement and ghost-control phases, ghost-control strength, and ghost angular offset.Source: §III · p. 3
- The structured action encodes target-enhancement and ghost-control phases plus ghost-control strength, then expands into an RIS phase mask with unit-modulus elements.The first two action pairs encode the phases; the ghost-control strength determines how many columns combine coherently.Source: §III · p. 4
- Training uses target-angle RMSE and inter-target separation error, with λsep setting the relative weight of the squared separation error.The simulator supplies true target azimuths during offline training; λsep = 1.5 emphasizes separation error.Source: §III · p. 4
- After offline convergence, the frozen actor generates the next-dwell RIS mask without true target angles, channel labels, or online gradient updates.The deployed controller forms its state from radar measurements and applies the actor's structured action.Source: §III · p. 4
IV. Numerical Results
The numerical evaluation uses matched Monte Carlo trials to compare SAC with monopulse, FMCW, and RIS-controller baselines. Fig. 2 tracks estimated against true target separation at 5 dB SNR; the FMCW curve remains at zero because the pair is unresolved in the tested interval.
IV. Numerical Results
- Each operating point averages 100 matched Monte Carlo trials, using identical scenes, calibration errors, and noise samples across compared methods.The simulations use two targets and a 28 GHz monostatic monopulse radar with 40×40 receive-array and RIS elements.Source: §IV · p. 4
- The comparison includes no-RIS monopulse, FMCW range-angle FFT, random-phase and deterministic RIS controllers, and a nondeployable exact-channel reference.Performance is assessed by mean absolute error of estimated inter-target azimuth separation; paired two-sided t-tests compare SAC with full calibration.Source: §IV · pp. 4–5
- At 5 dB SNR, Fig. 2 plots mean estimated azimuth separation against true separation; the FMCW estimate stays at zero because the pair is unresolved throughout the tested interval.The estimates average 100 matched Monte Carlo trials, and the diagonal marks ideal estimation.Source: Figure 2 · p. 5
A. Angular-Separation Accuracy and SNR Robustness
At the nominal 0.3° separation and 5 dB SNR, SAC improves separation accuracy over no-RIS monopulse and direction-only RIS, but full calibration performs better. Across separations and SNR levels, the comparison favors full calibration at the smallest separations and higher SNRs.
A. Angular-Separation Accuracy and SNR Robustness
- 0.402° SAC separation MAE versus 1.494° without RIS and 0.338° with full calibration at 0.3° and 5 dB; full calibration is 18.9% better (p = 4.41 × 10−2).SAC reduces MAE by 73.1% versus no-RIS monopulse and 36.1% versus direction-only RIS control.Source: A. Angular-Separation Accuracy and SNR… · p. 5
- Across 0.1°–1°, SAC estimates rise from 0.433° to 1.103° and full-calibration estimates from 0.351° to 1.067°; both approach the ideal relation as separation increases.At 0.3°, their mean estimates are 0.542° and 0.603°, respectively.Source: A. Angular-Separation Accuracy and SNR… · p. 5, Figure 3 · p. 5
- Fig. 3's FMCW estimate remains at 0° throughout the tested interval because the target pair is unresolved.Source: Figure 3 · p. 5
- Full calibration is significantly better than SAC at 0.1°, 0.2°, and 0.3° separation; the methods are statistically comparable at the remaining tested separations.Source: A. Angular-Separation Accuracy and SNR… · p. 5
- From −10 to 15 dB at 0.3° separation, SAC error falls from 1.258° to 0.310°, while full-calibration error falls from 1.335° to 0.212°; full calibration is significantly better at 5, 10, and 15 dB.The no-RIS error remains between 1.475° and 1.681°; SAC and full calibration are comparable at −10, −5, and 0 dB.Source: A. Angular-Separation Accuracy and SNR… · p. 5, Figure 4 · p. 5
B. Ghost-Strength Robustness
As the ghost-to-target amplitude ratio increases, SAC error rises more than full-calibration error, and full calibration performs significantly better at the stronger tested ghost levels.
B. Ghost-Strength Robustness
- As the ghost-to-target ratio rises from −15 to −3 dB, SAC error increases from 0.293° to 0.667°, versus 0.261° to 0.409° for full calibration; full calibration is significantly better at −9, −6, and −3 dB.The methods are statistically comparable at −15 and −12 dB; SAC does not provide uniform superiority as coherent ghosts strengthen.Source: B. Ghost-Strength Robustness · p. 6
V. Conclusion
The paper presents an SAC-controlled RIS-assisted monopulse radar for localizing two closely spaced vital-sign targets in multipath. At 0.3° separation and 5 dB SNR, SAC lowers separation MAE versus no RIS and direction-only RIS steering, though full calibration performs better in some conditions.
V. Conclusion
- At 0.3° separation and 5 dB SNR, separation MAE is 0.402° with SAC, versus 1.494° without RIS and 0.629° with direction-only RIS steering.The considered FMCW baseline does not resolve the target pair over the tested separation interval.Source: §V · p. 6
- The framework uses a 40 × 40 receive array and 40×40 RIS, with an offline-trained SAC policy deployed without access to true target angles.Source: §V · p. 6
- Full calibration performs better than SAC at moderate-to-high SNR and under strong ghost reflections.Future work includes broader multipath conditions, online policy refinement, and experimental validation.Source: §V · p. 6