DyRAD

Radar Novel View Synthesis for Dynamic Driving Scenes

Merav Keidar1 · Tomer Borreda1 · Rajalakshmi Nandakumar2 · Or Litany1,3
1Technion · 2Cornell Tech · 3NVIDIA
arXiv · 2026
DyRAD re-simulation: from a measured radar frame, DyRAD renders a laterally shifted sensor, a repositioned object and a higher-resolution sensor, and scores higher RAD PSNR and detection hit rate than RadarSplat and RadarFields. Ground Truth Range [m] Azimuth° Doppler Repositioned Object Range [m] Azimuth° Doppler Lateral Sensor Shift Range [m] Azimuth° Doppler High Res Sensor Range [m] Azimuth° Doppler 0.00.20.40.60.81.02324252627 Detection (Hit Rate) → RAD PSNR [dB] → DyRAD (ours) RadarSplat RadarFields on-path off-path, +2 m

From recorded radar, poses and object boxes, DyRAD reconstructs the scene and renders it from new viewpoints, with edited objects and for other sensors.

DyRAD reconstructs a dynamic driving scene from recorded radar and renders full range–azimuth–Doppler tensors from poses, sensors and scene layouts it never recorded.

Abstract

Reconstructing dynamic driving scenes from recorded sensor data supports closed-loop evaluation of autonomous driving systems by synthesizing observations beyond the original trajectory. Unlike cameras and LiDAR, radar measures radial velocity directly through Doppler. Yet existing radar novel-view synthesis fails to exploit this capability: methods addressing dynamic scenes reconstruct only range–azimuth tensors, while methods that render Doppler assume static scenes. Moreover, because radar processing spreads each reflection across multiple bins, existing representations absorb this spread into scene geometry, causing it to render incorrectly when the viewpoint moves.

We present DyRAD, which models dynamic driving scenes using static background reflectors and motion-tracked dynamic point reflectors to render complete range–azimuth–Doppler (RAD) tensors. Reflector velocities are derived from object tracks and projected onto the line of sight, making Doppler both a rendered output and supervision for those tracks. Crucially, we render reflectors through a fixed analytic point-spread function (PSF) derived from the radar’s signal-processing chain, preventing sensor-induced spread from being baked into the scene representation. Beyond improving scene reconstruction, this separation also enables zero-shot sensor-configuration transfer, allowing the same reconstructed scene to be rendered under different radar specifications without refitting.

We evaluate DyRAD on RADIal, Boreas, and a synthetic benchmark across both on-path poses and displaced viewpoints untested by prior work. On RADIal, DyRAD recovers radar detections in 90.7% of reference-detected objects, compared with 26.9% for the strongest baseline.

Method

paper §3 · Figs. 2–3
DyRAD overview. Inputs: recorded RAD tensors, sensor poses, object bounding boxes and the radar processing chain. Initialization seeds point reflectors and object tracks and fixes the sensor PSF. Optimization fits reflectors and tracks through the differentiable RAD renderer, with each object moving rigidly along its learned track. Inference renders RAD tensors at new times and poses. Radar Tensors {Yt, Pt}Tt=1 Object Bounding Boxes Radar Processing Chain Doppler / DDMA → Range FFT→ Azimuth beamforming SceneInitialization Fixed SensorModel Learned Scene RAD Renderer Radar Predictions {Ŷt, Pt}Tt=1 ℒ = ℒrec + λint ℒint Novel view Ŷt*, Pt* Data flow Gradient flow

Inputs initialize the scene, the differentiable RAD renderer fits it to the recordings, and the fitted scene renders new views. Solid grey arrows show data flow; dashed red arrows show gradient flow.

  1. InputRecorded RAD tensors, sensor poses, object bounding boxes, and the radar's processing chain.
  2. Scene initializationPoint reflectors are seeded from the measurements, the boxes split them into background and objects and initialize each object's track, and the processing chain fixes the sensor PSF.
  3. OptimizationThe differentiable RAD renderer lets reflectors and tracks be fit jointly to the recorded measurements, with 𝓛 = 𝓛rec + λint𝓛int.
  4. Novel viewsThe optimized scene renders RAD tensors at new times, sensor poses and configurations.

Novel views

paper §4.2 · supplementary videos

Ground truthDyRAD DyRAD-staticRadarSplat* RadarFields*

Every clip shows all methods on the same scene at the same instant, on one brightness scale. ρ is the Pearson correlation with the ground truth for that frame. White boxes mark annotated vehicles. DyRAD-static keeps our point-reflector representation and renderer but treats every reflector as static and does not use the bounding boxes to separate objects during initialization. Comparing it with the full model measures the combined benefit of object separation and motion modelling. * RadarSplat and RadarFields predict only RA measurements; their Doppler is assigned from a static ego-motion model.

Quantitative results

paper Tables 1–2
RADIal · RAD reconstructionRADIal · detectionBoreas · RA reconstruction
MethodPSNR ↑ρ ↑ρobj ↑CD [m] ↓F1 ↑Hit rate ↑PSNR ↑ρ ↑ρobj ↑

DyRAD-static keeps our point-reflector representation and renderer but treats every reflector as static and does not use the bounding boxes to separate objects during initialization. Comparing it with the full model measures the combined benefit of object separation and motion modelling.

Sensor-configuration transfer

paper §4.3 · Fig. 10 · Table 16
Measured native RA projection, the coarse-trained scene rendered directly at native resolution, and linear upsampling of the coarse render, with enlarged insets.
Because the PSF is fixed and kept out of the scene, a scene fit at a coarse configuration renders directly at a finer one without refitting. Left: measured native RA. Middle: the coarse-trained scene rendered with the native PSF and grid. Right: linear upsampling of its coarse render. Direct rendering keeps separate vehicle returns that upsampling blurs together: detection hit rate 0.960 against 0.906 and Chamfer distance 3.20 m against 7.27 m, while upsampling scores higher on full-frame correlation (paper Table 16).

Acknowledgments

Or Litany acknowledges support from the Israel Science Foundation (grant 624/25) and the Azrieli Foundation Early Career Faculty Fellowship. The authors gratefully acknowledge this support. This research was supported by the Council for Higher Education in Israel under the Moonshot Project.

BibTeX

@article{keidar2026dyrad,
  title         = {DyRAD: Radar Novel View Synthesis for Dynamic Driving Scenes},
  author        = {Keidar, Merav and Borreda, Tomer and Nandakumar, Rajalakshmi and Litany, Or},
  journal       = {arXiv preprint arXiv:2609.39841},
  year          = {2026},
  eprint        = {2609.39841},
  archivePrefix = {arXiv},
  url           = {https://arxiv.org/abs/2609.39841}
}