DyRAD: Radar Novel View Synthesis for Dynamic Driving Scenes

Merav Keidar; Tomer Borreda; Rajalakshmi Nandakumar; Or Litany

cs.CVSep 30, 2026

DyRAD: Radar Novel View Synthesis for Dynamic Driving Scenes

Authors: Merav Keidar, Tomer Borreda, Rajalakshmi Nandakumar, Or Litany

Organizations: Technion · Cornell Tech · NVIDIA

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.

Figures & tables

Figure 1: Radar re-simulation with DyRAD. From recorded radar measurements, poses, and object boxes, DyRAD reconstructs dynamic scenes supporting novel viewpoints , sensor-configuration changes , and scene editing . Right: Higher RAD PSNR and detection hit rate than both baselines on RADIal, on-path and after a $2\,\mathrm{m}$ lateral-shift round trip. See the project page at https://dyrad-nvs.github.io for videos, code, and more results.

Figure 2: DyRAD overview. Inputs: recorded RAD tensors, sensor poses, object bounding-boxes, and radar processing information when available. Initialization: measurements and annotations initialize the scene; processing information defines the fixed sensor response. Optimization: the differentiable RAD renderer jointly fits reflectors and object tracks to recorded measurements. Inference: the optimized scene produces RAD predictions $\hat{\mathbf{Y}}(t^{\star},\mathbf{P}^{\star})$ at query times and sensor poses. Solid gray arrows show data flow ; dashed red arrows show gradient flow .

Figure 3: Object motion along a learned track. Four neighboring track points estimate translational velocity. Reflectors follow the track’s translation and relative rotation, preserving their object-frame coordinates

\bm{\mu}_{i}

Figure 4: On-path reconstruction on RADIal and Boreas. The top two rows show held-out RADIal frames with RA (left) and RD (right) projections; the bottom row shows a held-out Boreas RA frame in Cartesian bird’s-eye view. White boxes mark objects, with selected regions enlarged in the insets. DyRAD preserves object returns that the static variant and baselines weaken or miss, even when surrounding scene structure appears similar.

Appendix figures & tables13 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 5: Point-reflector initialization. Object bounding boxes separate background and object regions. Selected returns are projected into world coordinates and accumulated across training frames. Dynamic returns are aligned by track translation and expressed in object coordinates. Deduplication yields the initial static and dynamic reflectors.

Hyperparameter	Range	RADIal	Boreas
Interpolation weight $\lambda_{\mathrm{int}}$	$10^{-3}$ – $1$	$0.1$	$0.01$
Position LR ( means_lr )	$10^{-5}$ – $10^{-2}$	$0.0002$	$0.00005$
Opacity LR ( opacities_lr )	$10^{-3}$ – $1$	$0.02$	$0.02$
SH LR ( sh_coeffs_lr )	$10^{-4}$ – $10^{-1}$	$0.01$	$0.02$
Motion LR ( deform_lr )	$10^{-4}$ – $10^{-1}$	$0.01$	$0.0005$

Appendix

Table 3: Search ranges and selected hyperparameters.

Method	PSNR [dB] $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	$\rho\uparrow$
DyRAD	$\mathbf{25.78}\,{\scriptstyle\pm 1.09}$	$0.690\,{\scriptstyle\pm 0.025}$	$\mathbf{0.234}\,{\scriptstyle\pm 0.017}$	$\mathbf{0.612}\,{\scriptstyle\pm 0.031}$
DyRAD-static	$25.16\,{\scriptstyle\pm 1.15}$	$0.686\,{\scriptstyle\pm 0.025}$	$0.249\,{\scriptstyle\pm 0.021}$	$0.533\,{\scriptstyle\pm 0.055}$
RadarSplat	$20.94\,{\scriptstyle\pm 0.66}$	$0.050\,{\scriptstyle\pm 0.011}$	$0.533\,{\scriptstyle\pm 0.022}$	$0.291\,{\scriptstyle\pm 0.100}$
RadarFields	$20.38\,{\scriptstyle\pm 0.71}$	$0.059\,{\scriptstyle\pm 0.009}$	$0.569\,{\scriptstyle\pm 0.021}$	$0.055\,{\scriptstyle\pm 0.024}$
Const	$23.71\,{\scriptstyle\pm 1.04}$	$\mathbf{0.723}\,{\scriptstyle\pm 0.023}$	$0.583\,{\scriptstyle\pm 0.025}$	$0.000\,{\scriptstyle\pm 0.000}$
Zero	$20.44\,{\scriptstyle\pm 0.65}$	$0.032\,{\scriptstyle\pm 0.004}$	$0.597\,{\scriptstyle\pm 0.024}$	$0.000\,{\scriptstyle\pm 0.000}$

Appendix

Table 4: Structureless controls. Full-measurement RA reconstruction on held-out RADIal frames. Mean

\pm

sample standard deviation across ten sequences; best means, including controls, are bold. Constant predictions are assigned

\rho=0

by the scorer.

RADIal — RAD reconstruction
	Full				Object
Method	$\rho\uparrow$	PSNR [dB] $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	$\rho\uparrow$	PSNR [dB] $\uparrow$	SSIM $\uparrow$
RadarSplat ∗	$0.068\,{\scriptstyle\pm 0.028}$	$23.56\,{\scriptstyle\pm 0.85}$	$0.215\,{\scriptstyle\pm 0.006}$	$0.634\,{\scriptstyle\pm 0.008}$	$-0.010\,{\scriptstyle\pm 0.057}$	$13.54\,{\scriptstyle\pm 1.22}$	$0.098\,{\scriptstyle\pm 0.021}$
RadarFields ∗	$0.015\,{\scriptstyle\pm 0.009}$	$23.13\,{\scriptstyle\pm 1.23}$	$0.203\,{\scriptstyle\pm 0.012}$	$0.637\,{\scriptstyle\pm 0.007}$	$-0.011\,{\scriptstyle\pm 0.051}$	$13.40\,{\scriptstyle\pm 1.28}$	$0.088\,{\scriptstyle\pm 0.021}$
DyRAD-static	$0.213\,{\scriptstyle\pm 0.054}$	$25.67\,{\scriptstyle\pm 0.95}$	$0.498\,{\scriptstyle\pm 0.021}$	$0.279\,{\scriptstyle\pm 0.026}$	$-0.013\,{\scriptstyle\pm 0.072}$	$13.76\,{\scriptstyle\pm 1.24}$	$0.199\,{\scriptstyle\pm 0.035}$
DyRAD	$\mathbf{0.272}\,{\scriptstyle\pm 0.062}$	$\mathbf{25.88}\,{\scriptstyle\pm 0.95}$	$\mathbf{0.502}\,{\scriptstyle\pm 0.022}$	$\mathbf{0.275}\,{\scriptstyle\pm 0.025}$	$\mathbf{0.658}\,{\scriptstyle\pm 0.070}$	$\mathbf{16.97}\,{\scriptstyle\pm 1.06}$	$\mathbf{0.374}\,{\scriptstyle\pm 0.028}$

Appendix

Table 5: On-path reconstruction: full results. Mean

\pm

sample standard deviation across ten RADIal sequences and three Boreas sequences. Best means within each projection are bold. ∗ denotes the ego-Doppler lift. LPIPS is not reported for RD.

	Entire area
Method	CD [m] $\downarrow$	CD-Full $\downarrow$	IoU@1 m $\uparrow$	Precision $\uparrow$	Recall $\uparrow$	F1 $\uparrow$
RadarSplat ∗	$6.53\,{\scriptstyle\pm 5.82}$	$0.1668\,{\scriptstyle\pm 0.0758}$	$0.141\,{\scriptstyle\pm 0.066}$	$0.084\,{\scriptstyle\pm 0.036}$	$0.087\,{\scriptstyle\pm 0.056}$	$0.073\,{\scriptstyle\pm 0.048}$
RadarFields ∗	$6.97\,{\scriptstyle\pm 0.96}$	$0.1612\,{\scriptstyle\pm 0.0521}$	$0.037\,{\scriptstyle\pm 0.012}$	$0.015\,{\scriptstyle\pm 0.006}$	$0.019\,{\scriptstyle\pm 0.010}$	$0.016\,{\scriptstyle\pm 0.007}$
DyRAD-static	$4.71\,{\scriptstyle\pm 1.29}$	$0.0783\,{\scriptstyle\pm 0.0117}$	$0.188\,{\scriptstyle\pm 0.043}$	$0.169\,{\scriptstyle\pm 0.055}$	$0.099\,{\scriptstyle\pm 0.034}$	$0.123\,{\scriptstyle\pm 0.041}$
DyRAD	$\mathbf{4.01}\,{\scriptstyle\pm 1.02}$	$\mathbf{0.0689}\,{\scriptstyle\pm 0.0092}$	$\mathbf{0.238}\,{\scriptstyle\pm 0.040}$	$\mathbf{0.233}\,{\scriptstyle\pm 0.051}$	$\mathbf{0.149}\,{\scriptstyle\pm 0.037}$	$\mathbf{0.179}\,{\scriptstyle\pm 0.041}$
	Entire area			Foreground

Appendix

Table 6: On-path detection on RADIal: full results. Mean

\pm

sample standard deviation across ten sequences. Best means, including ties, are bold. ∗ denotes the ego-Doppler lift.

Figure 6: Additional on-path examples on RADIal. Each row compares a held-out measurement with the corresponding reconstructions, showing RA (left) and RD (right) projections within each panel. Displayed correlations are computed separately for each projection. White boxes mark objects; insets enlarge selected RA object regions. DyRAD preserves object returns that the static variant and baselines miss, even where the surrounding scene is reconstructed similarly.

Figure 7: On-path examples on Boreas. Held-out RA measurements and reconstructions displayed in Cartesian bird’s-eye view. White boxes mark annotated objects; insets enlarge selected object regions. DyRAD preserves object returns that the static variant and baselines miss, even where the surrounding scene is reconstructed similarly.

Figure 8: Off-path rendering and round-trip reconstruction on RADIal. Each panel shows RA (left) and RD (right) projections. For each method, $M_{0}@T_{1}$ shows measurements synthesized along the trajectory shifted laterally by $2\,\mathrm{m}$ . A new model is fitted to these renders and evaluated at the original poses ( $M_{1}@T_{0}$ ), against the real measurements (GT@ $T_{0}$ ). White boxes mark annotated vehicles. The values in the return panels report RA and RD correlations with the reference, respectively.

Figure 9: Off-path rendering and round-trip reconstruction on Boreas. RA measurements shown in Cartesian bird’s-eye view. For each method, $M_{0}@T_{1}$ shows the initial model rendered along the trajectory shifted laterally by $2\,\mathrm{m}$ . A new model is fitted to these renders and evaluated at the original poses ( $M_{1}@T_{0}$ ), against the real measurements (GT@ $T_{0}$ ). White boxes mark annotated vehicles; $\rho$ denotes correlation of the round-trip reconstruction with the reference.

RADIal — RAD reconstruction
	Full				Object
Method	$\rho\uparrow$	PSNR [dB] $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	$\rho\uparrow$	PSNR [dB] $\uparrow$	SSIM $\uparrow$
RadarSplat ∗	$0.076\,{\scriptstyle\pm 0.032}$	$23.92\,{\scriptstyle\pm 0.85}$	$0.217\,{\scriptstyle\pm 0.006}$	$0.634\,{\scriptstyle\pm 0.008}$	$-0.004\,{\scriptstyle\pm 0.053}$	$13.57\,{\scriptstyle\pm 1.32}$	$0.099\,{\scriptstyle\pm 0.022}$
RadarFields ∗	$0.023\,{\scriptstyle\pm 0.018}$	$24.43\,{\scriptstyle\pm 0.75}$	$0.217\,{\scriptstyle\pm 0.006}$	$0.638\,{\scriptstyle\pm 0.008}$	$-0.011\,{\scriptstyle\pm 0.048}$	$13.48\,{\scriptstyle\pm 1.20}$	$0.099\,{\scriptstyle\pm 0.023}$
DyRAD-static	$0.335\,{\scriptstyle\pm 0.054}$	$26.41\,{\scriptstyle\pm 0.73}$	$0.492\,{\scriptstyle\pm 0.012}$	$0.271\,{\scriptstyle\pm 0.014}$	$0.056\,{\scriptstyle\pm 0.064}$	$13.86\,{\scriptstyle\pm 1.35}$	$0.194\,{\scriptstyle\pm 0.035}$
DyRAD	$\mathbf{0.363}\,{\scriptstyle\pm 0.052}$	$\mathbf{26.53}\,{\scriptstyle\pm 0.74}$	$\mathbf{0.498}\,{\scriptstyle\pm 0.012}$	$\mathbf{0.267}\,{\scriptstyle\pm 0.015}$	$\mathbf{0.536}\,{\scriptstyle\pm 0.077}$	$\mathbf{15.98}\,{\scriptstyle\pm 1.36}$	$\mathbf{0.344}\,{\scriptstyle\pm 0.041}$

Appendix

Table 7: Real-data off-path reconstruction: full results. Results at the original poses after a

2\,\mathrm{m}

lateral-shift round trip. Mean

\pm

sample standard deviation across ten RADIal sequences and three Boreas sequences; best means are bold. ∗ denotes the ego-Doppler lift. LPIPS is not reported for RD.

	Entire area
Method	CD [m] $\downarrow$	CD-Full $\downarrow$	IoU@1 m $\uparrow$	Precision $\uparrow$	Recall $\uparrow$	F1 $\uparrow$
RadarSplat ∗	$6.35\,{\scriptstyle\pm 4.75}$	$0.1699\,{\scriptstyle\pm 0.0702}$	$0.155\,{\scriptstyle\pm 0.076}$	$0.085\,{\scriptstyle\pm 0.042}$	$0.093\,{\scriptstyle\pm 0.061}$	$0.077\,{\scriptstyle\pm 0.050}$
RadarFields ∗	$9.19\,{\scriptstyle\pm 6.07}$	$0.2211\,{\scriptstyle\pm 0.0962}$	$0.072\,{\scriptstyle\pm 0.045}$	$0.198\,{\scriptstyle\pm 0.353}$	$0.033\,{\scriptstyle\pm 0.026}$	$0.027\,{\scriptstyle\pm 0.020}$
DyRAD-static	$3.47\,{\scriptstyle\pm 0.61}$	$0.0681\,{\scriptstyle\pm 0.0101}$	$0.239\,{\scriptstyle\pm 0.036}$	$0.221\,{\scriptstyle\pm 0.051}$	$0.205\,{\scriptstyle\pm 0.036}$	$0.210\,{\scriptstyle\pm 0.042}$
DyRAD	$\mathbf{3.25}\,{\scriptstyle\pm 0.57}$	$\mathbf{0.0629}\,{\scriptstyle\pm 0.0085}$	$\mathbf{0.268}\,{\scriptstyle\pm 0.034}$	$\mathbf{0.251}\,{\scriptstyle\pm 0.049}$	$\mathbf{0.227}\,{\scriptstyle\pm 0.034}$	$\mathbf{0.236}\,{\scriptstyle\pm 0.040}$

Appendix

Table 8: Real-data off-path detection on RADIal: full results. Results at the original poses after a

2\,\mathrm{m}

lateral-shift round trip. Mean

\pm

sample standard deviation across ten sequences; best means are bold. ∗ denotes the ego-Doppler lift.

Method	Training budget	Train [min]	Render [ms/frame]	Peak [MiB]	Rendered output
RadarSplat	10k steps	4.3	4.68	2138	RA map
RadarFields	800 iterations/image	1.4	15.10	4669	RA map
DyRAD	10k steps	7.2	5.65	1055	RAD tensor

Appendix

Table 11: Computational cost on RADIal. Median training time, rendering latency, and peak allocated GPU memory on an RTX 4090.

Figure 10: Changing the measurement configuration of a reconstructed scene. Left: measured native RA projection. Middle: the coarse-trained scene rendered with the native PSF and sampling grid. Right: linear upsampling of the same scene’s coarse render. The bottom row enlarges the dashed region; white boxes mark vehicles.

Explore similar work

May 25, 2026cs.CV

RadarSim: Simulating Single-Chip Radar via Multimodal Neural Fields

Radars are an ideal complement to cameras: both are inexpensive, solid-state sensors, with cameras offering fine angular resolution, while radars provide metric depth and robustness under adverse weather. However, radar data is more difficult to interpret than camera images and varies significantly between sensors, necessitating increased reliance on simulation for prototyping sensors and processing pipelines. Recent work treating radar reconstruction as a novel view synthesis problem has shown great promise in reconstructing radar-relevant geometry and simulating low-level radar data. However, such methods are constrained by the low spatial resolution of the underlying radar. To address this, we propose a unified differentiable renderer, RadarSim, which leverages the high angular resolution of RGB cameras to generate Doppler radar range images from a camera-initialized neural field. Using a novel data set of calibrated radar camera recordings from a custom hand-held rig, we demonstrate that RadarSim produces sharper geometry and Doppler range frames than radar-only reconstructions.

Chuhan Chen, Tianshu Huang, Akarsh Prabhakara +5

Carnegie Mellon University · Bosch Research · University of Wisconsin–Madison

Radar Doppler Velocity Log Differentiable Rendering Neural Fields

Ask PDF

Aug 4, 2026cs.CV

Radar4D-VLM: Proposal-Grounded Temporal 4D Radar Reasoning Across Frozen Language Models

Vision-language models for autonomous driving primarily rely on cameras and LiDAR, leaving 4D radar largely unexplored as a standalone perceptual modality despite its robustness to adverse visibility and direct measurement of radial velocity. We introduce Radar4D-VLM, a radar-only temporal vision-language model that reasons from ten consecutive 4D-radar point-cloud sweeps without camera or LiDAR input. Radar4D-VLM extracts geometrically grounded object proposals and organizes radar evidence into a compact hierarchy of object, scene, and kinematic tokens. A parameter-efficient projector maps these tokens into frozen language backbones, while auditable prediction heads jointly model object count, spatial distribution, motion state, collision risk, semantic category, and radial velocity. Radar4D-VLM combines proposal-grounded temporal object tokenization, global scene context, and explicit kinematic tokens within a unified frozen-backbone interface. On sequence-isolated K-Radar development validation, its Top-64 proposal recall reaches 98.13% at 4 m, exceeding fixed-lattice and uniform-random controls by 6.40 and 22.83 percentage points, respectively. We further evaluate 24 matched runs spanning eight frozen Qwen, Phi, Mistral, Llama, and Gemma backbones under an identical adaptation budget. The radar-token interface remains compatible across all five language-model families, while matched aligned, permuted, and no-language controls show sensor dependence but no stable direct-head gain from aligned language supervision. These results establish a reproducible foundation for radar-only multimodal scene and motion reasoning while separating interface compatibility from the benefit of language supervision.

Jiaju Han, Xuemeng Sun, Qike Zhang +6

China University of Petroleum-Beijing at Karamay · Shenzhen Research Institute of Big Data · Shanghai Jiao Tong University +2

4D Radar Autonomous Driving Perception Scene Reasoning Autonomous Driving Point Clouds

Ask PDF

Jun 22, 2026cs.CV

DrivingVoxels: Compositional Sparse Voxel Rasterization for Dynamic Driving Scene Reconstruction

Reconstructing dynamic urban scenes remains challenging due to the unbounded nature of driving environments and the presence of multiple dynamic objects. Currently, potentially faster sparse voxel methods are mainly designed for static scenarios. On the other hand, dynamic approaches based on 3D Gaussian Splatting, despite their high-fidelity, are often time-consuming for driving scenarios and exhibit uncontrollable memory growth in large scenes. To address these limitations, we present DrivingVoxels, a compositional sparse voxel rendering framework for dynamic driving scenes. Our method jointly rasterizes sparse voxels from multiple independent octrees within a single rendering pass. Each rigid dynamic object is represented by an octree defined in its local coordinate frame, while a separate static octree models the stationary background. DrivingVoxels adopts a fully explicit, neural-free representation together with a LiDAR-guided structural initialization that efficiently captures scene geometry. We evaluate our framework on the PandaSet benchmark, demonstrating that DrivingVoxels performs on par on perceptual metrics and better on structural metrics for NVS and reconstruction while requiring shorter training times than previous 3DGS-base methods to an efficient optimization workflow anchored by a strong LiDAR prior.

Tania Aguirre, Luis Roldão, Moussab Bennehar +4

Noah’s Ark, Huawei Paris Research Center, France · EURECOM, France

Dynamic Driving Scenes Voxel Scene Reconstruction Rasterization Octree

Ask PDF